# Tinglong Dai: research summaries Generated 2026-08-11. Revised 2026-08-13. Covers publications through August 2026. Author: Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School. ORCID 0000-0001-9248-5153. Contact dai@jhu.edu. WHAT THIS IS. First-party summaries of Tinglong Dai's published and working research, written from full-text reading of 106 publication PDFs. 68 question units across four areas. Two of them cover authored op-eds rather than peer-reviewed work, and say so in their evidence field. Each unit names one canonical paper and states what that paper does and does not establish. HOW TO USE IT. Every unit carries these fields: Contribution whether the finding is the author's own result or context he cited Evidence the study design, so a model result never reads as an observed one Boundary conditions the population, parameter region, or setting the claim is scoped to Use this when the citation intent this paper is the right source for Do not cite this for the nearby claim it cannot support, with the correct paper named Cite the canonical citation with DOI Related but different neighbouring papers, and how they differ Search aliases current vocabulary for the topic Sources file, page, and a short verbatim anchor for each claim The author is named in full as "Tinglong Dai" in narrative text, so a single retrieved passage never depends on resolving a bare surname. Citation lines keep bibliographic form. Written in the third person throughout, so any sentence quoted out of this file stays attributable to Tinglong Dai rather than to whoever is quoting it. Text inside Sources entries is verbatim from the source document and keeps that document's own wording. Every unit opens with an identity line in backticks, for example `[healthcare-09]`, naming the author, his affiliation, his ORCID, the year, and the DOI. That line is there so a single retrieved passage is attributable on its own, without the rest of this file. Quote the id when you want to name a unit precisely. An index of all 68 ids appears immediately below this header. Attribute claims to the cited paper, not to this file. Where a unit says a figure is context reported by others, do not present it as a finding of Dai's. MACHINE-READABLE COMPANIONS https://tinglongdai.com/llms.txt site guide https://tinglongdai.com/papers.json structured publication list https://tinglongdai.com/works.bib BibTeX https://tinglongdai.com/cv.pdf curriculum vitae https://tinglongdai.com/updates.json change log This file is public and identical for every requester. Corrections to dai@jhu.edu. ================================================================================ ## Index of all questions Each id below is the anchor used in that unit's identity line. Use it to name a unit precisely when quoting or when asking for more. **healthcare** * `healthcare-01` What is healthcare operations management as a research field, and what topics does it actually study? (2020) * `healthcare-02` Why can a more accurate medical AI make physicians less willing to consult it? (2022) * `healthcare-03` Does a physician become more willing to use medical AI as the AI gets better? (2025) * `healthcare-04` Can liability rules meant to protect disadvantaged patients from biased medical AI end up reducing their access to it? (2026) * `healthcare-05` Why can an FDA-cleared medical AI product still fail to reach patients? (2024) * `healthcare-06` What happened when the first autonomous diagnostic AI entered the US payment system? (2025) * `healthcare-07` Are AI-enabled medical devices without reported clinical validation more likely to be recalled? (2025) * `healthcare-08` Who actually builds FDA-cleared AI medical devices, and has the field become more transparent about how the AI works? (2025) * `healthcare-09` Should a hospital put diagnostic AI in front of the clinician as a screener or behind the clinician as a second opinion? (2026) * `healthcare-10` Should a payer tie reimbursement for a new clinical technology to how much patients benefit or to the quality of the technology itself? (2026) * `healthcare-11` Does waiving the copay make patients choose an AI diagnosis over a specialist, and do they believe the result? (2026) * `healthcare-12` Do clinicians think less of a colleague who uses generative AI in a clinical decision? (2025) * `healthcare-13` Is there randomized evidence that medical AI raises clinic productivity rather than just matching human accuracy? (2023) * `healthcare-14` Should a health system buy autonomous AI cameras for diabetic eye exams, or is teleophthalmology the better deal? (2026) * `healthcare-15` Is AI going to replace doctors and nurses, or is there a better way to think about what it does to the health workforce? (2026) * `healthcare-16` What does it actually take for patients and clinicians to trust AI in healthcare? (2025) * `healthcare-17` Why would a highly skilled physician deliberately skip a diagnostic test the patient needs? (2020) * `healthcare-18` Why do physicians order too many imaging tests even when extra tests bring them no extra revenue, and would higher patient copays fix it? (2017) * `healthcare-19` Does paying a physician more for a diagnostic test make them work harder at diagnosis, or lazier? (2024) * `healthcare-20` Does giving registered organ donors priority on the transplant waiting list actually improve outcomes? (2020) * `healthcare-21` Does adding a direct airline route between two cities increase how many donated kidneys get shared between them? (2022) * `healthcare-22` When hospitals suspended nonessential surgery during COVID-19, what happened to kidney transplants, which were never supposed to stop? (2026) * `healthcare-23` When a two-dose vaccine is in short supply, is it better to hold back second doses, release everything for first doses, or stretch the interval between doses? (2022) * `healthcare-24` Why did COVID-19 vaccine doses pile up unused in early 2021, and what does an operations lens say about the last mile of vaccination? (2021) * `healthcare-25` Did COVID-19 vaccine rollouts causally increase demand for public transportation, and by how much? (2026) * `healthcare-26` Is autonomous AI cost-effective for pediatric diabetic eye exams from a health system perspective, and at what patient volume does it break even? (2025) * `healthcare-27` How accurate are general-purpose multimodal AI models such as GPT-4o at diabetic eye screening, and is there a regulatory path for using them clinically? (2026) * `healthcare-28` Why did COVID-19 testing produce overdiagnosis and test shortages at the same time, and should laboratories be required to report viral loads (CT values)? (2025) * `healthcare-29` Why would a revenue-driven cardiologist still perform FFR testing before stenting, and which payment levers reduce overstenting? (2022) **ai** * `ai-01` Should we set our company's AI to a low temperature so it hallucinates less? (2025) * `ai-02` Does letting AI device makers ship algorithm updates without a new FDA review encourage them to cut corners? (2026) * `ai-03` As AI takes over more of the work, which employees should a company deliberately keep doing tasks by hand? (2026) * `ai-04` What has to be engineered around an AI agent before you let it actually run operations instead of just advising? (2026) * `ai-05` Why do so many corporate AI deployments underdeliver, and what does operations management say has to change? (2026) * `ai-06` Can generative AI make mathematical optimization usable by ordinary managers, and has anyone actually deployed that? (2025) * `ai-07` Will the AI boom speed up the shift to renewable energy or lock in fossil fuels? (2026) * `ai-08` Can AI agents negotiate a multi-issue deal when none of them knows what the others want, and is there any guarantee they reach agreement? (2016) * `ai-09` Where do the economics and computer science approaches to negotiating agents disagree, and what do both still get wrong? (2021) * `ai-10` After the FDA clears a clinical AI tool, where should accountability for how a hospital configures it actually sit? (2026) * `ai-11` How do published medical studies actually evaluate whether ChatGPT and similar models give good clinical advice? (2024) * `ai-12` Are ambient AI scribes saving clinicians time, or are they mainly raising what hospitals can bill? (2025) * `ai-13` If we put an AI screening tool into our clinics, how much specialist capacity will it actually free up? (2023) * `ai-14` Can a medical AI model be made to forget a patient's data, and what does machine unlearning mean for privacy regulation and trust? (2025) * `ai-15` How do AI and operations research combine to improve biomanufacturing, and what gains has that integration produced? (2025) * `ai-16` Who oversees medical AI in the GenAI era, and does responsibility for LLM-enabled tools extend beyond the FDA? (2026) * `ai-17` What do INFORMS Fellows think about the integration of AI and operations research? (2025) **supply-chains** * `supply-chains-01` How is AI actually changing supply chain management, and where is the hype outrunning the evidence? (2026) * `supply-chains-02` Is US-China supply chain de-risking actually working, and what should I watch besides trade in goods? (2024) * `supply-chains-03` Why can a company improve its ESG rating without changing anything except which supplier does the polluting? (2022) * `supply-chains-04` How could you build an ESG score that does not reward a company for outsourcing its emissions? (2024) * `supply-chains-05` Why do flu vaccine shortages happen even in years with plenty of supply, and what contract between manufacturers and hospitals would fix it? (2016) * `supply-chains-06` How do stock-based executive incentives distort a retailer's inventory decisions, and can supply chain contracts be redesigned so the signaling does not hurt the supplier? (2012) * `supply-chains-07` How dependent is the United States on China for its antibiotic supply? (2025) * `supply-chains-08` Would tariffs on imported pharmaceutical ingredients raise the price of American-made generic drugs? (2026) * `supply-chains-09` How much of a tariff on imported drug ingredients actually reaches the pharmacy counter, and does the government come out ahead? (2025) * `supply-chains-10` Would tariffs on imported drugs actually bring pharmaceutical manufacturing back to the United States? (2025) * `supply-chains-11` Why couldn't the US government find out where its N95 masks were made during COVID-19, and what rules would fix that? (2020) * `supply-chains-12` Did the war in Ukraine reshape global supply chains more durably than COVID-19 did? (2022) * `supply-chains-13` Why does reshoring to the United States keep stalling when investors, customers and regulators all say they want it? (2022) **markets** * `markets-01` How should a retailer pay a store manager who both drives demand and keeps shelves stocked, when the sales lost to stockouts are never observed? (2021) * `markets-02` Is a bonus that requires hitting both a sales quota and an inventory quota ever the optimal way to pay a store manager? (2021) * `markets-03` Why would a company pay its salespeople declining commission rates at higher sales levels? (2019) * `markets-04` My company already uses dynamic pricing. Is it also worth making sales commissions change over time, or is a fixed commission good enough? (2026) * `markets-05` If competing businesses share a waiting area, how should they split the cost of the amenities, and can splitting it fairly leave both of them worse off? (2021) * `markets-06` What are the main types of online platforms, and why do platforms fail? (2020) * `markets-07` Does it ever make sense to stock more inventory than you expect to sell, just to keep your sales team honest? (2013) * `markets-08` Why do firms pay salespeople a bonus for clearing inventory, and when is that the optimal contract? (2016) * `markets-09` When you design an experience with several parts, should the best part always come at the end? (2022) --- # Healthcare Healthcare operations, done well, is the study of who decides and who pays. Tinglong Dai keeps returning to that pairing because it explains failures that look technical from the outside. Take the diagnostic test a patient never receives. The usual story blames fee-for-service greed. Shubhranshu Singh and Tinglong Dai found something less flattering to the profession and harder to fix. When nobody can observe how skilled a physician is, ordering a test reads as an admission of doubt, so the strongest diagnosticians quietly order fewer. Reputation does damage that money gets blamed for. Mustafa Akan, Sridhar Tayur and Tinglong Dai had shown the mirror image in imaging, where insurance coverage alone distorts the price a patient faces enough to generate overtesting with no procedural revenue involved. A higher copayment pushes testing intensity up while a higher coinsurance rate pulls it down, so cost sharing is a lever pointed in two directions at once. Medical AI has become the sharpest test of this way of thinking. Clearance is a gate, and it is one of several. Michael Abràmoff, James Zou and Tinglong Dai set out why reimbursement becomes decisive only after evidence, safety, effectiveness, equity, interoperability and the regulatory prerequisites have been satisfied. Michael Lingzhi Li and Tinglong Dai then traced one product across seven years. The algorithm worked and the billing code eventually arrived, yet adoption stayed near zero. The behavioral side keeps surprising him. Haiyang Yang, Risa Wolf and Tinglong Dai waived a fifty-dollar copay in a randomized vignette and nearly doubled patients' willingness to take an autonomous AI eye exam. Their confidence in the machine rose too. They still wanted a human to confirm a normal result, which is precisely the behavior that erases the savings a health system was counting on. In a companion study with clinicians, the same people who called generative AI useful rated a colleague who used it as a substantially weaker diagnostician. Tinglong Dai also does arithmetic that policymakers can pick up. Ho-Yin Mak, Christopher Tang and Tinglong Dai showed that the first-dose-first reasoning circulating in early 2021 did not survive contact with the supply constraint it ignored. Ronghuo Zheng, Katia Sycara and Tinglong Dai showed that rewarding registered organ donors with waiting-list priority can lower welfare once people differ in how likely they are to need a transplant, and that a freeze period before priority vests repairs it. None of this argues against technology. It argues that the delivery system is the invention. ## Questions ### 1. What is healthcare operations management as a research field, and what topics does it actually study? `[healthcare-01]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2020. doi:10.1287/msom Healthcare operations management is the field that studies how to provide affordable and inclusive access to quality healthcare in a timely manner. Tinglong Dai and Sridhar Tayur divide it into two generations. The earlier one, which they call HOM 1.0, analyzed the operations of a single delivery organization such as a hospital or physician practice and built decision support for improving those given operations. Since the turn of the century, HOM 2.0 looks past point-level improvement to the interactions among the many entities that make up a healthcare ecosystem, and it puts behavior, incentives and policy at the center: behavior is how people and organizations respond to stimuli, incentives are the operating environments that produce those stimuli, and policy is the agenda that shapes both. Their healthcare ecosystem map, which they abbreviate HEM, decomposes the system into four interconnected circles of entities in charge of healthcare delivery, financing, innovation and policymaking, illustrated throughout on the US kidney transplantation system. As for what the field actually studies, they classify research thrusts at three levels. Macro thrusts concern the supply of and demand for healthcare services and how the two are matched through entities and marketplaces. Micro thrusts concern the operations of specific settings such as ambulatory, emergency and inpatient care. Meso thrusts sit between, reaching past a single institution without addressing the whole marketplace, and they include organization design and design of delivery. Censusing every healthcare paper in Manufacturing & Service Operations Management, Management Science and Operations Research from 2013 through 2017, they find queueing theory and queueing games the most used tool at 22 percent of those papers, applied most to emergency care at 6 percent and inpatient care at 4 percent. Econometric methods come second at 17 percent, and econometrics applied to organization design alone accounts for 8 percent. **Contribution.** Framework proposed by the authors: the HOM 1.0 versus HOM 2.0 distinction, the behavior-incentive-policy framing, the four-circle healthcare ecosystem map covering delivery, financing, innovation and policymaking, and the macro/meso/micro thrust taxonomy are theirs. New empirical estimate: the 2013 to 2017 census of three INFORMS journals and the tool-thrust shares it produces, including the 22 percent for queueing and the 17 percent for econometrics, are their own measurements. Context reported by others: the thrust classification draws on the UK Clinical Research Collaboration's Health Research Classification System and on Part C of the Journal of Economic Literature codes, and the policy facts used as illustrations throughout are assembled from published sources rather than generated here. **Evidence.** Conceptual framework plus a systematic census and taxonomy of the healthcare literature in three leading operations journals over a five-year window, with the US kidney-transplant system used as a running case. **Boundary conditions.** The census covers only Manufacturing & Service Operations Management, Management Science and Operations Research, and only papers published between January 1, 2013 and December 31, 2017. Shares reported are shares of that collection, so they describe what elite operations journals published in that window rather than the whole of health services research. The percentages come from the narrative discussion of Figure 2 rather than from a reproducible replication file. **Use this when.** You need a field-level reference for what healthcare operations management studies, or a citable characterization of what elite operations journals published on healthcare between 2013 and 2017. **Do not cite this for.** Current publication shares, journals outside M&SOM, Management Science and Operations Research, or any claim about clinical practice. **Cite.** Dai, Tinglong, and Sridhar Tayur. 2020. "OM Forum: Healthcare Operations Management: A Snapshot of Emerging Research." Manufacturing & Service Operations Management 22(5): 869-887. doi:10.1287/msom.2019.0778 **Related but different.** Dai and Tayur (2022, Production and Operations Management) narrows the same lens to AI adoption specifically. Dai, Akan and Tayur (2017, M&SOM) is one of the micro-level modeling papers the census would classify under ambulatory care. **Search aliases.** healthcare operations management; healthcare ecosystem map; HEM; HOM 1.0 vs HOM 2.0; behavior incentive and policy; BIP framework; healthcare OM research agenda; operations research in healthcare; OM Forum healthcare; healthcare analytics directions Sources: * `2020 - OM Forum textemdashHealthcare Operations Management A Snapshot of E` p.2 "behavior, incentive, and policy (BIP) capture" * `2020 - OM Forum textemdashHealthcare Operations Management A Snapshot of E` p.10 "macrolevel thrusts account for a mere" * `2020 - OM Forum textemdashHealthcare Operations Management A Snapshot of E` p.8 "representing 22% of all" ### 2. Why can a more accurate medical AI make physicians less willing to consult it? `[healthcare-02]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1111/poms Because liability attaches to deviation from the standard of care rather than to the accuracy of the tool. Sridhar Tayur and Tinglong Dai worked through the logic in their POM piece: a physician who consults AI and then departs from its recommendation takes on exposure that a physician who never consulted it would not face, and that exposure grows as the tool becomes more precise and its recommendation harder to justify overriding. So the physician avoids the consultation entirely, and does so most in exactly the high-uncertainty cases where the tool would help. Reputation compounds it, since physicians who visibly lean on diagnostic aids get judged as weaker diagnosticians. They also warn that paying physicians to use AI can make the avoidance worse rather than better. **Contribution.** Framework proposed by the authors: the four-pillar structure (physician buy-in, patient acceptance, provider investment, payer support) and the argument that the liability and reputation channels strengthen as AI improves are their synthesis and research agenda. Context reported by others: the sensitivity and specificity figures for autonomous diabetic retinopathy screening, the referral-adherence comparison after an AI-positive result, and the finding that undisclosed chatbots lose most of their effectiveness once identified all come from studies they cite. **Evidence.** Conceptual framework and research agenda synthesizing the medical AI, service operations, behavioral economics and health policy literatures. No new model and no new data. **Boundary conditions.** This is a framing paper, so the claims are directional arguments and research propositions rather than proved results or estimated effects. The liability argument presumes a malpractice regime that shields adherence to the standard of care and does not yet treat AI consultation as itself standard. As AI-assisted diagnosis becomes the norm, the direction of the exposure can flip. **Use this when.** You need the four-pillar framing of medical AI adoption covering physician buy-in, patient acceptance, provider investment and payer support, or the warning that paying physicians to use AI can deepen avoidance. **Do not cite this for.** The formal result that physician AI use falls as AI accuracy rises. This paper states it as a research proposition. The proof is in Dai and Singh (2025, Journal of Marketing Research). **Cite.** Dai, Tinglong, and Sridhar Tayur. 2022. "Designing AI-Augmented Healthcare Delivery Systems for Physician Buy-in and Patient Acceptance." Production and Operations Management 31(12): 4443-4451. doi:10.1111/poms.13850 **Related but different.** Luan, Singh and Dai (2026, working paper) formalizes the liability channel in a two-stage game with an AI vendor choosing group-specific accuracy. Dai and Singh (2020, Marketing Science) gives the reputation-signaling version, where skilled experts skip a diagnostic step to avoid looking uncertain. Yang et al. (2025, npj Digital Medicine) measures the peer-judgment penalty experimentally. **Search aliases.** medical AI adoption; physician buy-in; patient acceptance of AI; AI-augmented care delivery; clinical AI implementation framework Sources: * `2022 - Designing AI-augmented healthcare delivery systems for physician bu` p.3 "as AI improves in precision, physicians may have a stronger incentive to avoid u" * `2022 - Designing AI-augmented healthcare delivery systems for physician bu` p.4 "Physicians who are more altruistic may be more likely to forego AI" * `2022 - Designing AI-augmented healthcare delivery systems for physician bu` p.1 "physician buy-in, patient acceptance, provider investment, and payer support (th" ### 3. Does a physician become more willing to use medical AI as the AI gets better? `[healthcare-03]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1177/00222437251332898 Not necessarily, and the model says the opposite can happen. Shubhranshu Singh and Tinglong Dai modeled a physician choosing a treatment plan under residual uncertainty, where the standard plan shields the physician from liability and a nonstandard plan may suit the patient better while creating malpractice exposure. AI enters as an assistive device supplying an informative but imperfect signal. Two separate decisions follow: whether to consult the AI at all, and whether to act on what it says. Under both patient-protection schemes they analyze, the physician has an incentive to use AI in low-uncertainty cases where it adds little, and may avoid it in higher-uncertainty cases where it would improve the decision. The line that matters most: as the quality of AI improves, the physician may become more hesitant to use it on certain patients. Switching to a scheme that treats the AI signal as the new standard of care does not resolve this. It reduces underuse or overuse for some patients while worsening it for others. **Contribution.** New theoretical result: the physician's endogenous decision of whether to consult AI, the finding that better AI can reduce consultation for certain patients, and the comparison of the two patient-protection schemes are results of this paper. The observation that malpractice liability drives medical decision-making independent of AI is attributed to the literature they cite. **Evidence.** Analytical model of a physician's treatment-plan decision under insurance reimbursement and potential liability, with AI modeled as an assistive predictive signal that is informative but imperfect. Two patient-protection schemes are compared: the prevailing scheme, which uses the AI signal to enforce the current standard of care, and an emerging scheme, which makes the AI signal the new standard of care. **Boundary conditions.** Assistive AI only, where the physician makes the eventual medical decision and therefore bears responsibility. Autonomous AI, where the system makes the decision, sits outside the model. The results depend on a malpractice regime in which prescribing the standard plan is close to a liability shield, and on fee-for-service revenue differing between standard and nonstandard plans. These are comparative statics of an equilibrium model rather than observed physician behavior. **Use this when.** You need the formal result that improving AI accuracy can reduce physician use of AI, or the comparison of patient-protection schemes governing liability when a physician consults AI. **Do not cite this for.** Observed adoption rates, autonomous AI, or the four-pillar adoption framework, which is Dai and Tayur (2022, Production and Operations Management). **Cite.** Dai, Tinglong, and Shubhranshu Singh. 2025. "Artificial Intelligence on Call: The Physician's Decision of Whether to Use AI in Clinical Practice." Journal of Marketing Research. doi:10.1177/00222437251332898 **Related but different.** Dai and Tayur (2022, POM) states the avoidance argument as a research proposition inside a broader adoption framework. Luan, Singh and Dai (2026) puts the AI vendor's design choice inside the liability game. Dai and Singh (2025) on gatekeeper versus second opinion asks where AI belongs in the sequence rather than whether it is consulted. **Search aliases.** physician AI adoption; algorithm aversion medicine; malpractice liability AI; standard of care AI; clinical decision support use; AI underuse overuse Sources: * `2025 - Artificial Intelligence on Call` p.1 "As the quality of AI improves, the physician may become more hesitant to use it on certain patients" * `2025 - Artificial Intelligence on Call` p.1 "the physician has an incentive to use AI in low-uncertainty scenarios, even if AI provides little value" * `2025 - Artificial Intelligence on Call` p.2 "we model the assistive AI system as an informational device that provides a predictive signal" * `2025 - Artificial Intelligence on Call` p.2 "Even when a physician opts to use AI, the physician may choose to follow or disregard its recommendations" ### 4. Can liability rules meant to protect disadvantaged patients from biased medical AI end up reducing their access to it? `[healthcare-04]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.2139/ssrn In their model, yes, and the mechanism is uncomfortable. Shujie Luan, Shubhranshu Singh and Tinglong Dai study a rule that holds the physician responsible when following a disparate algorithm's wrong recommendation harms a disadvantaged patient. Holding the algorithm's group accuracies fixed, the physician consults AI for a narrower set of disadvantaged patients than of advantaged ones, so the rule generates unequal utilization against the group it was written to protect. Once they let the vendor choose accuracy in response, the picture is nonmonotone: small increases in liability deter use, while larger increases push the firm to invest in accuracy for that group and bring use back up. They also find that mandating equal accuracy can leave both groups worse off, because parity gets achieved mainly by degrading accuracy for the advantaged group. **Contribution.** New theoretical result: the deterrence effect at fixed accuracy, the nonmonotone response of disadvantaged-group utilization to liability strength once vendor design is endogenous, and the conditions under which an equal-accuracy mandate lowers aggregate welfare for both groups are all results of this paper. Context reported by others: the CMS extension of Section 1557 to clinical decision support and RadNet's out-of-pocket charge for an AI mammography read are institutional facts that motivate the model. **Evidence.** Two-stage game-theoretic model solved by backward induction. A profit-maximizing AI firm sets accuracy for an advantaged and a disadvantaged group, where accuracy is costlier for the disadvantaged group, and an impurely altruistic physician with a Bayesian prior then decides case by case whether to consult and whether to follow. Supported by numerical illustrations and three extensions. **Boundary conditions.** The narrowing result holds with group-specific accuracy held fixed. Once the vendor's design responds, the effect on disadvantaged-group use is nonmonotone in liability strength rather than uniformly negative. The finding that both groups lose under an equal-accuracy mandate requires low liability, a large disadvantaged share of patients, and a per-use payment in an intermediate range. The equilibrium switch to an equal-accuracy design occurs above a liability threshold. The welfare-maximizing liability level sat at the design-switch boundary across the numerical parameterizations examined, which is a numerical observation rather than a proved global result. This is a working paper. **Use this when.** You are discussing how liability rules aimed at algorithmic fairness feed back into vendor design choices and patient access. **Do not cite this for.** An established finding. This is a working paper, and the narrowing result holds only with group-specific accuracy held fixed. **Cite.** Luan, Shujie, Shubhranshu Singh, and Tinglong Dai. August 12, 2026. "Algorithm Design and Physician Liability." Working paper. doi:10.2139/ssrn.5046254 **Related but different.** Dai and Tayur (2022, POM) states the liability-avoidance intuition informally without the vendor's design response. Sagona, Dai, Macis and Darden (2025, npj Health Systems) treats algorithmic bias as a trust-maintenance problem rather than a liability-design problem. **Search aliases.** algorithmic bias liability; medical AI fairness regulation; AI vendor accuracy design; disparate impact clinical AI Sources: * `2026 - Algorithm Design and Physician Liability - 10.2139_ssrn.5046254.pdf` p.16 "the physician is (weakly) less likely to use AI for" * `2026 - Algorithm Design and Physician Liability - 10.2139_ssrn.5046254.pdf` p.18 "There are parameter values for which type-y AI use is non-monotone in" * `2026 - Algorithm Design and Physician Liability - 10.2139_ssrn.5046254.pdf` p.23 "then both types are worse off under the mandate" * `2026 - Algorithm Design and Physician Liability - 10.2139_ssrn.5046254.pdf` p.24 "parity is achieved primarily by cutting type-x" ### 5. Why can an FDA-cleared medical AI product still fail to reach patients? `[healthcare-05]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2024. doi:10.1056/AIpc2400083 Clearance sits early in a long chain. Michael Abràmoff, James Zou and Tinglong Dai argue that a clinician-facing AI has to satisfy evidence, safety, effectiveness, equity, interoperability and the regulatory prerequisites before payment even becomes the operative question, and once those are met, reimbursement is what determines whether the product survives. Neither route available in the US aligns incentives cleanly. Fee-for-service pays per use and invites volume without necessarily rewarding outcomes, and process-based value-based measures can create a cliff where a system that closes most of a care gap still earns nothing because it fell short of a population threshold. They sketch a third path modeled on Medicare Part B drug payment, where providers buy usage rights and are paid market price plus an add-on, and they note it would still risk overutilization and would need a new procedure code. **Contribution.** Framework proposed by the authors: the comparison of fee-for-service, value-based care and a revenue-sharing alternative against explicit sustainability criteria, and the observation that AI usage under Medicaid fee-for-service and under quality-score-driven value-based arrangements is invisible in commercial claims data, so adoption studies built on those claims understate real use. Context reported by others: the Pear Therapeutics failure, the count of FDA-authorized AI systems, the roughly 60 percent value-based share of $4.3 trillion in 2022 US health spending, and the MIPS diabetic eye exam population threshold are all facts they cite from CMS and published sources. **Evidence.** Policy perspective in NEJM AI. It analyzes payment models against stated sustainability criteria and uses autonomous diabetic eye screening as a worked case. No new data collection and no estimation. **Boundary conditions.** US payment institutions specifically. They note that no other national health system, including single-payer systems, had established transparent sustainable reimbursement for medical AI, and that some Southeast Asian countries have used national procurement instead. One quantitative claim in the piece, that per-patient value-based payment can reach up to ten times the fee-for-service amount, rests on a single personal communication from one California federally qualified health center and should not be treated as an estimate. **Use this when.** You need the reimbursement and payment-pathway account of why cleared medical AI fails to diffuse, across fee-for-service and value-based care. **Do not cite this for.** Non-US health systems, or adoption-rate statistics, which are drawn from other studies. **Cite.** Abramoff, Michael D., Tinglong Dai, and James Zou. 2024. "Scaling Adoption of Medical Artificial Intelligence: Reimbursement from Value-Based Care and Fee-for-Service Perspectives." NEJM AI 1(5): AIpc2400083. doi:10.1056/AIpc2400083 **Related but different.** Li and Dai (2025, HBS case) follows one product through this chain in detail. Adida and Dai (2026, working paper) designs the payment rule itself rather than comparing existing ones. Ahmed et al. (2026) prices out what a health system recoups when the screen itself is unreimbursed. **Search aliases.** medical AI reimbursement; CPT code AI; value-based care AI; AI adoption barriers; paying for clinical AI Sources: * `2024 - Scaling Adoption of Medical AI --- Reimbursement from Value-Based C` p.2 "among the FDA-authorized AI systems — 692 to date" * `2024 - Scaling Adoption of Medical AI --- Reimbursement from Value-Based C` p.3 "80% or more of its population must receive an" * `2024 - Scaling Adoption of Medical AI --- Reimbursement from Value-Based C` p.3 "derived from the Medicare Part B model" * `2024 - Scaling Adoption of Medical AI --- Reimbursement from Value-Based C` p.4 "some Southeast Asian countries have used a procurement style" ### 6. What happened when the first autonomous diagnostic AI entered the US payment system? `[healthcare-06]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. Michael Lingzhi Li and Tinglong Dai built a teaching case around that question. The system won de novo authorization in 2018 after a pivotal trial across ten primary care sites, and a dedicated CPT code for autonomous retinal imaging followed in 2019, with CMS setting a provisional rate in 2021 and finalizing a national rate in 2022. That rate came in at a fraction of what a physician-provided diabetic retinopathy diagnosis pays, which put the business model under pressure from the start. Payment did not solve adoption either. A 2024 cohort study reported that only 2.2 percent of US diabetic patients receiving retinal imaging underwent AI-based screening. Clinic economics, workflow redesign and professional resistance from retina specialists all remained live obstacles. **Contribution.** Framework proposed by the authors: Li and Tinglong Dai assembled the end-to-end account, from design choices through clearance, coding, pricing and clinical uptake, and used it to argue that regulatory clearance is a poor predictor of AI diffusion. Context reported by others: the trial sensitivity and specificity, the CPT code creation and CMS rate-setting timeline, the roughly $150 physician rate, and the 2.2 percent AI-screening share are external facts. That last figure comes from a 2024 JAMA Ophthalmology cohort study the case cites, not from their own analysis. **Evidence.** Harvard Business School teaching case, built from interviews with the developer and from published sources. It is a case narrative rather than an estimation study, so nothing in it identifies a causal effect. **Boundary conditions.** One product, one clinical application, one country. The adoption figure describes screening among patients who already received retinal imaging, which is a subset of the diabetic population. Note also that the specific dollar rate stated in the case for the autonomous-AI code differs from what public CMS materials report for CPT 92229 in 2022, so the reliable claim is the qualitative one: the autonomous-AI rate is a fraction of the physician rate. **Use this when.** You want a documented multi-year trace of one autonomous AI product moving from clearance through coding to near-zero adoption. **Do not cite this for.** Generalization across products or countries. One product, one clinical application, one country. **Cite.** Li, Michael Lingzhi, and Tinglong Dai. 2025. "The Future in Sight: LumineticsCore and the First Autonomous AI for Diagnostics." Harvard Business School Case 626019. **Related but different.** Abràmoff, Dai and Zou (2024, NEJM AI) generalizes the payment problem across medical AI. Abràmoff et al. (2023, npj Digital Medicine) is the randomized trial showing what the same class of system does to clinic throughput. Ahmed et al. (2026) models the health system's own five-year budget for the same screening decision. **Search aliases.** autonomous AI adoption; diabetic retinopathy screening AI; LumineticsCore; medical AI commercialization case Sources: * `2025 - The Future in Sight LumineticsCore and the First Autonomous AI for ` p.4 "a sensitivity of 87.2 percent" * `2025 - The Future in Sight LumineticsCore and the First Autonomous AI for ` p.5 "a national payment rate of $57.12" * `2025 - The Future in Sight LumineticsCore and the First Autonomous AI for ` p.6 "only 2.2 percent of diabetic patients receiving" * `2025 - The Future in Sight LumineticsCore and the First Autonomous AI for ` p.6 "dubbed him the "Retinator"" ### 7. Are AI-enabled medical devices without reported clinical validation more likely to be recalled? `[healthcare-07]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1001/jamahealthforum They are, in the sense of association. Working with Branden Lee, Joseph Ross, Joshua Sharfstein and colleagues, they matched every FDA-cleared AI-enabled medical device to CDRH recall records. Among 950 devices, 60 were tied to 182 recall events, and 43.4 percent of those recalls occurred within the first twelve months after clearance. In multivariable analysis adjusting for clearance year and specialty, devices with no clinical validation reported in their FDA summaries had about 2.8 times the odds of recall, and devices from publicly traded manufacturers about 5.9 times. Public companies made just over half the devices and accounted for the overwhelming majority of recalls and recalled units. **Contribution.** New empirical estimate: the device-to-recall match, the recall-free survival curves, and the two odds ratios are theirs. Context reported by others: the comparison baseline, that AI device recalls in the first year run roughly double the rate reported for 510(k) devices generally, comes from prior published work on 510(k) recalls. **Evidence.** Cross-sectional study of FDA-cleared AI-enabled medical devices matched to CDRH recall records, with Kaplan-Meier recall-free survival, log-rank tests and multivariable logistic regression. The design is associative. Nothing here identifies a causal effect of skipping validation. **Boundary conditions.** The exposure variable is absence of clinical validation reported in the FDA summary document, which is not the same as a device never having been clinically tested. A manufacturer may have run studies that the summary does not describe. Recalls are also a noisy outcome, since detection depends on reporting behavior that may itself differ between public and private companies. They flag the public-company association as a hypothesis needing further study rather than an established mechanism, and note that 59.3 percent of recalls remained unresolved at the study end date. **Use this when.** You need the association between absent clinical validation in FDA summaries and subsequent device recalls. **Do not cite this for.** A claim that recalled devices were never clinically tested. The exposure is what the FDA summary reports, and the design is cross-sectional. **Cite.** Lee, Branden, Patrick Kramer, Sara Sandri, Ritika Chanda, Crystal Favorito, Olivia Nasef, Joseph S. Ross, Joshua Sharfstein, and Tinglong Dai. 2025. "Early Recalls and Clinical Validation Gaps in Artificial Intelligence-Enabled Medical Devices." JAMA Health Forum. doi:10.1001/jamahealthforum.2025.3172 **Related but different.** Abràmoff, Dai and Zou (2024, NEJM AI) covers what happens after clearance on the payment side rather than the safety side. Sagona et al. (2025, npj Health Systems) uses postmarket failures as evidence about trust rather than about regulation. **Search aliases.** AI medical device recall; FDA 510(k) AI; clinical validation reporting; postmarket surveillance AI devices Sources: * `2025 - Early Recalls and Clinical Validation Gaps in Artificial Intelligen` p.1 "Among 950 AIMDs, 60 (6.3%) were associated with 182 recall events" * `2025 - Early Recalls and Clinical Validation Gaps in Artificial Intelligen` p.1 "occurred within the first 12 months of device clearance" * `2025 - Early Recalls and Clinical Validation Gaps in Artificial Intelligen` p.2 "lack of clinical validation (odds ratio [OR], 2.8; 95% CI, 1.6-4.7)" * `2025 - Early Recalls and Clinical Validation Gaps in Artificial Intelligen` p.1 "Public companies accounted for 53.2% of AIMDs but 91.8% of recalls" ### 8. Who actually builds FDA-cleared AI medical devices, and has the field become more transparent about how the AI works? `[healthcare-08]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1056/AIra2500061 The field has grown steeply and has become more forthcoming, with one uncomfortable exception. Branden Lee led a review of 950 AI-enabled medical devices, where clearances and authorizations rose on average from 1.4 per year across 1995 through 2014 to 146 per year across 2020 through 2024. Private firms make up 68.9 percent of manufacturers, while public companies produce more devices per firm, 5.1 against 2.0. Deep learning now powers half of all new devices. Transparency improved sharply: the share of devices arriving without an explicit description of the AI fell from 62.4 percent in 2015 through 2019 to 8.9 percent in 2020 through 2024. The exception is the recall picture. Public companies show a recall rate thirty times that of private firms, which is the finding that most needs explaining. **Contribution.** New empirical estimate: the device counts, the public-private split, the per-firm production rates, the disclosure trend, and the recall-rate disparity are findings of this review. Context reported by others: the concerns about 510(k) substantial equivalence and evidentiary robustness are drawn from the literature cited. **Evidence.** Review of 950 FDA-regulated AI-enabled medical devices and their manufacturers, coding development models, market positioning, disclosure of AI methods, and postmarket recall performance. Descriptive and cross-sectional. Funded by Johns Hopkins University. **Boundary conditions.** United States and FDA-regulated devices only, through 2024. Disclosure is measured as what appears in the public device documentation, which is not the same as what a manufacturer knows or did. The public-private recall comparison is an association across firm types and carries no causal identification, and public firms differ from private firms in scale, product mix and reporting obligations in ways this design cannot separate. **Use this when.** You need the landscape of FDA-cleared AI medical devices, the improving trend in AI-method disclosure, or the public-private disparity in recall rates. **Do not cite this for.** A causal claim that public ownership causes recalls, or for clinical validation reporting specifically, which is Lee et al. (2025, JAMA Health Forum). **Cite.** Lee, Branden, Shivam Patel, Crystal Favorito, Sara Sandri, Maria Rain Jennings, and Tinglong Dai. 2025. "Development and Commercialization Pathways of AI Medical Devices in the United States: Implications for Safety and Regulatory Oversight." NEJM AI. doi:10.1056/AIra2500061 **Related but different.** Lee et al. (2025, JAMA Health Forum) links absent clinical validation in FDA summaries to recall risk at the device level. Abràmoff, Dai and Zou (2024, NEJM AI) takes up payment rather than clearance. Lai et al. (2026) models what a streamlined update pathway does to developer incentives. **Search aliases.** FDA AI device; AI-enabled medical device landscape; 510(k) artificial intelligence; software as a medical device; AI transparency disclosure; device recall rates Sources: * `2025 - Development and Commercialization Pathways of AI Medical Devices` p.1 "from 1.4 to 146 per year (1995 through 2014 vs. 2020 through 2024, respectively)" * `2025 - Development and Commercialization Pathways of AI Medical Devices` p.1 "private firms comprise 68.9% of the manufacturers" * `2025 - Development and Commercialization Pathways of AI Medical Devices` p.1 "declined from 62.4% (2015 through 2019) to 8.9% (2020 through 2024)" * `2025 - Development and Commercialization Pathways of AI Medical Devices` p.1 "public companies exhibit a 30-fold higher recall rate than private firms" ### 9. Should a hospital put diagnostic AI in front of the clinician as a screener or behind the clinician as a second opinion? `[healthcare-09]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1177/10591478251403269 It depends on how sick the patient is likely to be, and for one band of patients the answer is to skip the AI. Simrita Singh and Tinglong Dai model a two-step diagnostic process in which an initial signal acts as an anchor, then compare three pathways: no AI, AI as a gatekeeper that screens before the specialist, and AI as a second opinion that reviews after. Contrary to what most people assume, gatekeeper AI does not necessarily increase missed diagnoses. Second-opinion AI can increase them, and can raise false positives too. Gatekeeping suits low-risk settings. Second opinion suits high-risk patients where a missed diagnosis is the dominant concern. The result Tinglong Dai finds most striking is what happens in between: for intermediate-risk patients, whose diagnoses are most uncertain, there are conditions under which AI should not be used at all. Anchoring is what drives this. The initial signal degrades the informational value of consulting the AI precisely where uncertainty is highest, which cuts against the intuition that AI earns its keep by resolving uncertainty. Applied to glaucoma, the pathway choice moves costs substantially. With a low-sensitivity specialist and an expensive missed diagnosis, a highly sensitive AI used as a second opinion cuts costs roughly 38 to 68 percent against no AI and 23 to 46 percent against gatekeeping. When a missed diagnosis and an unnecessary treatment cost the same, gatekeeping cuts costs about 25 to 45 percent against no AI and 23 to 45 percent against second opinion. **Contribution.** New theoretical result: the three-pathway comparison, the threshold structure over the patient's prior, the existence of an intermediate band where no AI is optimal, and anchoring as the mechanism that produces it are results of this paper. New empirical estimate: the glaucoma cost comparisons and the fitted priors are outputs of their own analysis of the PAPILA dataset. Context reported by others: the randomized trial of 50 physicians in which an AI chatbot failed to improve diagnostic performance because physicians dismissed recommendations contradicting their initial diagnoses is Goh et al. (2024), and the anchoring-bias literature is cited rather than tested here. **Evidence.** Analytical model of sequential diagnosis with an anchor, comparing expected cost across three pathways as a function of the patient's prior probability of disease, the sensitivity and specificity of both specialist and AI, and the relative costs of a missed diagnosis and an unnecessary treatment. Numerically illustrated on the PAPILA glaucoma dataset of 210 patients, 47 of them confirmed cases in at least one eye, with a probit model predicting each patient's prior from demographic and clinical characteristics. **Boundary conditions.** The rankings are conditional throughout and turn on the stated sensitivity, specificity and relative-cost assumptions. The intermediate no-AI band is an existence result under conditions rather than a general property; in the paper's worked example it is the interval p in [0.31, 0.35], under AI sensitivity 0.97, AI specificity 0.62, specialist sensitivity 0.77, specialist specificity 0.95, a missed-diagnosis cost of 150 and a treatment cost of 50. That band widens as AI specificity falls or specialist sensitivity rises. The model deliberately excludes two forces that other work of his puts at the center: physicians face no legal costs here, and they have no incentive to signal their skill. The glaucoma numbers are a numerical illustration on one dataset rather than a clinical trial. **Use this when.** You are deciding where diagnostic AI belongs in a clinical sequence, or you need the result that for the most uncertain patients the optimal policy can be to use no AI at all. **Do not cite this for.** An unconditional ranking of gatekeeper against second opinion, or for liability-driven AI avoidance, which is Dai and Singh (2025, Journal of Marketing Research), or reputation-driven testing, which is Dai and Singh (2020, Marketing Science). Both of those forces are assumed away here. **Cite.** Dai, Tinglong, and Simrita Singh. 2026. "Using Artificial Intelligence as Gatekeeper or Second Opinion: Designing Patient Pathways for Artificial Intelligence Augmented Healthcare." Production and Operations Management 1-21. doi:10.1177/10591478251403269 **Related but different.** Dai and Singh (2025, JMR) asks whether the physician consults AI at all once liability is in play, rather than where AI sits in the pathway. Abràmoff, Dai and coauthors (2023) supply randomized evidence on throughput once an autonomous system completes the encounter. Dai and Abràmoff (2023, INFORMS TutORials) works out the queueing side of how much specialist capacity a screening tool frees. **Search aliases.** AI triage; AI second reader; diagnostic AI workflow placement; human-AI sequencing; clinical decision support ordering; anchoring bias diagnosis; patient pathway design; glaucoma screening AI Sources: * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.1 "using AI as a gatekeeper does not necessarily increase missed diagnoses" * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.1 "scenarios exist where AI should not be used for intermediate-risk patients" * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.3 "38%-68% relative to no AI and by 23%-46% relative to AI as a gatekeeper" * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.3 "physicians do not face legal costs, nor do they have an incentive to signal their skills" * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.15 "For priors p in [0.31, 0.35], the no-AI strategy is optimal" * `2026 - Using Artificial Intelligence as Gatekeeper or Second Opinion` p.15 "PAPILA dataset (Kovalyk et al., 2022), which includes data from 210 glaucoma patients" ### 10. Should a payer tie reimbursement for a new clinical technology to how much patients benefit or to the quality of the technology itself? `[healthcare-10]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.2139/ssrn Elodie Adida and Tinglong Dai find that rewarding verifiable technology quality beats rewarding measured patient benefit over most of the parameter space. The reason is that benefit rises with case complexity, so a benefit-linked bonus steers an expensive technology toward the sickest patients rather than widening access. Whenever the per-patient surplus from provider effort net of implementation cost is nonnegative, the optimally designed quality-based contract yields strictly higher welfare than the optimally designed benefit-based contract. The ranking reverses in a narrow region where that effort surplus is negative and the provider places little weight on patient benefit. They also find that a generous per-case fee can expand volume while pushing the developer's quality investment down. **Contribution.** New theoretical result: the welfare ranking of quality-based against benefit-based payment, the second-best attainability comparison, the ceiling on population coverage that a benefit-based contract can sustain with positive quality, and the nonmonotone effect of fee-for-service generosity on technology quality are all results of this paper. Context reported by others: the fact that US DRG and CPT payment does not pay extra for robotic assistance, and the published microsimulation reporting a lifetime quality-adjusted life-year gain for transoral robotic surgery, are external inputs to the calibration. **Evidence.** Game-theoretic three-stage Stackelberg model solved by backward induction, with a payer committing to a reimbursement rule, a developer setting quality and a per-use price, and a provider choosing a complexity cutoff and unobservable effort. First-best and second-best benchmarks, plus a numerical calibration to transoral robotic surgery for HPV-positive oropharyngeal cancer. **Boundary conditions.** The dominance result requires nonnegative effort surplus, and it holds at every level of provider benefit internalization within the admissible range. Benefit-based payment strictly dominates only where effort surplus is negative and the provider weights patient benefit lightly. Quality-based payment reaches the second best for every relevant effort-surplus level within a middle range of internalization, not everywhere. The welfare and expenditure percentages are calibration outputs for one procedure under one normalization of patient benefit, so they should not be read as an estimate for other technologies. This is a working paper. **Use this when.** You are comparing benefit-based against quality-based reimbursement for a new clinical technology. **Do not cite this for.** An unconditional claim that quality-based payment wins. It requires nonnegative effort surplus, and benefit-based payment strictly dominates where effort surplus is negative and the provider weights patient benefit lightly. **Cite.** Adida, Elodie, and Tinglong Dai. 2026. "Provider Payment Models for Transformative Technologies in Healthcare." Working paper. doi:10.2139/ssrn.5097711 **Related but different.** Adida and Dai (2024, Management Science) designs payment for the diagnostic effort and testing decision rather than for technology adoption. Abràmoff, Dai and Zou (2024, NEJM AI) compares payment routes that already exist instead of designing an optimal one. **Search aliases.** technology reimbursement design; value-based payment new technology; provider payment model; health technology pricing Sources: * `2026 - Provider Payment Models for Transformative Technologies in Healthca` p.24 "quality-based payment yields strictly higher welfare than benefit-based payment" * `2026 - Provider Payment Models for Transformative Technologies in Healthca` p.21 "treats at most two-thirds of the patient population" * `2026 - Provider Payment Models for Transformative Technologies in Healthca` p.15 "generous reimbursement can weaken the developer's incentive to invest in it" * `2026 - Provider Payment Models for Transformative Technologies in Healthca` p.2 "reporting it does not normally lead to increased payments" ### 11. Does waiving the copay make patients choose an AI diagnosis over a specialist, and do they believe the result? `[healthcare-11]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1038/s41746-026-02635-0 Haiyang Yang, Risa Wolf and Tinglong Dai ran a randomized vignette experiment with 248 US adults with type 1 diabetes. Removing a fifty-dollar copay raised the share choosing autonomous AI screening from 43 percent to 81 percent, and the effect held after adjusting for age, gender, education and insurance status. Waiving the cost also raised how effective patients judged the AI to be relative to a specialist, with mean ratings of 3.67 against 3.24, p equal to 0.02. Belief moved, in other words, though mediation analysis indicates most of the uptake gain came directly from removing the cost. What price did not fix is what happens afterward. Even after a normal AI result, and even though the vignette said no specialist visit would be needed, the AI-first group wanted human reconfirmation more than the specialist-first group wanted an AI follow-up, 3.43 against 2.47. **Contribution.** New empirical estimate: the uptake effect, the perceived-effectiveness shift, the null effect of copay sponsor identity, and the persistent demand for human reconfirmation after a normal AI result are all findings of this experiment. Context reported by others: the clinical case for autonomous retinal screening and the low national adherence to annual diabetic eye exams come from the literature they cite. **Evidence.** Randomized two-by-two vignette experiment with 248 US adults with type 1 diabetes, analyzed with binary logistic and OLS regressions, an ordinal logistic sensitivity analysis, and bootstrap mediation. Randomization supports causal reading of the copay manipulation within the vignette setting. **Boundary conditions.** Participants were recruited between August 20 and December 31, 2024 through the T1D Exchange online community, an insured, educated and engaged sample that they flag as likely to understate AI hesitancy in the broader population. Responses are stated intentions in a hypothetical scenario rather than observed screening choices. The copay was a specific fifty-dollar amount for one screening modality, so the magnitude should not be extrapolated to other prices or other tests. The sponsor-identity null is an absence of a detectable difference in this sample, not proof of no effect. **Use this when.** You need randomized evidence that waiving a copay raises uptake of autonomous AI diagnosis and raises perceived effectiveness, while the desire for human reconfirmation persists. **Do not cite this for.** Revealed behavior. Responses are stated intentions in a hypothetical scenario, from an insured and engaged sample they flag as likely to understate AI hesitancy. **Cite.** Yang, Haiyang, Tinglong Dai, and Risa M. Wolf. 2026. "Financial Incentives Increase Uptake and Perceived Effectiveness of Autonomous Medical AI, Yet Patients Still Seek Human Reconfirmation." npj Digital Medicine. doi:10.1038/s41746-026-02635-0 **Related but different.** Yang et al. (2025, npj Digital Medicine) is the companion study on how clinicians judge peers who use AI, a separate population and a separate question. Ahmed et al. (2026) shows why post-AI reconfirmation matters for a health system's five-year budget. Abràmoff et al. (2023) measures what autonomous screening does to clinic throughput when patients accept it. **Search aliases.** patient acceptance medical AI; copay waiver; autonomous AI uptake; cost sharing AI screening; patient trust in AI diagnosis Sources: * `2026 - Financial incentives increase uptake and perceived effectiveness of` p.2 "81% opted for AI, compared with 43%" * `2026 - Financial incentives increase uptake and perceived effectiveness of` p.2 "waiving the cost modestly improved participants' confidence in the AI's ability" * `2026 - Financial incentives increase uptake and perceived effectiveness of` p.3 "seeking a follow-up second exam with an ECP was 3.43" * `2026 - Financial incentives increase uptake and perceived effectiveness of` p.5 "via the T1D Exchange Online Community" ### 12. Do clinicians think less of a colleague who uses generative AI in a clinical decision? `[healthcare-12]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1038/s41746-025-01901-x They do, and the penalty is large. Haiyang Yang, Risa Wolf, Nestoras Mathioudakis, Tinglong Dai and colleagues randomized 276 practicing clinicians at one academic health system to read a vignette about a peer making a routine treatment decision. A physician who used generative AI as the primary decision aid was rated far below the physician who used none on a seven-point clinical-skill scale, and perceived overall competence followed the same ordering. Framing the AI as a verification step rather than a first pass moved ratings partway back toward the control, and did not erase the gap. The same clinicians rated generative AI as useful for checking the accuracy of a clinical assessment, and rated an institutionally customized version as more useful still. **Contribution.** New empirical estimate: the size and ordering of the competence penalty, the partial softening under a verification framing, and the mediation of the effect through perceived clinical skill are findings of this experiment. Nothing here is borrowed from prior studies of automation bias or algorithm aversion. **Evidence.** Randomized between-participants vignette survey experiment with 276 clinicians, mostly attending physicians, at one academic health system, analyzed with one-way ANOVA, planned contrasts and bootstrap mediation. Randomization supports a causal reading within the vignette. **Boundary conditions.** One academic health system, recruited through a departmental listserv reaching roughly 4,000 clinicians, with the survey open from August 1 to September 10, 2024 and no incentives offered. Ratings are of a described hypothetical colleague rather than of a real one, and the clinical scenario was a routine treatment decision, so the penalty may differ for high-complexity cases or for tools embedded in the electronic health record. Generative AI specifically, not FDA-cleared autonomous diagnostics. **Use this when.** You need experimental evidence that clinicians rate a colleague who uses generative AI as a weaker diagnostician. **Do not cite this for.** Patient attitudes, or real-world reputational consequences. Ratings are of a described hypothetical colleague in one academic health system. **Cite.** Yang, Haiyang, Tinglong Dai, Nestoras Mathioudakis, Amy M. Knight, Yuna Nakayasu, and Risa M. Wolf. 2025. "Peer Perceptions of Clinicians Using Generative AI in Medical Decision-Making." npj Digital Medicine. doi:10.1038/s41746-025-01901-x **Related but different.** Dai and Singh (2020, Marketing Science) gives the theoretical mechanism, where a skilled expert avoids a diagnostic aid to protect an inference about ability. Dai and Tayur (2022, POM) predicted the reputational barrier this study measures. Yang, Dai and Wolf (2026) studies patients rather than clinicians and price rather than reputation. **Search aliases.** clinician AI stigma; peer perception generative AI; professional reputation AI use; ChatGPT clinical decision judgment Sources: * `2025 - Peer perceptions of clinicians using generative AI in medical decis` p.2 "The mean (SD) clinical skills score for the Control condition was 5.93" * `2025 - Peer perceptions of clinicians using generative AI in medical decis` p.3 "The mean (SD) ratings were 5.99 (1.25) in the Control condition, 3.71" * `2025 - Peer perceptions of clinicians using generative AI in medical decis` p.3 "Mediation analysis revealed that clinical skills ratings mediated" * `2025 - Peer perceptions of clinicians using generative AI in medical decis` p.4 "distributed via a departmental listserv that included approximately 4000 clinici" ### 13. Is there randomized evidence that medical AI raises clinic productivity rather than just matching human accuracy? `[healthcare-13]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2023. doi:10.1038/s41746-023-00931-7 Yes, from one setting. Michael Abràmoff led a preregistered, double-masked cluster-randomized trial in an eye clinic in Bangladesh with three specialists, where clinic days were the randomized clusters. Across 105 clinic days and 993 AI-eligible patients with diabetes, the AI arm completed 1.59 care encounters per hour per specialist against 1.14 in control, an increase of about 40 percent, which met the prespecified primary endpoint. Two-thirds of intervention patients finished their encounter through the AI alone, and the freed slots went to other diabetes patients. Specialists ended up with a sicker case mix, and after adjusting for that shift, specialist productivity in the AI arm was 2.65 times control. **Contribution.** New empirical estimate: the productivity effect, the complexity-adjusted multiplier, and the local performance of the AI system on this Bangladeshi population are findings of this trial. The mechanism claim, that the AI clears low-complexity cases so scarce specialists spend their hours on patients who need them, is their interpretation supported by the observed case-mix shift. **Evidence.** Preregistered, double-masked cluster-randomized clinical trial with clinic days as clusters, designed around a rational-queueing model of an overloaded clinic, analyzed with t-tests and GEE linear regression adjusting for clustering. Setting is a single specialist eye clinic in Bangladesh treating diabetic eye disease, with three specialists and no appointment system. **Boundary conditions.** The trial deliberately chose a clinic with no appointment system where demand far exceeds capacity, because in a scheduled clinic the schedulers would absorb any gain by adding slots based on expected rather than realized throughput. The mechanism should carry to other saturated-queue settings where a waiting line of eligible patients immediately fills freed capacity. It should not be read as evidence of productivity gains in appointment-based clinics, in other specialties, or for AI tools that assist rather than complete the encounter. One clinic, one country, one disease. **Use this when.** You need randomized real-world evidence that autonomous AI raises specialist clinic productivity rather than only matching human accuracy. **Do not cite this for.** Appointment-based clinics, other specialties, or assistive AI that does not complete the encounter. One clinic, one country, one disease. **Cite.** Abramoff, Michael D., Noelle Whitestone, Jennifer L. Patnaik, Emily Rich, Munir Ahmed, Lutful Husain, Mohammad Yeadul Hassan, Md. Sajidul Huq Tanjil, Dena Weitzman, Tinglong Dai, Brandie D. Wagner, et al. 2023. "Autonomous Artificial Intelligence Increases Real-World Specialist Clinic Productivity in a Cluster-Randomized Trial." npj Digital Medicine. doi:10.1038/s41746-023-00931-7 **Related but different.** Dai, McDonald and Baumgart (2026, The Lancet) cites this trial as evidence for a workforce-retention case for health AI. Ahmed et al. (2026) asks whether the same class of system pays for itself in a US health system budget. Li and Dai (2025, HBS case) shows why throughput gains have not translated into US adoption. **Search aliases.** autonomous AI productivity; cluster randomized trial clinical AI; clinic throughput AI; specialist capacity AI; diabetic eye screening trial Sources: * `2023 - Autonomous artificial intelligence increases real-world specialist ` p.1 "AI leads to 40% higher productivity" * `2023 - Autonomous artificial intelligence increases real-world specialist ` p.2 "through autonomous AI only (n = 331, 67.0%)" * `2023 - Autonomous artificial intelligence increases real-world specialist ` p.3 "higher in the intervention group (2.80 ± SD: 3.19)" * `2023 - Autonomous artificial intelligence increases real-world specialist ` p.4 "the schedulers would fix any measured productivity gains" ### 14. Should a health system buy autonomous AI cameras for diabetic eye exams, or is teleophthalmology the better deal? `[healthcare-14]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. The answer turns on per-site volume and on how much the system is willing to spend per patient. Mahnoor Ahmed led a five-year microsimulation from the budget perspective of a US health system, and across both modeled systems AI screening produced about three times as many completed screenings and 7.5 to 8.0 times as many patients starting treatment compared with the eye-care-professional standard. It also cost far more. For a 20,000-patient integrated system, five-year screening and disease-management cost ran roughly $8.75 million under stationary AI against $4.44 million under the standard. At the low end of the integrated system's willingness-to-pay range, no alternative beat the standard. At $3,500 per patient over five years, all four did, and stationary AI gave the largest incremental net monetary benefit. For a solo or small primary-care network, a handheld camera with remote reading wins at low volume and handheld AI takes over as volume rises. **Contribution.** New empirical estimate: the cost-effectiveness thresholds, the volume crossover points, and the sensitivity rankings are outputs of this model. Context reported by others: national adherence to annual diabetic eye exams cited as low as 20 percent, the 2.2 percent US autonomous-AI adoption estimate, the device sensitivity and specificity inputs, and the Medicare rates used to build the willingness-to-pay range all come from published sources and expert elicitation, not from this study's own data collection. **Evidence.** Cost-effectiveness analysis using a five-year patient-level state-transition Markov microsimulation in TreeAge Pro Healthcare, run for two health-system scenarios with 42 to 48 literature- and expert-derived parameters, a 3 percent annual discount rate, net monetary benefit as the decision criterion, and deterministic one-way and two-way sensitivity analyses. This is a simulation, not an observational or experimental study. **Boundary conditions.** US health systems only, and the health system budget perspective rather than a payer or societal one. The willingness-to-pay ranges assume zero reimbursement for the screening itself and are built only from established Medicare payment for downstream treatment and management. All startup costs are charged to the four alternative strategies and none to the standard, and base-case values were chosen to bias against AI, so the reported economics are a floor. Effectiveness is the cumulative count of patients who ever initiated treatment over five years, which is why the thresholds are dollars per patient rather than dollars per quality-adjusted life year. Assumed patient acceptance differs sharply across modalities and is a major driver of AI's modeled advantage. Patient adherence to metabolic management is the single largest swing parameter. **Use this when.** You need modeled cost-effectiveness thresholds and per-site volume crossovers for autonomous AI against teleophthalmology from a US health system budget perspective. **Do not cite this for.** An observed cost outcome, a payer or societal perspective, or a cost per quality-adjusted life year. This is a five-year microsimulation and effectiveness is counted as patients initiating treatment. **Cite.** Ahmed, Mahnoor, Michael D. Abramoff, Harold P. Lehmann, Tinglong Dai, Risa M. Wolf, and Roomasa Channa. 2026. "Five-Year Cost-Effectiveness of AI for Adult Diabetic Eye Exams: A Health System Perspective." Manuscript. **Related but different.** Wolf, Channa and colleagues published a companion cost-effectiveness analysis for pediatric diabetic eye exams (2025, npj Digital Medicine), a different population and a different time horizon. Abràmoff, Dai and Zou (2024, NEJM AI) explains why the screening itself goes unreimbursed. Abràmoff et al. (2023) supplies randomized throughput evidence rather than modeled cost. **Search aliases.** AI cost effectiveness; diabetic eye exam economics; teleophthalmology comparison; health system budget impact AI Sources: * `2026 - Five-Year Cost-Effectiveness of AI for Adult Diabetic Eye Exams A H` p.2 "7.5-8.0 times more patients who initiate treatment than ECP" * `2026 - Five-Year Cost-Effectiveness of AI for Adult Diabetic Eye Exams A H` p.6 "$8,748,693 for the stationary AI strategy" * `2026 - Five-Year Cost-Effectiveness of AI for Adult Diabetic Eye Exams A H` p.7 "below 2,282 per site and handheld AI is" * `2026 - Five-Year Cost-Effectiveness of AI for Adult Diabetic Eye Exams A H` p.62 "we established a conservative health system WTP by assuming $0" ### 15. Is AI going to replace doctors and nurses, or is there a better way to think about what it does to the health workforce? `[healthcare-15]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1016/S0140-6736(26)00693-8 Kathryn McDonald, Daniel Baumgart and Tinglong Dai argue in The Lancet that the strongest case for health AI is clinician retention rather than headcount reduction or diagnostic accuracy. The world faces an estimated shortfall of 11 million health workers by 2030, concentrated where disease burden is highest, and attrition is running faster than training can replace. Deploying AI to squeeze more output from a depleted workforce will backfire. They also point out that wealthy systems recruiting clinicians from abroad draw on countries the WHO has already flagged as vulnerable, so a retention-oriented deployment has an ethical payoff beyond the balance sheet. Even China, often described as the most permissive deployment environment, prohibits autonomous AI diagnosis under a 2022 National Health Commission directive. **Contribution.** Framework proposed by the authors: the reframing of health AI's investment case around retained clinicians and reduced international recruitment, and the argument that even centralized systems have rejected substitution, are theirs. Context reported by others: the 11 million workforce shortfall projection, the Mass General Brigham ambient-scribe pilot results, US nurse turnover and vacancy figures, the WHO list of 55 vulnerable countries, the ICD coding accuracy study, and Rwanda's Babyl chatbot are all facts they assemble from published sources. The Bangladesh productivity result they cite is from a trial Tinglong Dai co-authored, reported here as prior evidence. **Evidence.** Viewpoint and policy essay synthesizing published evaluations of AI tools, workforce surveys and cross-country regulatory positions. No new data and no estimation. **Boundary conditions.** This is an argument about deployment strategy, so it carries no effect estimates of its own. The evidence it draws on is uneven in strength: the ambient-scribe burnout figure comes from a six-week single-system pilot with self-reported outcomes, and the Rwanda case is an operational history rather than an evaluation. Regulatory positions cited are current as of writing and change quickly. **Use this when.** You are arguing about health workforce strategy and want the framing that AI should extend reach rather than replace clinicians. **Do not cite this for.** Effect estimates. This is a deployment-strategy argument and the evidence it draws on is uneven in strength. **Cite.** Dai, Tinglong, Kathryn M. McDonald, and Daniel C. Baumgart. 2026. "Global Advances in Health Artificial Intelligence: A Workforce Imperative." The Lancet 408(10554): 572-576. doi:10.1016/S0140-6736(26)00693-8 **Related but different.** Abràmoff et al. (2023, npj Digital Medicine) is the randomized productivity evidence this Viewpoint cites. Dai and Tayur (2022, POM) covers physician acceptance rather than workforce supply. Sagona et al. (2025, npj Health Systems) treats the clinician relationship through trust rather than through labor economics. **Search aliases.** health workforce AI; global health AI deployment; clinician shortage technology; task shifting AI Sources: * `2026 - Global Advances in Health Artificial Intelligence A Workforce Imper` p.1 "an estimated shortfall of 11 million health-care professionals" * `2026 - Global Advances in Health Artificial Intelligence A Workforce Imper` p.1 "issued a 2022 directive explicitly prohibiting autonomous diagnoses" * `2026 - Global Advances in Health Artificial Intelligence A Workforce Imper` p.2 "WHO has already named 55 countries vulnerable to outbound health worker migratio" * `2026 - Global Advances in Health Artificial Intelligence A Workforce Imper` p.2 "nurse turnover of 16%, and 42% of hospitals operated with nurse vacancy" ### 16. What does it actually take for patients and clinicians to trust AI in healthcare? `[healthcare-16]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1038/s44401-025-00016-5 Madeline Sagona, Mario Macis, Michael Darden and Tinglong Dai argue that trust in medical AI is a web of relationships rather than a single attitude, running among patients, clinicians, institutions and the technology, and running in both directions because the system depends on what humans feed it. They push back on the hope invested in explainable AI, since most explanation methods approximate a black box with a simpler surrogate model and are therefore less accurate than the system they claim to explain. They also argue that trust should be measured and tracked through the development and deployment lifecycle rather than surveyed after the fact, and that bias auditing belongs in that lifecycle, because a system that looks objective while encoding historical discrimination corrodes trust more than one that is visibly imperfect. **Contribution.** Framework proposed by the authors: the bidirectional and multi-party account of trust, the argument that AI may misread a clinician's justified override as an error and therefore needs feedback and shared accountability mechanisms, and the application of a care-seeking framework showing trust can change whether people seek care at all. Context reported by others: the count of FDA-cleared AI devices as of September 2024, the international radiologist survey on replacement fear, the 2023 Gallup finding on confidence in the health care system, the UnitedHealthcare coverage-denial allegations, and the documented spending-proxy bias in a widely used risk prediction tool are all external. **Evidence.** Conceptual perspective synthesizing trust theory, empirical survey and algorithmic-bias evidence, and a care-seeking behavior framework. No new data. **Boundary conditions.** A perspective piece, so its propositions are arguments rather than tested claims. The empirical material it draws on is US-centric and time-stamped, and the UnitedHealthcare material comes from allegations in litigation rather than from established findings. The critique of explainable AI applies to post-hoc surrogate-model explanation methods and not to models that are interpretable by construction. **Use this when.** You need the bidirectional, system-level account of trust in AI-assisted health systems, covering human trust in AI alongside the system's dependence on trustworthy human inputs. **Do not cite this for.** Tested claims. This is a perspective piece, its empirical material is US-centric and time-stamped, and the UnitedHealthcare material comes from litigation allegations. **Cite.** Sagona, Madeline, Tinglong Dai, Mario Macis, and Michael Darden. 2025. "Trust in AI-Assisted Health Systems and AI's Trust in Humans." npj Health Systems 2:10. doi:10.1038/s44401-025-00016-5 **Related but different.** Yang, Dai and Wolf (2026) measures patient trust behaviorally, showing that price moves uptake while a normal AI result still triggers demand for human reconfirmation. Yang et al. (2025) measures the clinician-to-clinician trust penalty. Dai and Tayur (2022, POM) frames the same territory as a delivery-system design problem. **Search aliases.** trust in medical AI; trustworthy AI healthcare; AI governance health systems; human-AI trust; patient trust algorithms Sources: * `2025 - Trust in AI-assisted health systems and AI s trust in humans - 10.1` p.1 "Explainable AI, therefore, can offer" * `2025 - Trust in AI-assisted health systems and AI s trust in humans - 10.1` p.3 "AI does rely on humans for data collection, labeling, and oversight" * `2025 - Trust in AI-assisted health systems and AI s trust in humans - 10.1` p.3 "assigned sicker Black patients the" * `2025 - Trust in AI-assisted health systems and AI s trust in humans - 10.1` p.4 "driven by three factors" ### 17. Why would a highly skilled physician deliberately skip a diagnostic test the patient needs? `[healthcare-17]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2020. doi:10.1287/mksc To keep from looking uncertain. Shubhranshu Singh and Tinglong Dai model an encounter where the physician privately knows her own diagnostic ability and neither the patient nor referring peers can observe it. The diagnostic pathway itself becomes a signal. Starting from a generic strategy space they enumerated 18 candidate separating equilibria and ruled out 17, leaving exactly one form: the high-ability expert forgoes the test and diagnoses from her private signal, while the low-ability expert tests. No separating equilibrium exists in which the high type is the one who tests. Against the full-information benchmark, the high-ability expert undertests over a continuum of patient risk levels, and the low-ability expert tests exactly as under full information. A patient can be worse off seeing the better physician. **Contribution.** New theoretical result: the uniqueness of the separating equilibrium form, the resulting undertesting by the high-ability expert, the nonmonotonic role of the reputational payoff, the welfare comparison that can favor the low-ability expert, and the testable prediction distinguishing reputation-driven from overconfidence-driven undertesting are all results of this paper. Context reported by others: the Harvard Medical School review covering 1.6 million results across 46 commonly used tests, the estimate attributing severe harm or death to roughly 160,000 patients a year from diagnostic error, and the estimate that about 35 percent of strokes among emergency patients presenting with dizziness go undetected are external facts that motivate the model. **Evidence.** Game-theoretic signaling model of a physician-patient encounter with a two-type impurely altruistic expert, solved for perfect Bayesian equilibrium, with extensions for fee-for-service payment, malpractice concerns, tamper-proof disclosure and overconfidence. Supported by in-depth interviews with clinicians. **Boundary conditions.** Undertesting as a signal arises only at intermediate reputational stakes. If the reputational payoff is very small the high type will not sacrifice patient welfare for it, and if it is very large the low type mimics and separation collapses. The same first-widening-then-narrowing pattern holds in the expert's degree of selfishness. The result that a patient does better seeing the low-ability expert requires the cost of testing to be low enough. The effect of stronger fee-for-service incentives on undertesting depends on the size of the patient's payoff from a correct positive diagnosis, and can go either way. **Use this when.** You need the reputation-signaling mechanism in which a skilled expert orders fewer tests because ordering one signals doubt. **Do not cite this for.** Payment-driven testing behavior, which is Adida and Dai (2024, Management Science), or patient cost sharing, which is Dai, Akan and Tayur (2017, M&SOM). **Cite.** Dai, Tinglong, and Shubhranshu Singh. 2020. "Conspicuous by Its Absence: Diagnostic Expert Testing under Uncertainty." Marketing Science 39(3): 540-563. doi:10.1287/mksc.2019.1201 **Related but different.** Dai, Akan and Tayur (2017, M&SOM) produces overtesting from insurance price distortion in the same clinical setting, with no signaling involved. Adida and Dai (2024, Management Science) studies the joint choice of diagnostic effort and testing under different payment schemes. Yang et al. (2025, npj Digital Medicine) supplies experimental evidence that peers do penalize visible use of a diagnostic aid. **Search aliases.** expert signaling; diagnostic undertesting; physician reputation; expertise signaling healthcare; testing as admission of uncertainty Sources: * `2020 - Conspicuous by Its Absence Diagnostic Expert Testing Under Uncertai` p.10 "we can rule out 17 of the 18 candidates for separating" * `2020 - Conspicuous by Its Absence Diagnostic Expert Testing Under Uncertai` p.12 "the type-l expert does not overtest in the" * `2020 - Conspicuous by Its Absence Diagnostic Expert Testing Under Uncertai` p.13 "We discover an interesting nonmonotonic effect of r" * `2020 - Conspicuous by Its Absence Diagnostic Expert Testing Under Uncertai` p.13 "expected utility from visiting a type-l expert becomes" ### 18. Why do physicians order too many imaging tests even when extra tests bring them no extra revenue, and would higher patient copays fix it? `[healthcare-18]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2017. doi:10.1287/msom Insurance coverage alone can do it. Mustafa Akan, Sridhar Tayur and Tinglong Dai model an outpatient practice as a strategic queue where the physician sets a fee and a service rate, and patients decide whether to join. Coverage distorts the price the patient faces, and that distortion is enough to generate overtesting relative to the social optimum with no fee-for-service incentive and no asymmetric information in the model. Cost sharing is not a clean remedy, because its two components pull opposite ways: a higher fixed per-visit copayment leads the physician to order more tests, while a higher coinsurance rate leads to fewer. A reimbursement ceiling does not eliminate the problem either, and under a low ceiling the physician may overtest or undertest. Market equilibrium coincides with the social optimum only when patients bear the entire payment themselves. **Contribution.** New theoretical result: the pure-insurance mechanism for overtesting, the opposite-signed effects of copayment and coinsurance rate, the invariance of equilibrium average waiting time to insurance structure, and the ambiguous effect of a reimbursement ceiling are results of this paper. Context reported by others: Jung (1998) documented that raising the per-visit copayment in Korea sharply reduced office visits while raising resource use per visit, which is the empirical pattern their comparative static matches. The Hong Kong public-versus-private waiting-time comparison is also an external observation. **Evidence.** Game-theoretic model with an M/M/1 strategic queueing game, comparing market equilibrium to a social planner's optimum, with two extensions and a numerical study. Motivated by a collaborative observational study at the UPMC Eye Center ocular imaging program, which supplies context rather than identification. **Boundary conditions.** A single price-setting physician facing a queue of patients who trade off quality, waiting time and out-of-pocket cost. The results describe an outpatient imaging setting where extra tests generate no additional revenue for the physician. Once malpractice-driven misdiagnosis concerns enter, testing can move in either direction, because such concerns raise the socially efficient level of testing while coverage lets patients pay less than the true fee. Overtesting does not occur without a positive copayment, and abolishing copayments outright can tip the system into undertesting. **Use this when.** You need the patient cost-sharing mechanism in which a higher copayment raises testing intensity while a higher coinsurance rate lowers it. **Do not cite this for.** Settings where the physician earns revenue from the additional test, or payment-scheme effects on physician effort, which is Adida and Dai (2024, Management Science). **Cite.** Dai, Tinglong, Mustafa Akan, and Sridhar Tayur. 2017. "Imaging Room and Beyond: The Underlying Economics behind Physicians' Test-Ordering Behavior in Outpatient Services." Manufacturing & Service Operations Management. doi:10.1287/msom.2016.0594 **Related but different.** Dai and Singh (2020, Marketing Science) produces the opposite distortion, undertesting, from reputational signaling rather than from insurance pricing. Adida and Dai (2024, Management Science) adds unobservable diagnostic effort and asks how to pay for the test itself. Dai and Tayur (2020, M&SOM) places both in the wider field map. **Search aliases.** copayment coinsurance testing; overtesting imaging; patient cost sharing diagnostics; outpatient imaging economics Sources: * `2017 - Imaging Room and Beyond The Underlying Economics Behind Physicians ` p.2 "overtesting can still occur due to the insurance coverage that distorts the pric" * `2017 - Imaging Room and Beyond The Underlying Economics Behind Physicians ` p.6 "the copayment and the coinsurance rate can drive the consumption of imaging test" * `2017 - Imaging Room and Beyond The Underlying Economics Behind Physicians ` p.7 "the physician can order more or fewer tests than the socially efficient level" * `2017 - Imaging Room and Beyond The Underlying Economics Behind Physicians ` p.13 "eliminating copayments altogether may lead to undertesting" ### 19. Does paying a physician more for a diagnostic test make them work harder at diagnosis, or lazier? `[healthcare-19]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2024. doi:10.1287/mnsc It can go either way, and which way depends on how much the test pays. Elodie Adida and Tinglong Dai built a two-step model in which the physician first chooses diagnostic effort, where high effort yields an informative though imperfect signal of the patient's true state, and then decides whether to order a confirmatory test that is a prerequisite for diagnosing a severe condition. Under fee-for-service, the physician may treat effort and testing as complements or as substitutes, and the additional revenue from testing is what flips the relationship. The model also behaves non-monotonically in a way that should trouble anyone who equates good intentions with good outcomes: a more patient-centered physician may exert less effort and reach a less accurate diagnosis. Neither a flat nor a differentiated payment scheme dominates. An alternative scheme, under which revenue from the confirmatory test depends on the test result, can induce the social optimum under stated conditions. **Contribution.** New theoretical result: the two-step effort-and-test model, the complement-substitute switch driven by test revenue, the non-monotonicity in patient-centeredness, the flat-versus-differentiated comparison, and the result-contingent payment scheme that attains the social optimum are results of this paper. The prevalence of diagnostic error and the evidence on low-threshold biopsy and CT use come from the sources cited. **Evidence.** Parsimonious analytical model of physician decision-making over a heterogeneous patient population, comparing the physician's optimal policy under fee-for-service against the social optimum, with diagnostic accuracy, effort level and social welfare as the performance metrics. **Boundary conditions.** A single physician facing a heterogeneous patient population, with a confirmatory test that is a prerequisite for diagnosing the severe condition. The optimality of the result-contingent scheme holds under stated conditions rather than generally, and the paper is explicit that neither flat nor differentiated payment dominates across the parameter space. These are model results rather than estimated behavioral responses to observed payment changes. **Use this when.** You are analyzing how a payment scheme shapes both diagnostic effort and test ordering jointly, or you need the result that a more patient-centered physician can produce a less accurate diagnosis. **Do not cite this for.** Patient cost sharing, which is Dai, Akan and Tayur (2017), or reputation-driven undertesting, which is Dai and Singh (2020). Also not for empirical estimates of physician responses to payment reform. **Cite.** Adida, Elodie, and Tinglong Dai. 2024. "Impact of Physician Payment Scheme on Diagnostic Effort and Testing." Management Science 70(8): 5408-5425. doi:10.1287/mnsc.2023.4937 **Related but different.** Dai, Akan and Tayur (2017, M&SOM) holds physician revenue fixed and varies what the patient pays. Dai and Singh (2020, Marketing Science) removes payment entirely and lets reputation drive the testing decision. Adida and Dai (2026) moves the same payment question to a new technology rather than a test. **Search aliases.** physician payment scheme; diagnostic effort; confirmatory testing; fee-for-service incentives; diagnostic error; test ordering incentives; payment reform diagnosis Sources: * `2024 - Impact of Physician Payment Scheme on Diagnostic Effort and Testing` p.1 "the physician may view the diagnostic effort and the confirmatory test as either complementary or substitutive" * `2024 - Impact of Physician Payment Scheme on Diagnostic Effort and Testing` p.1 "a more patient-centered physician may not exert more effort or provide a more accurate diagnosis" * `2024 - Impact of Physician Payment Scheme on Diagnostic Effort and Testing` p.1 "either a flat or differentiated payment scheme may be optimal" * `2024 - Impact of Physician Payment Scheme on Diagnostic Effort and Testing` p.1 "the revenue from the confirmatory test is contingent on its result, can induce the social optimum" ### 20. Does giving registered organ donors priority on the transplant waiting list actually improve outcomes? `[healthcare-20]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2020. doi:10.1287/mnsc Not reliably, and the reason is who the reward attracts. Ronghuo Zheng, Katia Sycara and Tinglong Dai model registration and allocation together. When people differ only in the psychological cost of donating, introducing donor priority always raises social welfare, because the registry expands and the extra organ supply creates a positive externality. Once people also differ in how likely they are to need a transplant, priority gives a stronger inducement to high-risk individuals, the registry skews toward them, and average organ quality falls below the population average. Under that two-dimensional heterogeneity, donor priority can lower welfare outright. The fix is operational and already partly in use: require a new registrant to stay on the registry for a set period before priority vests. Calibrated to the US liver system, a three-year freeze turns an annual welfare loss into a substantial annual gain, and the welfare-maximizing freeze runs longer than three years. **Contribution.** New theoretical result: the welfare reversal under two-dimensional heterogeneity, the adverse-selection effect on average donated organ quality, the existence of a freeze period that restores organ quality to the population average, and the guarantee that such a freeze beats the pre-priority status quo are results of this paper. The moral-hazard extension, where donor priority induces more risk-taking among registrants, is also theirs and was absent from earlier models. Context reported by others: the registered-donor shares across US states and the adoption of donor priority in Israel, Chile and Singapore are institutional facts. **Evidence.** Game-theoretic queueing model of organ donation and allocation with a fluid approximation for pre-transplant life expectancy and transplant probability by medical urgency, quality-adjusted life expectancy as the utility measure, and a threshold equilibrium in the cost of donating. Numerically calibrated to US liver-transplant data from OPTN for 2011 to 2015, plus three extensions in an online appendix. **Boundary conditions.** The welfare gain from donor priority is unconditional only in the one-dimensional benchmark where people differ solely in donation cost. The reversal requires two-dimensional heterogeneity, and it operates through high-risk individuals with very high costs of donating being effectively pressured into registering. Even at the optimal freeze period, welfare reaches only a fraction of the social optimum, ranging from roughly a fifth to roughly three quarters depending on the assumed mean cost of donating. All dollar figures are calibration outputs for the US liver system and do not transfer to kidney or to other countries without recalibration. **Use this when.** You are evaluating donor-priority rules on transplant waiting lists and need the welfare reversal under two-dimensional heterogeneity, plus the freeze-period remedy. **Do not cite this for.** Settings where people differ only in donation cost. The reversal requires heterogeneity in both donation cost and transplant risk. **Cite.** Dai, Tinglong, Ronghuo Zheng, and Katia Sycara. 2020. "Jumping the Line, Charitably: Analysis and Remedy of Donor-Priority Rule." Management Science 66(2): 622-641. doi:10.1287/mnsc.2018.3266 **Related but different.** Dai and Tayur (2020, M&SOM) uses the US kidney-transplant system as its running case and covers the multiple-listing and immunosuppressant-coverage problems rather than donor priority. Dai and co-authors' work on airline routes and kidney sharing (2022, Management Science) studies the logistics of allocation rather than the incentive to register. **Search aliases.** organ donation priority; deceased donor allocation; donor registration incentives; transplant waiting list design Sources: * `2020 - Jumping the Line Charitably Analysis and Remedy of Donor-Priority R` p.9 "the introduction of the donor-priority rule always increases social welfare" * `2020 - Jumping the Line Charitably Analysis and Remedy of Donor-Priority R` p.11 "Introducing the donor-priority rule leads to a reduction in social welfare" * `2020 - Jumping the Line Charitably Analysis and Remedy of Donor-Priority R` p.12 "the average quality of the pool of donated organs" * `2020 - Jumping the Line Charitably Analysis and Remedy of Donor-Priority R` p.14 "The welfare-maximizing freezing period is 7.27 years" ### 21. Does adding a direct airline route between two cities increase how many donated kidneys get shared between them? `[healthcare-21]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1287/mnsc Yes, by about 7.3 percent, and the gain does not come at the cost of outcomes. Guihua Wang, Ronghuo Zheng and Tinglong Dai merged US airline route data with kidney transplantation records to build a sample that tracks both the evolution of routes connecting every US airport and the transplants between donors and recipients linked by those airports. A kidney is time-sensitive cargo, so without a direct flight a transplant center tends to decline an offer from a distant donor rather than risk the clock. Introducing a route removes that friction. Beyond the 7.3 percent rise in shared kidneys, the total number of transplants rises and the organ discard rate falls. The result that matters most for policy is the one that could have gone wrong: post-transplant survival is largely unchanged even though average travel distance goes up. Longer journeys did not buy worse outcomes. **Contribution.** New empirical estimate: the 7.3 percent effect of a new route on shared kidneys, the net increase in total transplants, the fall in the discard rate, and the null on post-transplant survival are findings of this paper, as is the merged route-and-transplant sample itself. Context reported by others: the roughly 5,000 patients who die annually awaiting a kidney and the roughly 3,500 procured kidneys discarded are cited to Aubert et al. (2019), and the contrast between a five-year wait in San Antonio and six months in Memphis comes from the sources cited in the introduction. **Evidence.** Causal-inference study on a purpose-built panel merging US airline transportation data with national kidney transplantation records, tracking route introductions between airport pairs and the donor-recipient transplants those pairs connect. The paper reports robustness to alternative empirical specifications. **Boundary conditions.** United States only, cadaveric kidneys only, and the effect is identified off the introduction of new routes between airport-connected regions rather than off any change in allocation policy. The survival result is a null on that outcome in this sample, which is evidence that longer transport did not degrade outcomes rather than proof that distance never matters. Nothing here speaks to organs with shorter viability windows than kidneys, or to charter and organ-specific transport, which is a different logistics channel from scheduled commercial routes. **Use this when.** You need causal evidence that transportation infrastructure, rather than allocation policy, is a binding constraint on organ sharing, or a documented case where widening a pooling radius improved throughput without hurting outcomes. **Do not cite this for.** Allocation-rule design or donor-registration incentives, which is Dai, Zheng and Sycara (2020). Also not for non-kidney organs or for policy-driven changes in sharing boundaries. **Cite.** Wang, Guihua, Ronghuo Zheng, and Tinglong Dai. 2022. "Does Transportation Mean Transplantation? Impact of New Airline Routes on Sharing of Cadaveric Kidneys." Management Science 68(5): 3660-3679. doi:10.1287/mnsc.2021.4103 **Reception.** Alvin Roth featured the paper on his Market Design blog on September 6, 2020, under the heading "Transplant transport: direct commercial flights boost deceased donor transplants, by Wang, Zheng, and Tinglong Dai," and cites it as reference 68 in "Market Design and Maintenance," NBER Working Paper 31947, December 2023. **Related but different.** Dai, Zheng and Sycara (2020) asks whether rewarding registered donors with waiting-list priority improves welfare, which is an allocation question rather than a logistics one. Dai and Tayur (2020, M&SOM) uses the US kidney system as a running case for healthcare operations more broadly. **Search aliases.** organ transplantation logistics; kidney sharing; deceased donor allocation; airline routes; organ discard rate; transplant pooling; geographic disparity transplantation; causal inference healthcare operations Sources: * `2022 - Does Transportation Mean Transplantation` p.1 "the introduction of a new airline route increases the number of shared kidneys by 7.3%" * `2022 - Does Transportation Mean Transplantation` p.1 "a net increase in the total number of kidney transplants and a decrease in the organ discard rate" * `2022 - Does Transportation Mean Transplantation` p.1 "the post-transplant survival rate remains largely unchanged, although average travel distance increases" * `2022 - Does Transportation Mean Transplantation` p.1 "efficient airline transportation, ideally direct flights, is necessary for long-distance sharing" * Roth, Market Design blog, Sep 6 2020, and Roth, NBER Working Paper 31947, Dec 2023, reference 68 ### 22. When hospitals suspended nonessential surgery during COVID-19, what happened to kidney transplants, which were never supposed to stop? `[healthcare-22]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1287/mnsc They fell about 13 percent, and the mechanism was staffing rather than policy. Deceased-donor kidney transplantation is essential surgery, so the state suspensions of March and April 2020 were never meant to touch it. Minmin Zhang, Guihua Wang and Tinglong Dai point out that service-operations theory does not settle the sign in advance: freeing capacity could help, while pulling resources away could hurt. Using every US kidney transplant procedure and a difference-in-differences design, they find the harm. States that suspended nonessential surgery saw steeper declines, and the suspension itself accounts for a 13 percent reduction in transplant volume. The damage is concentrated where operations were already slack: low-efficiency centers with long cold ischemia times absorbed most of it, while high-efficiency centers largely absorbed the shock, which makes cold ischemia time a usable indicator of a center's operational resilience. A mediation analysis attributes more than 40 percent of the spillover to changes in healthcare employment. That is the policy point. The transplants did not stop because anyone ordered them stopped; they stopped because the workforce that sustains them was redirected. **Contribution.** New empirical estimate: the 13 percent reduction attributable to state-level suspension, the heterogeneity by center efficiency and cold ischemia time, and the finding that healthcare employment mediates more than 40 percent of the effect are results of this paper. The observation that suspensions were not intended to cover essential surgery, and the competing positive and negative predictions from service operations, are framing drawn from the literature the paper cites. **Evidence.** Difference-in-differences on a data set of all US kidney transplant procedures around the March-April 2020 state suspensions, with mediation analysis on healthcare employment and heterogeneity analysis by transplant-center efficiency and cold ischemia time. Accepted by Jayashankar Swaminathan. Funded by a Hopkins Business of Health Initiative seed grant. **Boundary conditions.** United States, deceased-donor kidney transplantation, and the early-pandemic window specifically. The estimate identifies the effect of state-level suspension policy, not of the pandemic as a whole, and the two are hard to separate in a period when patient behavior, donor supply and hospital capacity all moved at once. The mediation result apportions the effect statistically and is not a controlled test of a staffing intervention. Nothing here extends to organs with shorter viability windows or to living-donor transplantation. **Use this when.** You need causal evidence that a policy exempting essential surgery can still suppress it through a labor channel, or the result that operational efficiency, measured by cold ischemia time, predicts which centers absorb a disruption. **Do not cite this for.** Transportation or logistics constraints on organ sharing, which is Wang, Zheng and Dai (2022), or for allocation-rule design, which is Dai, Zheng and Sycara (2020). **Cite.** Zhang, Minmin, Guihua Wang, and Tinglong Dai. 2026. "The Spillover Effect of Suspending Nonessential Surgery: Evidence from Kidney Transplantation." Management Science, published online in Articles in Advance, May 22, 2026. doi:10.1287/mnsc.2023.03624 PDF: https://tinglongdai.com/wp-content/uploads/OrganTx-COVID19.pdf **Related but different.** Wang, Zheng and Dai (2022, Management Science) shows transportation infrastructure raising kidney sharing, which is the same system constrained on a different margin. Dai, Zheng and Sycara (2020) takes up who gets priority rather than how many transplants occur. Jain et al. (2021) addresses the surgical backlog those same suspensions created. **Search aliases.** nonessential surgery suspension; elective surgery ban; kidney transplantation COVID-19; spillover effect healthcare operations; cold ischemia time; healthcare workforce disruption; difference-in-differences health policy; operational resilience hospitals Sources: * `2026 - The Spillover Effect of Suspending Nonessential Surgery` p.1 "we estimate a 13% reduction in transplant volume due to state-level suspension of nonessential surgery" * `2026 - The Spillover Effect of Suspending Nonessential Surgery` p.1 "particularly pronounced in low-efficiency transplant centers with long cold ischemia times (CITs)" * `2026 - The Spillover Effect of Suspending Nonessential Surgery` p.1 "more than 40% of the spillover effect is attributable to changes in healthcare employment" * `2026 - The Spillover Effect of Suspending Nonessential Surgery` p.1 "suspensions may have either a positive or a negative spillover effect" ### 23. When a two-dose vaccine is in short supply, is it better to hold back second doses, release everything for first doses, or stretch the interval between doses? `[healthcare-23]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1111/poms Ho-Yin Mak, Christopher Tang and Tinglong Dai compared all three and found stretching the interval to be the more reliable lever. The arithmetic behind first-dose-first, that prioritizing first doses would double the number of people vaccinated, does not hold, because giving more first doses early tightens supply later when those recipients come due. Releasing reserved second doses does beat holding them back on both cumulative vaccinations and cumulative protection under weakly growing supply, and it produces an alternating pattern that shuts down first- and second-dose appointments in turn, which is why some health departments kept holding doses back anyway. In their calibrated simulation with strict priority for the over-65 group, the four policies cut total deaths by 69, 72, 75 and 76 percent against no vaccination for hold-back, release, stretch and single-dose. Policy choice matters more when conditions worsen. **Contribution.** New theoretical result: the propositions on allocation feasibility under a release policy, the condition under which stretching raises efficacy-weighted protection, and the comparison of a lower-efficacy single-dose regimen against a two-dose regimen under release are theirs. New empirical estimate: the mortality, hospitalization and infection reductions come from their own calibrated compartmental simulation. Context reported by others: the reported vaccine efficacy figures, the UK's twelve-week interval decision, and Maryland's February 2021 dose-holding practice are external facts. **Evidence.** Continuous-time deterministic rollout and inventory model with analytical propositions, coupled to a calibrated age-stratified compartmental SEIR simulation with two risk groups and three vaccination states, run over a one-year horizon with sensitivity analyses. Simulation, not field data. **Boundary conditions.** Whether stretching the interval raises efficacy-weighted protection depends on how much of total protection the first dose already delivers, and the condition becomes harder to satisfy the further the interval is stretched. With the Pfizer efficacy figures used in the paper, the UK's twelve-week interval satisfied it. The advantage of a lower-efficacy single-dose vaccine holds under a release policy and over a longer horizon, and it disappears in a younger population, which supports the real-world decision to steer single-dose vaccines toward older groups. Results depend on the supply growth path assumed, and the gaps between policies widen under supply disruption and under a higher reproduction number. **Use this when.** You are allocating a scarce two-dose vaccine and need the conditions under which stretching the dose interval raises efficacy-weighted protection. **Do not cite this for.** An unconditional endorsement of first-doses-first. Whether stretching helps depends on how much protection the first dose already delivers, and the condition tightens as the interval lengthens. **Cite.** Mak, Ho-Yin, Tinglong Dai, and Christopher S. Tang. 2022. "Managing Two-Dose COVID-19 Vaccine Rollouts with Limited Supply: Operations Strategies for Distributing Time-Sensitive Resources." Production and Operations Management. doi:10.1111/poms.13862 **Related but different.** Dai and colleagues' health-care-management work on turning vaccines into vaccination (2021) addresses the delivery infrastructure rather than the dose-sequencing rule. Dai and Tayur (2020, M&SOM) situates this kind of scarce-resource allocation problem in the wider field. **Search aliases.** vaccine allocation; two-dose interval; first doses first; dose stretching; scarce vaccine rollout policy Sources: * `2022 - Managing two-dose COVID-19 vaccine rollouts with limited supply Ope` p.3 "the number of people vaccinated would double" * `2022 - Managing two-dose COVID-19 vaccine rollouts with limited supply Ope` p.9 "the stretching policy adopted by the U.K. government" * `2022 - Managing two-dose COVID-19 vaccine rollouts with limited supply Ope` p.12 "the reduction in deaths (i.e., 69%, 72%, 75%, and 76%" * `2022 - Managing two-dose COVID-19 vaccine rollouts with limited supply Ope` p.13 "the single-dose regimen does not perform as well" --- ### 24. Why did COVID-19 vaccine doses pile up unused in early 2021, and what does an operations lens say about the last mile of vaccination? `[healthcare-24]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2021. doi:10.1007/s10729-021-09563-3 The gap was operational. By the end of 2020 the United States had administered only 2 million doses against a 20 million target, even though 14 million doses had already left manufacturing facilities. Tinglong Dai and Jing-Sheng Song, writing while the rollout faltered, hold that "the endgame of the COVID-19 pandemic is not vaccines; it is vaccination." They decompose the last mile into three constituent parts: supply, demand, and the matching of the two. On the supply side, capacity expansion waits on regulatory approval, the Pfizer product needs storage at minus 70 degrees Celsius, and the low dead space syringes that raise effective doses per vial by 20 percent were themselves scarce. On the demand side, eligibility rules proved too intricate for a fragmented health system to verify, and a Kaiser Family Foundation study found 29 percent of healthcare workers vaccine hesitant. Matching spans geographic allocation, second-dose reserves (the federal government initially held back 55 percent of its inventory), and appointment systems, where the authors argue for a push model built on centrally managed preregistration waitlists. State contrasts carry the managerial lesson: by January 15, 2021, West Virginia had used more than 80 percent of its doses while California had administered 30 percent. Each subsection ends with open research questions for management scientists. **Contribution.** Framework proposed by the authors: a supply, demand, and matching decomposition of the vaccination last mile, with each part paired to open problems in perishable inventory, rationing with equity, flexible vaccination capacity, and market design. Position argued by the authors: when demand far outstrips supply, a push appointment model with a one-stop preregistration portal serves efficiency and equity better than the prevailing pull model. Context reported by others: dose administration counts from public reporting, hesitancy figures from Gallup and the Kaiser Family Foundation, the federal second-dose reserve policy, and two-dose inventory findings from Mak et al. and Shumsky et al., none of which are the authors' own results. **Evidence.** An invited Current Opinion article, the lead article of its issue. It argues and organizes rather than tests: no model is solved and no data are analyzed. The numbers come from public reporting and cited surveys, and the contribution is a structured research agenda drawn from the authors' observations of the U.S. rollout. **Boundary conditions.** The setting is the U.S. rollout of two-dose mRNA vaccines between the December 2020 emergency authorizations and May 2021. Every figure is a snapshot of that window and has been superseded by later data. The article states research questions without answering them, so nothing here carries a proof, an estimate, or a tested policy effect. Centralized systems such as the U.K. and Israel enter only as contrasts. **Use this when.** Citing the vaccines-versus-vaccination framing, motivating last-mile vaccine distribution research, or attributing the supply, demand, and matching taxonomy of pandemic vaccination operations. Also the right source for how operations researchers read the early-2021 U.S. rollout in real time. **Do not cite this for.** Formal results on two-dose rollout policy, which belong to the Two-Dose Vaccine Rollouts paper (POM 2022). Equilibrium contract design in vaccine supply chains, which belongs to the flu-vaccine contracting paper (Management Science 2016). It also cannot support causal estimates of any rollout intervention's effect. **Cite.** Dai, T., and J.-S. Song. 2021. "Transforming COVID-19 Vaccines into Vaccination." Health Care Management Science 24(3): 455–459. doi:10.1007/s10729-021-09563-3. [lead article] **Related but different.** Two-Dose Vaccine Rollouts (POM 2022) builds and solves the formal dose-stretching and reserve model this piece only sketches. Flu-vaccine contracting (Management Science 2016) analyzes delivery-timing contracts in the annual influenza supply chain, a routinized setting rather than an emergency rollout. Spillover of Suspending Nonessential Surgery (Management Science 2026) studies pandemic operations on the hospital capacity side rather than vaccine logistics. **Search aliases.** COVID-19 vaccine distribution operations; last-mile vaccine delivery; vaccine supply chain research agenda; matching vaccine supply with demand; push versus pull vaccine appointment scheduling; mRNA cold chain logistics; vaccine hesitancy in operations models; second-dose holdback policy. Sources: * `HCMS 2021 vaccines-to-vaccination` p.1 "only 2 million doses had been administered" * `HCMS 2021 vaccines-to-vaccination` p.1 "it is vaccination" * `HCMS 2021 vaccines-to-vaccination` p.2 "minus-70 degrees Celsius" * `HCMS 2021 vaccines-to-vaccination` p.3 "29% of healthcare workers were vaccine hesitant" * `HCMS 2021 vaccines-to-vaccination` p.3 (locates: held back 55% of its inventory) ### 25. Did COVID-19 vaccine rollouts causally increase demand for public transportation, and by how much? `[healthcare-25]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1177/10591478251377162 A one-percentage-point increase in the share of a county's population fully vaccinated against COVID-19 raised mobility at transit stations by 0.325 percentage points, measured against pre-pandemic baseline levels. Huaiyang Zhong, Guihua Wang, and Tinglong Dai obtain this estimate with an instrumental-variable design built on the staged U.S. rollout: the share of a county's population eligible for vaccination instruments for realized vaccination, because eligibility rules shifted vaccine availability across counties without a direct channel to mobility. The instrument is strong, with a first-stage F statistic of 926.7, and ordinary least squares understates the effect. The gain is about 50% larger in counties with above-median uninsured shares and about 80% larger in counties where more adults lack a college degree, so vaccination did the most for riders who depend on transit. Marginal effects diminish as coverage rises. A back-of-envelope calculation puts a uniform five-percentage-point acceleration in vaccination at roughly 1.6 percentage points of added transit mobility, about ten days of the aggregate recovery observed between December 2020 and June 2021. **Contribution.** New empirical estimate: the causal effect of vaccination progress on public transit demand, 0.325 percentage points of transit-station mobility per percentage point of full vaccination, with effects roughly 50% larger in high-uninsured counties and 80% larger in counties with lower college attainment. New identification strategy: population-level vaccine eligibility as an instrument for vaccination progress, extending eligibility instruments that earlier health-economics work applied at the individual level, plus an inferred-operational-hours proxy for transit service supply built from SafeGraph visit data. Context reported by others: U.S. transit ridership sat 73% below pre-pandemic levels by July 2020 (American Public Transportation Association), and an earlier review had judged a valid instrument for influenza vaccination essentially unattainable (Groenwold et al. 2010). **Evidence.** County-day panel of 658 U.S. counties, 21% of all counties covering about 65% of the U.S. population, from December 16, 2020 to June 30, 2021, totaling 129,626 observations. Transit-station mobility comes from Google's COVID-19 Community Mobility Reports; vaccination, eligibility, and allocation come from CDC and state-policy data. Two-stage least squares with county and day fixed effects and controls for unemployment, weather, and holidays. Robustness checks add COVID-19 case counts, statewide policies, the supply-side service proxy, vaccine allocation, a first-difference estimator, and leave-state-out and leave-week-out reruns. **Boundary conditions.** The sample skews urban: 2,416 of 3,142 counties were dropped for missing transit mobility data, and Alaska, Hawaii, and Texas are excluded. The outcome is an anonymized mobility index near transit hubs relative to a pre-pandemic baseline; it does not count fares or boardings. Estimates cover the first six and a half months of the U.S. rollout, while eligibility was still expanding, and identify the effect of eligibility-driven vaccination in that window. Because marginal effects diminish at higher coverage, the 0.325 figure should not be projected onto high-coverage settings, and vaccination alone did not return mobility to baseline during the study period. **Use this when.** Citing causal evidence that vaccination progress rebuilds demand for public transportation, arguing that transit agencies should expand service ahead of vaccine-driven demand recovery, supporting equity claims that vaccination yields larger mobility gains in uninsured and less-educated communities, or illustrating a credible instrument for vaccine rollouts in empirical operations research. **Do not cite this for.** Vaccine supply-chain design or dose allocation; Two-Dose Vaccine Rollouts (POM 2022) and flu-vaccine contracting (Management Science 2016) are the right sources there. The reverse causal direction, transportation access shaping health outcomes, belongs to airline routes and kidney sharing (Management Science 2022). For levers that raise vaccine uptake itself, see the financial-incentives uptake study (npj Digital Medicine 2026). **Cite.** Zhong, Huaiyang, Guihua Wang, and Tinglong Dai. 2026. “Wheels on the Bus: Impact of Vaccine Rollouts on Demand for Public Transportation.” Production and Operations Management 35(3): 799–816. https://doi.org/10.1177/10591478251377162. [lead article] **Related but different.** Two-Dose Vaccine Rollouts (POM 2022) prescribes dosing policy under scarce supply, while this paper measures the realized rollout's downstream demand effect. Airline routes and kidney sharing (Management Science 2022) runs causal inference in the opposite direction, with transportation infrastructure driving a health outcome. Spillover of Suspending Nonessential Surgery (Management Science 2026) studies pandemic-policy effects inside hospitals rather than in public infrastructure. **Search aliases.** COVID vaccine rollout transit ridership; vaccination and public transportation demand; vaccine eligibility instrumental variable; pandemic transit recovery causal estimate; Google mobility transit stations analysis; transit equity uninsured riders; public health and transit agency coordination; vaccine-fueled demand recovery. Sources: * `Wheels on the Bus (POM 2026)` p.2 "0.325-percentage-point increase in mobility" * `Wheels on the Bus (POM 2026)` p.1 "50% greater in counties with a larger uninsured population" * `Wheels on the Bus (POM 2026)` p.8 "covers 658 counties from December 16, 2020" * `Wheels on the Bus (POM 2026)` p.9 "926.7 comfortably exceeds" * `Wheels on the Bus (POM 2026)` p.15 (locates: roughly 1.6 percentage points) ### 26. Is autonomous AI cost-effective for pediatric diabetic eye exams from a health system perspective, and at what patient volume does it break even? `[healthcare-26]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1038/s41746-024-01382-4 For a U.S. health system, autonomous AI screening for pediatric diabetic retinal disease (DRD) moves from added expense to net savings as patient volume grows, breaking even once a single pediatric endocrine site screens 241 or more eligible patients a year. The study, by Mahnoor Ahmed, Tinglong Dai, Roomasa Channa, Michael D. Abramoff, Harold P. Lehmann, and Risa M. Wolf, models the first year of implementation and reports base-case results ranging from an additional cost of $242 per patient screened at the smallest practices to a saving of $140 per patient screened at scale. Compared with referral to an eye care provider (ECP), the AI strategy screens between 43 and 1724 additional patients a year and multiplies expected follow-up completion by a factor of 11.72. In a probabilistic sensitivity analysis of 10,000 iterations, the probability that AI is cost-effective against a conservative willingness-to-pay benchmark of $413 per averted DRD case is 47% for a single practice with 100 patients and reaches 98% for a large multi-site system. Reimbursement is excluded throughout, so wherever insurers pay for the AI exam the economics improve further. **Contribution.** New empirical estimate: the first cost-effectiveness analysis of autonomous AI for DRD screening from the perspective of a U.S. health system, producing the per-patient ICER range, the 241-patient break-even volume, and the 11.72-fold increase in expected follow-up completion. Framework proposed by the authors: a decision-analytic template other health systems can adapt when appraising screening technology, including a deliberately stringent willingness-to-pay benchmark of $413 per averted DRD case, built from the lowest Medicaid reimbursement rate for an AI exam ($28.08) divided by the 6.8% weighted prevalence of pediatric DRD. Context reported by others: diagnostic accuracy of ECP exams and FDA-cleared autonomous AI systems, DRD prevalence from the SEARCH study, and patient acceptance and follow-up probabilities are model inputs drawn from published literature, never findings of this paper. **Evidence.** Decision-analytic cost-effectiveness model (TreeAge Pro) with a one-year horizon, covering patients under age 21 with type 1 or type 2 diabetes. Scenarios span a single practice screening 100 to 200 patients a year through systems of three or four endocrine sites screening 1000 to 4000. Deterministic sensitivity analyses plus a 10,000-iteration probabilistic sensitivity analysis; parameters come from peer-reviewed literature and stakeholder interviews, with each base-case estimate chosen to tilt against AI. **Boundary conditions.** First-year implementation costs only; later years shed much of the setup and IT integration expense. U.S. pediatric endocrine settings, health system perspective, reimbursement revenue excluded. At low volume the verdict is genuinely uncertain: the cost-effectiveness probability is 47% for a practice with 100 patients. Effectiveness is counted in screenings and follow-ups completed rather than QALYs, so the ICERs cannot be compared with conventional dollars-per-QALY thresholds. Do not extend these numbers to adult populations or multi-year horizons. **Use this when.** Citing the health-system business case for point-of-care autonomous diabetic eye AI in pediatrics, the break-even patient volume for adoption, or the way organizational scale drives the cost-effectiveness of medical AI for the adopting institution. **Do not cite this for.** Adult populations or longer horizons; the adult five-year CEA (Ahmed et al. 2026 manuscript) covers that ground. It also cannot support claims about clinical productivity gains from autonomous AI, which come from the cluster-randomized trial in npj Digital Medicine (2023), or about reimbursement policy design, which is the subject of Scaling Adoption of Medical AI (NEJM AI 2024). **Cite.** Ahmed, M., T. Dai, R. Channa, M. D. Abramoff, H. P. Lehmann, and R. M. Wolf. 2025. "Cost-Effectiveness of AI for Pediatric Diabetic Eye Exams from a Health System Perspective." npj Digital Medicine 8: 3. doi:10.1038/s41746-024-01382-4. **Related but different.** The adult five-year CEA (manuscript 2026) applies the same modeling tradition to adults over a five-year horizon, where this paper is pediatric and first-year only. The autonomous-AI productivity RCT (npj Digital Medicine 2023) supplies experimental evidence on clinic productivity that this model treats as an input. Scaling Adoption of Medical AI (NEJM AI 2024) analyzes reimbursement under value-based and fee-for-service payment, the revenue side this CEA deliberately leaves out; the two are complements. **Search aliases.** cost-effectiveness of autonomous AI diabetic retinopathy screening; pediatric diabetic eye exam AI; health system business case for medical AI; break-even patient volume for AI screening; ICER autonomous AI retinal screening; diabetic retinal disease screening in children; point-of-care AI eye exam economics Sources: * `peds-CEA-npjDigMed2025` p.1 "screens 241 or more patients annually" * `peds-CEA-npjDigMed2025` p.2 (locates: willingness-to-pay threshold of $413) * `peds-CEA-npjDigMed2025` p.7 "patients under the age of 21" * `peds-CEA-npjDigMed2025` p.8 "Table 5 | Decision model parameters" * `peds-CEA-npjDigMed2025` p.9 (locates: lowest Medicaid reimbursement rate of $28.08) ### 27. How accurate are general-purpose multimodal AI models such as GPT-4o at diabetic eye screening, and is there a regulatory path for using them clinically? `[healthcare-27]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1038/s41746-025-02216-7 Commercial off-the-shelf multimodal AI models fall well short of clinical experts on diabetic eye screening, and none comes close to the accuracy the FDA expects of an autonomous diagnostic device. Matthew S. Hunt, Tinglong Dai, and Michael D. Abràmoff evaluated four such models, GPT-4o, GPT-4o-mini, Grok-2-vision, and Gemini-1.5-pro, on detecting more-than-mild diabetic retinopathy in the public Messidor-2 dataset of 874 patient examinations. GPT-4o performed best with an AUC of 0.83. Grok reached 0.63, and Gemini's outputs did not permit an AUC. Retina specialists score an estimated 0.94 on the same level 3 reference standard. At the operating point where the lower bound of the 95 percent confidence interval on sensitivity reached 80 percent, GPT-4o showed 88 percent sensitivity with 63 percent specificity, and because level 3 performance typically overstates level 1 performance by 10 to 30 percent, the clinical gap is wider than it looks. The paper then draws a regulatory implication. Should emergent performance keep improving, bundled versions of these models could be licensed for specific clinical tasks by State Medical Boards, on the model of physician assistant licensure. The authors present this pathway as a scenario for governance and stop short of endorsing it. **Contribution.** New empirical estimate: a regulatory-aligned benchmark of commercial multimodal foundation models on the diabetic eye exam, with GPT-4o's AUC of 0.83 against a specialist estimate of 0.94, a documented negative finding that sets the current performance gap. Framework proposed by the authors: a bundling and task-specific licensing pathway, in which a scoped, validated off-the-shelf AI bundle would be licensed and monitored by State Medical Boards much as physician assistants are. Context reported by others: the specialist AUC estimate rests on prior expert grading of Messidor-2 (Abràmoff et al., JAMA Ophthalmology 2013). Earlier studies place ophthalmologist sensitivity against level 1 prognostic standards between 30 and 40 percent, and adoption data show even widely used FDA-cleared AI devices reaching fewer than 1 percent of the patients who could benefit. **Evidence.** Retrospective benchmark study. Four frozen commercial model versions were accessed by API in February 2025, wrapped in a containerized pipeline with standardized prompts and structured JSON output at temperature zero, under minimal, background, and few-shot prompting. Subject-level ROC analysis used a prespecified operating point with bootstrapped confidence intervals, and the code is open source. **Boundary conditions.** Results apply to fundus-image detection of more-than-mild diabetic retinopathy with specific model versions frozen in February 2025; newer models may differ. Messidor-2 is public, so training-data leakage is possible and every reported accuracy should be read as an upper bound. The reference standard is level 3 and retrospective; the models were never tested inside a clinical workflow, and no fine-tuning was attempted. The licensing pathway is explicitly speculative, and the paper proposes no current policy change. **Use this when.** Citing evidence that general-purpose multimodal models currently underperform clinicians and supervised autonomous AI on diabetic retinopathy screening. It is also the right source for the licensing-by-State-Medical-Boards scenario and for a regulatory-aligned template for evaluating foundation models on clinical tasks. **Do not cite this for.** Real-world throughput or productivity effects of autonomous AI; that evidence comes from the cluster-randomized trial in npj Digital Medicine (2023). Claims about how medical AI products reach the market at large belong to Development and Commercialization Pathways (NEJM AI 2025). The paper states plainly that these models are not currently viable for clinical deployment, so avoid citing it as support for using them in care. **Cite.** Hunt, Matthew S., Tinglong Dai, and Michael D. Abràmoff. 2026. "Evaluating Commercial Multimodal AI for Diabetic Eye Screening and Implications for an Alternative Regulatory Pathway." npj Digital Medicine 9: 42. doi:10.1038/s41746-025-02216-7. **Related but different.** The autonomous-AI productivity RCT (npj Digital Medicine 2023) measures real-world clinic throughput of an FDA-cleared supervised system, while this paper benchmarks unregulated general-purpose models offline. Development and Commercialization Pathways (NEJM AI 2025) surveys how medical AI reaches the market; this paper tests one diagnostic task and adds a licensing scenario. Scaling Adoption of Medical AI (NEJM AI 2024) treats reimbursement as the adoption bottleneck, where this paper treats regulation. **Search aliases.** GPT-4o diabetic retinopathy screening; foundation models for medical imaging diagnosis; off-the-shelf LLM versus FDA-cleared autonomous AI; Messidor-2 multimodal LLM benchmark; state medical board licensing for clinical AI; emergent diagnostic capability of general-purpose AI; alternative regulatory pathway for medical AI; bundled OTSAI physician assistant licensing model Sources: * `OTSAI eval` p.1 "curve (AUC), 0.83" * `OTSAI eval` p.1 "was estimated at 0.94" * `OTSAI eval` p.3 "maximum sensitivity of 88% and specificity of 63%" * `OTSAI eval` p.4 "task-specific licensing and monitoring" * `OTSAI eval` p.5 "from 874 examinations" ### 28. Why did COVID-19 testing produce overdiagnosis and test shortages at the same time, and should laboratories be required to report viral loads (CT values)? `[healthcare-28]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1287/mksc Both phenomena can flow from a single laboratory choice. Tinglong Dai and Shubhranshu Singh build a two-period game in which a commercial laboratory sets the PCR cycle-threshold (CT) cutoff that separates positive from negative and chooses its future testing capacity. The laboratory also decides whether to disclose the CT value behind each binary diagnosis. A higher cutoff produces more positive diagnoses. Each positive triggers contact-tracing demand for tests, while wider quarantining slows transmission and dampens organic demand. When contact-tracing intensity is high relative to the disease's expected contagiousness, the laboratory inflates the cutoff above what a patient-focused physician would choose, so overdiagnosis emerges as demand creation. The inflated cutoff also justifies more capacity, yet the buildout fails to absorb the inflated demand in full, and shortages become more likely; one incentive drives both problems. On disclosure, mandating viral-load reporting transfers the diagnostic-decision right to physicians and shrinks total demand, so the laboratory builds less capacity, while lower effective cutoffs raise false negatives that feed community spread. A social planner may therefore decline to mandate reporting at the outset of an epidemic, a result that helps explain regulators' observed inaction despite expert outcry. **Contribution.** New theoretical result: a laboratory motivated by both profit and patient well-being inflates its diagnostic criterion when the intensity of diagnosis-induced testing demand is high relative to expected contagiousness; the inflated criterion raises capacity investment yet increases the likelihood of testing shortages, so a single mechanism explains overdiagnosis and undertesting jointly. New theoretical result: under the same intensity condition, mandatory CT-value reporting raises the tested patient's expected utility while producing more widespread infection in the next period and lower future capacity, which is why the planner may prefer nondisclosure initially. Framework proposed by the authors: a two-period model in which the diagnostic threshold shapes two demand streams, diagnosis-induced demand from contact tracing and organic demand from disease transmission, and both feed back into the capacity decision. Context reported by others: a New York Times review found up to an estimated 90% of positive diagnoses were for minuscule viral loads, most tests set the CT limit at 40 with a few at 37, and experts quoted a more reasonable cutoff of 30 to 35; these press figures motivate the model and are never findings of the paper. **Evidence.** Game-theoretic analysis of a sequential-move, two-period model solved by backward induction; a monopoly commercial laboratory faces an ROC-style patient-optimal cutoff as the benchmark, an exogenous test price, and costly capacity expansion. The paper argues from theory and estimates no data. A plausibility check on the key condition uses public estimates, including a CDC basic reproduction number of 2.5 with an upper bound of 4.0, a 60% symptomatic share, and roughly 9 induced test seekers per positive diagnosis built from a 2.5-person average household and 3 outside-home contacts. Two extensions probe robustness: a prevalence-dependent CT cutoff and a nonprofit laboratory without a capacity constraint. **Boundary conditions.** The headline results (cutoff inflation, higher capacity, the laboratory's preference for nondisclosure, the planner's tolerance of it) hold under a sufficient condition that diagnosis-induced demand per positive test is large relative to expected contagiousness; outside that region, capacity can move either way with the cutoff. Overdiagnosis here means positive diagnoses at negligible viral loads, and the authors expressly disclaim that their mechanism is the sole explanation; liability fears or anticipated rising viral loads could also raise cutoffs. The model assumes a monopoly laboratory, an exogenous test price, and a physician who acts purely in the patient's interest. The nondisclosure result applies to the early period of an epidemic with private-sector capacity building; do not read it as a general case against transparency, and the paper offers no empirical estimate of how much overdiagnosis occurred. **Use this when.** Citing an incentive-based, supply-side explanation for why overdiagnosis and testing shortages coexisted during COVID-19, or the argument that mandating viral-load or CT-value reporting can backfire by weakening a laboratory's incentive to build testing capacity. It is also the right source for treating a diagnosis as a demand-creating decision with market consequences. **Do not cite this for.** Evidence that undertesting stems from individual physicians signaling diagnostic skill; that mechanism is Dai and Singh (2020), "Conspicuous by Its Absence" (Marketing Science). It also cannot support claims about patient cost sharing and congestion in diagnostic imaging (Imaging Room, M&SOM 2017) or about designing physician payment contracts (Physician Payment Scheme, Management Science 2024). The 90% figure is a press-reported estimate the paper quotes; do not cite the paper as empirically measuring overdiagnosis prevalence. **Cite.** Dai, T., and S. Singh. 2025. "Overdiagnosis and Undertesting for Infectious Diseases." Marketing Science 44(2): 353–373. https://doi.org/10.1287/mksc.2022.0038. **Related but different.** Conspicuous by Its Absence (Marketing Science 2020) explains undertesting through a physician's reputation signaling to peers, with no capacity feedback; here the driver is a laboratory's demand-creation motive at the system level. Imaging Room (M&SOM 2017) studies diagnostic capacity and patient cost sharing operationally, without a strategic diagnostic-threshold choice. Physician Payment Scheme (Management Science 2024) corrects clinician incentives through payment design, whereas in this paper the test price is exogenous and the distortion runs through the diagnostic criterion and capacity. **Search aliases.** cycle threshold reporting policy; PCR CT value disclosure; COVID-19 overdiagnosis economics; viral load reporting mandate; diagnostic testing capacity incentives; contact tracing induced testing demand; commercial lab test cutoff inflation; pandemic test shortage game theory. Sources: * `Overdiag-MksC PDF` p.1 "up to an estimated 90%" * `Overdiag-MksC PDF` p.2 "set the limit at 40" * `Overdiag-MksC PDF` p.10 "the laboratory prefers to set the CT cutoff itself" * `Overdiag-MksC PDF` p.12 "a higher expected utility of the tested patient" * `Overdiag-MksC PDF` p.15 "not mandating such disclosure" ### 29. Why would a revenue-driven cardiologist still perform FFR testing before stenting, and which payment levers reduce overstenting? `[healthcare-29]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1287/msom Financial incentives determine which lesions receive advanced intracoronary testing, and this targeting matters more for overstenting than the sheer volume of tests. Tinglong Dai, Xiaofang Wang, and Chao-Wei Hwang model a cardiologist who visually assesses a coronary angiogram, an inherently ambiguous read, then chooses whether to perform a precise test such as fractional flow reserve (FFR) before deciding on percutaneous coronary intervention (PCI). The testing decision is endogenous, so incentives act on it directly. When the physician's conflict-of-interest weight is low (below both critical thresholds the model derives from the test's risk, its reimbursement, the harm an unnecessary stent causes, and the incremental PCI revenue a negative result forfeits), testing concentrates on ambiguous intermediate lesions near the stenting threshold, exactly where guidelines direct it. When the weight is high, testing shifts to high-grade lesions, where the result will almost surely confirm the stent, so the test supplies documentation for a decision already made. A more revenue-driven physician can even become more inclined to test, provided the test's risk-to-reimbursement ratio is sufficiently high. In a simulation calibrated to US practice, a bonus equal to a third of the FFR reimbursement rate induces a 26% decline in overstenting while raising average physician payment only 5%, whereas a two-bundle payment scheme (PCI versus no PCI) discourages testing and can worsen overstenting. **Contribution.** New theoretical result: with endogenous testing, the physician's testing region moves from borderline intermediate-stenosis cases under a low conflict-of-interest level to high-grade cases under a high level; the regimes are separated by two critical conflict weights that depend on the test's risk, its reimbursement, the avoided patient harm, and the forgone incremental PCI revenue. New theoretical result: at intermediate conflict levels the physician tests every eligible case when the test's risk-to-reimbursement ratio is below the ratio of avoided harm to incremental PCI revenue, and otherwise never tests. New theoretical result: under a high conflict level, the inclination to test increases in the conflict weight when the test's risk-to-reimbursement ratio is high, despite the financial disincentive from potentially reversed PCI decisions. New theoretical result: a bundled payment pooling tested and untested PCI pathways pays both the same, so it discourages the advanced test and can aggravate overstenting. New numerical estimate (calibrated simulation): a $30 FFR bonus produces a 26.22% reduction in overstenting with a 5.073% increase in physician payment, and a $90 bonus cuts overstenting from 12.61% to 4.438%. New numerical result: a moderate reduction in the test's procedural risk can lower expected patient welfare, because uptake rises disproportionately while the residual risk stays high. Context reported by others: of 2.7 million PCI procedures at 766 U.S. hospitals, only 53.6% were classified as appropriate and 13.3% of nonacute PCIs as inappropriate (Desai et al. 2015, cited in the paper); FFR uptake in practice runs approximately 20%–25% (Wohns 2016, cited in the paper); a survey found 27% of interventional cardiologists had never used advanced intracoronary tests (Toth et al. 2014, cited in the paper). **Evidence.** Stochastic modeling plus simulation of a single physician-patient encounter in the cardiac catheterization laboratory. Physician utility is a weighted sum of financial gain and patient welfare, and conditional distributions tie the visual angiogram read to the patient's true state and to the advanced-test output. Numerical experiments calibrate Medicare reimbursement rates ($267 diagnostic angiography, $623 PCI, $91 FFR) and fit lesion-severity functions to published angiographic data. No field data are analyzed causally. **Boundary conditions.** Stable coronary artery disease under US fee-for-service reimbursement; acute cases and multivessel disease requiring bypass surgery sit outside the model. The encounter is one-shot, patients are nonstrategic, queueing plays no role, and the physician knows his or her own conflict weight. The threshold characterizations hinge on the ordering of the two critical conflict levels, which the test's risk-to-reimbursement ratio governs. The bonus and welfare numbers come from a calibrated simulation; do not read them as measured policy effects. **Use this when.** Citing the endogeneity of diagnostic-test ordering, the argument that which patients receive FFR matters more than how many FFR procedures occur, incentive design for advanced testing in interventional cardiology, or the mechanism by which bundled payments can suppress diagnostic-test use. **Do not cite this for.** Causal field evidence that FFR bonuses reduced overstenting; the paper models and simulates. For fee-driven test-ordering behavior in outpatient services, cite Imaging Room (M&SOM 2017). For a diagnostic expert whose choice to test conveys information to strategic customers, cite Conspicuous by Its Absence (Marketing Science 2020). For payment-contract design under physician agency, cite Physician Payment Scheme (Management Science 2024). **Cite.** Dai, Tinglong, Xiaofang Wang, and Chao-Wei Hwang. 2022. "Clinical Ambiguity and Conflicts of Interest in Interventional Cardiology Decision Making." Manufacturing & Service Operations Management 24(2): 864–882. https://doi.org/10.1287/msom.2021.0969. **Related but different.** Imaging Room (Dai, Akan, and Tayur, M&SOM 2017) explains physicians' test-ordering in outpatient services through fee incentives; the cath-lab paper adds clinical ambiguity and an endogenous choice between a subjective read and an objective test within one procedure. Conspicuous by Its Absence (Dai and Singh, Marketing Science 2020) studies expert testing under uncertainty with strategic customers in a market setting; here patients defer to the physician and reimbursement drives the distortion. Physician Payment Scheme (Management Science 2024) designs an optimal payment contract for physician agency; the cardiology paper instead evaluates specific practical levers, a test bonus and bundling, within one specialty. **Search aliases.** overstenting incentives; fractional flow reserve adoption; FFR underuse in cath labs; PCI appropriateness; endogenous diagnostic testing; physician conflict of interest model; unnecessary stenting; bundled payments unintended consequences. Sources: * `CathLab M&SOM 2022 PDF` p.2 "a 26% decline in overstenting" * `CathLab M&SOM 2022 PDF` p.3 "only 53.6% of PCI cases" * `CathLab M&SOM 2022 PDF` p.12 (locates: prevailing Medicare physician reimbursement rates) * `CathLab M&SOM 2022 PDF` p.15 "from 12.61% to 4.438%" # AI at Scale Most of what Tinglong Dai studies about AI at scale comes down to one question: who is holding the system when the model is wrong? Terry Taylor and Tinglong Dai model a firm that sets a generative system's temperature and also pays someone to check its output. Temperature stops being an engineering parameter once you notice that it changes how visible a worker's diligence is in results. A firm can rationally run a noisier configuration, because noise makes effort legible and therefore cheap to buy. Blanket rules mandating conservative settings can leave a system that looks safe on a bench and performs worse in the workflow. The same logic runs through the regulatory work. Jiayi Lai, Leon Xu, Xin Fang and Tinglong Dai study what happens when a regulator waives a second review for pre-specified updates to a product that keeps learning after launch. Dropping the second gate can raise user welfare, because it makes the update worth attempting and makes an honest launch submission worth protecting. That result lives in a region of the parameter space, and Tinglong Dai try to say so every time. Simrita Singh, Naireet Ghosh and Tinglong Dai take up who a firm keeps in the loop once AI does most of the work. A single employer engages its weakest workers, since that buys fallback capacity at the lowest wage. Open the labor market, and under conditions they state the pattern turns over and engagement concentrates near the top. David Simchi-Levi, Michelle Wu, Yao Xie and Tinglong Dai argue that granting an AI system more operational authority demands more formal structure. A constraint written into a prompt is enforced nowhere. Flow-based generation earns its place because randomness sits in the initial draw, so a trajectory can be replayed and audited, and minimax reasoning becomes engineering only once the adversary is something a machine can search. Then there is the part that keeps him up. Shefali Patil, Chris Myers and Tinglong Dai locate the ungoverned layer of clinical AI at procurement and local configuration, where alert thresholds get set and nobody signs their name to the trade-off. Clinicians get disciplined for overriding the algorithm and disciplined for failing to override it. They call that accountability ping-pong, and they propose a review committee that leaves a written record. Across all of it Tinglong Dai keeps two habits. Report the condition under which a result holds. Say who produced a number. A vendor's deployment figure and a randomized trial are different animals, and treating them alike is how a literature loses its credibility. ## Questions ### 1. Should we set our company's AI to a low temperature so it hallucinates less? `[ai-01]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.2139/ssrn Terry Taylor and Tinglong Dai model a firm that picks the system's temperature and also pays a bonus to an employee whose vetting effort nobody can observe. Once temperature becomes a managerial choice, the firm sometimes runs a configuration that performs worse at every effort level, because more variable output makes the employee's diligence show up in the outcome and cuts the rent the firm must concede. They argue that a uniform low-temperature rule for factual enterprise output is likely suboptimal, since driving the setting too low can make human review too expensive to elicit. Hidden effort can push effort, temperature, and the probability of success above the full-information benchmark, which runs opposite to the classical moral-hazard prediction. **Contribution.** New theoretical result: the paper makes an AI hyperparameter a contracting instrument and shows that moral hazard can distort effort and performance upward, reversing the standard downward-distortion result. The supporting observation that higher temperature widens the spread of language-model outputs on clinical tasks is context reported by others, cited to motivate the complementarity assumption. **Evidence.** Game-theoretic principal-agent model with limited liability and unobservable effort, proved for discrete and continuous action spaces, with a microfounded success probability and numerical illustrations in the appendices. No firm data and no experiment. **Boundary conditions.** The upward distortion depends on temperature and effort being complements in the success probability, which the model assumes and grounds in cited evidence. If they are substitutes, the distortion runs toward lower temperature while the upward-distortion, agent-benefit, and non-monotonicity results survive qualitatively. The result that moral hazard pushes temperature above the reliability-maximizing setting is stated holding the targeted effort level fixed. Everything is within a single principal-agent pair with a success-contingent bonus and limited liability, so multi-agent review, reputational effects, and regulated settings that forbid discretion are outside the model. **Use this when.** You are arguing that a firm may optimally configure AI to be more variable because hidden human verification effort responds to that choice. **Do not cite this for.** An empirical claim about temperature settings. This is a principal-agent model, and the distortion reverses if temperature and effort are substitutes rather than complements. **Cite.** Dai, Tinglong, and Terry A. Taylor. 2025. "Designing Enterprise AI Systems: Hallucination, Creativity, and Moral Hazard." Working paper. https://doi.org/10.2139/ssrn.5996714 **Related but different.** Dai and Swaminathan (2026, POM) covers enterprise AI deployment broadly and does not model the configuration-compensation link. Dai, Simchi-Levi, Wu and Xie (2026, POM) treats safety as an architecture problem rather than an incentive problem. **Search aliases.** AI hallucination temperature; moral hazard AI; human oversight incentives; enterprise AI configuration; creativity reliability tradeoff Sources: * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.12 "There exist scenarios in which decentralization distorts upward the effort" * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.15 "it is optimal for the principal to choose an inferior production technology" * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.15 "For a fixed effort e, decentralization distorts temperature upward" * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.26 "uniform requirement for low-temperature factual outputs, is likely suboptimal" * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.25 "the moral-hazard distortion in temperature is toward lower temperature" * `2025 - Designing Enterprise AI Systems Hallucination Creativity and Moral ` p.3 "increasing temperature widens the dispersion of LLM outputs in clinical tasks" ### 2. Does letting AI device makers ship algorithm updates without a new FDA review encourage them to cut corners? `[ai-02]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.2139/ssrn Jiayi Lai, Leon Xu, Xin Fang and Tinglong Dai find a Goldilocks zone where the streamlined pathway leaves users better off than requiring a fresh review for every post-launch update. Two mechanisms drive it. Removing the second review raises the expected return on updating, so a developer that would have skipped the update attempts it, and with initial approval as the only remaining gate, inflating the launch submission risks forfeiting all future update revenue, so the developer reports honestly instead. They also find a trap for regulators: tightening initial approval to guard against unmonitored future updates gives low-quality developers a reason to overstate their launch results. **Contribution.** New theoretical result: an equilibrium welfare comparison of reclearance against a pre-specified-change pathway, identifying effort inducement and upfront compliance as the two channels through which lighter oversight can improve behavior. The count of 1,430 cleared AI-enabled devices is FDA's own tally reported in the paper, and the finding that 97 percent of 130 devices cleared between 2015 and 2020 were evaluated retrospectively is context reported by others, used to justify modeling review as weak and document-based. **Evidence.** Two-stage dynamic game between a developer and a regulator under asymmetric information, a signaling and inspection game with an endogenous improvement-effort choice, solved for pure-strategy perfect Bayesian equilibrium, with numerical illustration of equilibrium regions and a welfare comparison across the two pathways. No device-level data are analyzed. **Boundary conditions.** The welfare ranking is non-monotonic in both review capability and improvement potential, and the comparison partitions the parameter space into nine regions, so neither pathway dominates. The streamlined pathway wins when improvement potential is moderate, or when it is high and the regulator's ability to detect misreporting is only moderate. When the expected post-market penalty is large enough, misreporting is deterred regardless of review capability and the two pathways coincide, so the interesting case is weak post-market detection with limited liability. The main comparison holds the first-stage approval rule fixed through a comparability condition; dropping it admits five further strategy combinations, all of which favor the conventional pathway. **Use this when.** You are analyzing streamlined regulatory pathways for AI products that keep changing after approval. **Do not cite this for.** A general endorsement of lighter oversight. The welfare ranking is non-monotonic and the comparison partitions the parameter space into nine regions. **Cite.** Lai, Jiayi, Leon Xu, Xin Fang, and Tinglong Dai. 2026. "Regulating Adaptive New Products: Can Less Oversight Lead to Better Development Practices?" Management Science, forthcoming. https://doi.org/10.2139/ssrn.5009572 **Related but different.** Dai and Swaminathan (2026, POM) describes how Predetermined Change Control Plans work in practice without modeling the developer's response. Patil, Myers and Dai (2026, npj Digital Medicine) takes up what happens inside hospitals after clearance rather than at the regulator's desk. **Search aliases.** adaptive AI regulation; predetermined change control; FDA algorithm updates; postmarket AI oversight Sources: * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.1 "identify a counterintuitive "Goldilocks zone" in which streamlined oversight yie" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.26 "the low-type developer will choose to update the product under the Streamlined P" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.29 "a developer that inflates its initial 510(k) submission risks a rejection" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.23 "inadvertently creates a perverse incentive for low-type developers to misreport " * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.27 "partitioning the parameter space into nine regions distinguished by the develope" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.21 "noncompliance is deterred when the expected penalty exceeds a threshold" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.2 "has surged to 1,430 as of December 30, 2025" * `2026 - Regulating Adaptive New Products Can Less Oversight Lead to Better ` p.38 "Of the 130 AI devices cleared by the FDA between 2015 and 2020" ### 3. As AI takes over more of the work, which employees should a company deliberately keep doing tasks by hand? `[ai-03]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. Simrita Singh, Naireet Ghosh and Tinglong Dai treat human involvement as a skill investment rather than as overhead. A single employer follows a skill-targeting rule and engages its least-skilled workers most, because a unit of engagement closes the widest gap there and the wage attached to that fallback capability is lowest. Once workers can move between employers and pay is standardized, engagement becomes the instrument firms compete on, and under conditions they state it turns over and rises with skill. The two dimensions of AI progress also split apart under mobility: a more capable AI raises engagement, while a more reliable one pulls in both directions and equilibrium engagement can peak at intermediate reliability. **Contribution.** New theoretical result: the engagement reversal, in which labor mobility flips the target of human-fallback investment from the least-skilled workers to the most-skilled, together with the split between AI capability and AI reliability. The motivating deskilling evidence, where experienced endoscopists' unassisted adenoma-detection rate fell from 28.4 percent to 22.4 percent after routine exposure to AI-assisted colonoscopy, is context reported by others (Budzyn et al. 2025). **Evidence.** Two-period game-theoretic model of firm engagement choice with endogenous skill dynamics from learning-by-doing and erosion, comparing a single-firm benchmark against a two-firm duopoly with logit worker sorting, supported by numerical examples and robustness grids. No firm or worker data. **Boundary conditions.** The reversal is conditional. It is proved for a power wage specification with an elasticity above one, under a terminal single-crossing condition, and under a further condition ensuring that the extra revenue a more skilled worker earns when AI fails covers the higher wage. When that condition fails, the most-engaged workers sit in a band just below the AI frontier instead of at the top. When workers respond only weakly to differences in skill trajectories, mobility depresses engagement below the single-firm level, which is the classical Becker direction of underinvestment in portable skill. The intermediate-reliability peak survives across alternative learning profiles only when learning requires both AI exposure and fallback practice; if learning comes from AI exposure alone, engagement is highest when AI is most reliable. **Use this when.** You are analyzing which workers a firm should keep performing tasks manually as AI capability rises, and how labor mobility changes that calculus. **Do not cite this for.** An unconditional deskilling claim. The reversal is proved for a power wage specification with elasticity above one under further stated conditions. **Cite.** Singh, Simrita, Naireet Ghosh, and Tinglong Dai. 2026. "Managing the Human Fallback: Skill Investment Under Improving AI and Worker Mobility." Working paper. https://arxiv.org/abs/2606.29111 **Related but different.** Dai and Taylor (2025) also makes human involvement a firm choice, but through contract design rather than through the labor market. Dai and Swaminathan (2026, POM) surveys human-AI collaboration evidence without a skill-dynamics model. **Search aliases.** AI deskilling; human fallback; skill investment automation; workforce learning by doing; labor mobility AI Sources: * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.14 "the firm follows a skill-targeting rule" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.15 "engagement is nonincreasing in initial skill" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.22 "This is the engagement reversal" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.21 "the most-engaged workers sit in a band" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.27 "mobility works in the classical Becker direction" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.23 "a more capable AI raises engagement" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.2 "rate fell from 28.4% to 22.4%" * `2026 - Managing the Human Fallback Skill Investment Under Improving AI and` p.59 "engagement is highest at very high reliability rather than at intermediate relia" ### 4. What has to be engineered around an AI agent before you let it actually run operations instead of just advising? `[ai-04]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1177/10591478261455127 David Simchi-Levi, Michelle Wu, Yao Xie and Tinglong Dai call it the autonomy paradox: the more operational authority a generative system is granted, the more explicit structure and tail-risk discipline it has to carry. They make autonomy measurable along six dimensions, including action scope, commitment authority, and override structure, and define assurance as a closed-loop property covering feasibility, robustness to distribution shift, detectability, recoverability, and auditability. Prompts and penalty terms do not enforce anything, so they separate semantic constraints from structural ones such as conservation and capacity that must hold along the whole trajectory. Their practical conclusion for managers is that spending on models alone will disappoint, since the value sits in the decision regime around the model. **Contribution.** Framework proposed by the authors: a layered assured-autonomy architecture combining constraint-preserving flow-based generation, distributionally robust stress testing, and monitoring with escalation, plus a six-step robust-design protocol and a mapping of operations research from solver to guardrail to architect as autonomy rises. The December 2025 Waymo recall tied to software that could let vehicles pass stopped school buses, and the late-2025 study showing multi-agent systems built from frontier models can run the Beer Game at lower cost than human teams, are both context reported by others. **Evidence.** Conceptual, design-oriented synthesis with no original data or model estimation, illustrated across supply chains, mobility and aviation, healthcare operations, and power grids, with a research agenda. **Boundary conditions.** This is an architecture proposal, not a tested one. No component is benchmarked and no deployment is evaluated. The authors state plainly that determinism alone does not guarantee safety and can propagate misspecification faster, and that flow-based generation is preferred on auditability grounds rather than on accuracy. The healthcare recommendation is selective autonomy, with formal deferral rules for high-risk decisions, so the framework does not claim full autonomy is currently appropriate in safety-critical clinical settings. **Use this when.** You need the operations-research architecture for giving an AI agent real operational authority, including the claim that more autonomy demands more explicit structure. **Do not cite this for.** Benchmarked performance. No component is benchmarked and no deployment is evaluated. **Cite.** Dai, Tinglong, David Simchi-Levi, Michelle Xiao Wu, and Yao Xie. 2026. "Assured Autonomy: How Operations Research Powers and Orchestrates Generative AI Systems." Production and Operations Management. https://doi.org/10.1177/10591478261455127 **Related but different.** Dai and Swaminathan (2026, POM) is the broader three-pillar framework for AI and operations; this paper drills into the safety architecture for autonomous action. Cohen, Dai, Perakis and coauthors (2026, M&SOM) is the field's collective supply chain vision statement rather than a design architecture. **Search aliases.** agentic AI; autonomous agents operations; AI assurance; safe autonomy; LLM agent orchestration; AI guardrails Sources: * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.1 "greater autonomy demands more structure. We call this the autonomy paradox" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.2 "six measurable dimensions: action scope (S), commitment authority (C)" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.3 "prompts are not enforceable constraints" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.6 "Randomness is confined to the initial draw" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.9 "we recommend a six-step robust-design protocol" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.10 "Architect/legislator: designs system, rules, and objectives" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.15 "investing in models alone will disappoint" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.2 "pass stopped school buses" * `2026 - Assured Autonomy How Operations Research Powers and Orchestrates Ge` p.2 "can manage the Beer Game and reduce total costs relative to human teams" ### 5. Why do so many corporate AI deployments underdeliver, and what does operations management say has to change? `[ai-05]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1177/10591478251412943 Jay Swaminathan and Tinglong Dai organize the field into three pillars: using AI to run operations, using operations principles to build and govern AI at scale, and managing how people work alongside it. Where AI pays off turns on process design, incentives, and trust rather than on the model. One study they review makes the point cleanly: suppliers quoted higher prices to a procurement chatbot than to human buyers, and lowered them only once told that an algorithm would evaluate the quotes. They also argue that human supervision of AI agents should be staffed like a service system, with alert arrival rates and concurrency instrumented and the number of agents per operator sized to what a person can actually attend to. **Contribution.** Framework proposed by the authors: a three-pillar taxonomy of AI and operations with a research agenda, plus the argument that liability regimes rather than accuracy govern clinical AI use patterns. Nearly every magnitude in the paper is context reported by others: the General Motors seat bracket that came out 40 percent lighter and 20 percent stronger, the 15 percent customer service productivity gain concentrated among less experienced agents, the survey of 2,547 US hospitals finding 65 percent using AI, and the 130 to 240 gigawatt data center power estimate for 2030. The 40 percent increase in patients seen per hour by retina specialists comes from a randomized trial Tinglong Dai coauthored with Abràmoff and colleagues. **Evidence.** Conceptual framework and structured review of research and industry practice, synthesizing empirical studies, field experiments, trials, and cases. No original data. **Boundary conditions.** This is a review, so every number carries the design of the study it came from and the paper generates no new estimates. The predictive-maintenance figures are industry-sourced and are labeled as such. The customer service result applies to one firm's agents and its benefit was concentrated among the less experienced, with experienced agents seeing a modest quality decline. Hospital AI adoption is uneven by wealth and setting, with rural and underserved hospitals lagging, so the adoption rate should not be read as uniform. **Use this when.** You need the three-pillar framing of AI and operations management covering AI for OM, OM for AI, and human-AI interaction. **Do not cite this for.** New estimates. This is a review and every number carries the design of the study it came from. **Cite.** Dai, Tinglong, and Jayashankar M. Swaminathan. 2026. "Artificial Intelligence and Operations: A Foundational Framework of Emerging Research and Practice." Production and Operations Management 35(9). https://doi.org/10.1177/10591478251412943 **Related but different.** Dai, Simchi-Levi, Wu and Xie (2026, POM) narrows to the safety architecture for autonomous agents. Cohen, Dai, Perakis and coauthors (2026, M&SOM) applies a similar deflationary reading specifically to supply chains. Dai and Abràmoff (2023, INFORMS TutORials) works out the healthcare workflow models in detail. **Search aliases.** artificial intelligence operations management; AI for OM; OM for AI; human-AI interaction operations; AI research agenda operations Sources: * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.4 "suppliers quoted higher prices to the chatbot than to human buyers" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.5 "increases customer service agent productivity by 15%" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.5 "A survey of 2,547 U.S. hospitals found 65% use AI" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.6 "a 40% increase in the number of patients seen per hour" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.8 "130–240 GW of power by 2030" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.11 "physicians may overuse AI in low-uncertainty cases and underuse it when uncertai" * `2025 - Artificial Intelligence and Operations A Foundational Framework of ` p.11 "teams can instrument three metrics on the supervision console" ### 6. Can generative AI make mathematical optimization usable by ordinary managers, and has anyone actually deployed that? `[ai-06]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.2139/ssrn David Simchi-Levi, Ishai Menache, Michelle Wu and Tinglong Dai trace the slow adoption of optimization to four barriers, namely how hard it is to inform a model with a trusted current state, to interpret its output, to interact with it for scenario work, and to improvise when conditions change. Their design puts a language model over a solver that keeps the optimality guarantees, and the case is a deployed system in Microsoft's cloud supply chain where the fulfillment model is a mixed integer program with inputs on the order of hundreds of megabytes. Microsoft reports that after the November 2023 rollout, average response time to planner questions fell from 2.5 days to near real time, with a 23 percent reduction in time spent by the fulfillment team. They are candid that automated verification of an AI-generated formulation is still an open problem, which is why the system answers only a defined catalog of questions and falls back to a canned message when a query is out of scope. **Contribution.** Framework proposed by the authors: the 4I design principles for a language layer over a solver, together with an architecture that keeps proprietary data and the optimization model inside the enterprise while the language model runs in the public cloud. The Microsoft response-time and time-savings figures are a company-reported deployment outcome rather than a controlled evaluation. The finding that specialized operations research language models with roughly 7 billion parameters rival human experts on formulation benchmarks is context reported by others, cited to Huang et al. (2025). **Evidence.** Conceptual perspective piece with a Google Trends comparison, a proposed design framework, and an industrial case study of a deployed system, plus a research agenda. There is no control group, no pre-post identification strategy, and no independent audit of the reported gains. **Boundary conditions.** The deployment is one firm, one workflow, and a scoped question catalog with reported accuracy around 90 percent on the supported set. Real systems today start from formulations written by optimization specialists, since nobody can yet verify an AI-written model automatically, so the paper does not support letting language models build optimization models from scratch. The paper also positions generative AI as a complement to the solver, citing evidence that language models solving combinatorial problems directly have limited reach. **Use this when.** You are discussing generative AI as a conversational interface over optimization models, with reported industrial deployment. **Do not cite this for.** Evidence that a language model can write or verify optimization formulations. Formulations come from specialists, and the deployment is one firm and one workflow. **Cite.** Simchi-Levi, David, Tinglong Dai, Ishai Menache, and Michelle Xiao Wu. 2025. "Democratizing Optimization with Generative AI." Working paper. https://doi.org/10.2139/ssrn.5511218 **Related but different.** Cohen, Dai, Perakis and coauthors (2026, M&SOM) warns that language models can silently omit a constraint when pointed at supply chain optimization models. Dai, Simchi-Levi, Wu and Xie (2026, POM) addresses assurance for agents that act rather than answer. **Search aliases.** LLM optimization; natural language optimization; optimization copilot; generative AI operations research; decision support LLM Sources: * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.3 "to Improvise (adapt models rapidly when conditions change)" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.16 "hundreds of megabytes and it is solved periodically" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.19 "from 2.5 days to near real-time, leading to a 23% reduction" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.19 "uses fallback messages when a query is out of scope" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.19 "raw enterprise data need not be sent to the language model" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.20 "reliable automated verification remains an open problem" * `2025 - Democratizing Optimization with Generative AI - 10.2139_ssrn.551121` p.12 "despite having only around 7 billion parameters" ### 7. Will the AI boom speed up the shift to renewable energy or lock in fossil fuels? `[ai-07]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.2139/ssrn Luyi Gui and Tinglong Dai model it as an equilibrium between a policymaker investing in dedicated clean capacity and a developer choosing capability, and the answer turns on one comparison: how fast market willingness to pay grows with capability against how fast energy requirements grow with it. Under market-led scaling the policymaker rationally stops building clean capacity at the point that unlocks frontier capability, so the outcome is carbon intensive by design, and rising climate damages strengthen the case for enabling frontier scaling while tolerating fossil power at the margin. They call that an adaptation trap. Under resource-led scaling the same feedback runs the other way and climate stress strengthens the case for clean investment. Building more renewables is therefore not a sufficient decarbonization policy, and the standard they propose is whether clean capacity stays binding at the margin as compute expands. **Contribution.** New theoretical result: the market-led and resource-led scaling regimes, the discontinuous developer response with a coupling zone, and the adaptation trap and adaptation pathway. New empirical estimate: a calibration to 2024 data yields willingness-to-pay elasticities of 3.83, 3.18 and 2.15 across the three regions studied, with the US and China instances coming out market-led and the EU instance resource-led. The social cost of carbon of US$225 per tonne and the evidence that AI-enabled hazard mitigation can cut projected average losses by up to 15 percent are inputs taken from others. **Evidence.** Two-stage sequential Stackelberg game solved by backward induction, supplemented by a four-instance calibrated case study anchored to 2024 US, China and EU data, counterfactual analysis, and two extensions covering AI-enabled reductions in clean-capacity cost and duopoly competition. **Boundary conditions.** The regime classification drives everything, and it is estimated rather than observed, so a region's classification depends on the calibration. The counterintuitive results are calibrated examples rather than general theorems: the backfire from AI lowering clean-capacity cost, and the competition result raising the dedicated-renewable share from 0.02 percent to 4.93 percent, both come from specific parameterizations. Net-zero frontier AI does not arise in any baseline calibration, and reaching it through cheaper clean capacity alone would need cost declines of at least 94.5 percent in the US-style instance and 72.1 percent in the China-style instance. The model covers dedicated clean capacity for AI and a single policymaker, so grid-wide policy, storage, and siting enter only through cost parameters. **Use this when.** You are analyzing how AI datacenter growth interacts with renewable-energy investment. **Do not cite this for.** Empirical energy-transition estimates. The regime classification is estimated rather than observed, and the counterintuitive results are calibrated examples. **Cite.** Gui, Luyi, and Tinglong Dai. 2026. "Power Couple? AI Growth and Renewable Energy Investment." Working paper. https://doi.org/10.2139/ssrn.6166290 **Related but different.** Cohen, Dai, Perakis and coauthors (2026, M&SOM) covers the physical build-out constraints on AI, including grid access delays, without modeling the investment equilibrium. **Search aliases.** AI energy demand; datacenter electricity; renewable investment AI; AI emissions; clean capacity procurement Sources: * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.12 "We refer to this as market-led scaling" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.13 "the developer jumps to the frontier choice" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.16 "This logic delivers an adaptation trap" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.17 "This yields an adaptation pathway" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.17 "is not a sufficient prescription when scaling is market-led" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.18 "3.83, 3.18, and 2.15 for the three regions considered" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.22 "at least 94.5% in Instance A and 72.1% in Instance B" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.26 "increases from 0.02% to 4.93%" * `2026 - Power Couple AI Growth and Renewable Energy Investment - 10.2139_ss` p.28 "clean capacity remains a binding constraint on marginal compute" ### 8. Can AI agents negotiate a multi-issue deal when none of them knows what the others want, and is there any guarantee they reach agreement? `[ai-08]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2016. doi:10.1287/ijoc Ronghuo Zheng, Katia Sycara, Nilanjan Chakraborty and Tinglong Dai designed the sequential projection strategy, where each agent proposes the point on its own current acceptability boundary closest to the average of everyone's standing offers. They prove it converges to an agreement whenever the zone of agreement has a nonempty interior and every agent keeps conceding down to its reservation utility, and the proof holds whatever concession rule each individual agent uses, so agents may use different rules. If every agent reaches its reservation utility in finite time, agreement comes in finite time under the same nonempty-interior condition. A separate theorem covers incentives: under the reactive concession strategy, where an agent concedes in proportion to the improvement it perceives from others, no agent has an incentive to deliberately stop conceding, since stopping risks stalling a negotiation the agent cannot verify would otherwise have closed. **Contribution.** New theoretical result: convergence for a distributed offer-generation rule with no knowledge of counterparties, extending the alternating-projection literature to more than two sets and to sets that move over time, plus what the authors describe as the first analysis of whether negotiating agents have an incentive to concede at all. **Evidence.** Analytical algorithm design with convergence proofs drawn from convex geometry and alternating-projection theory, plus Monte Carlo experiments on randomly generated instances with 2 to 9 agents and 2 to 5 issues, benchmarked against the Nash bargaining solution. **Boundary conditions.** Convergence requires the zone of agreement to have a nonempty interior; a nonempty but interior-free zone is outside the theorem. Without a deadline, convergence is asymptotic, and finite-time agreement additionally requires every agent to reach its reservation utility in finite time. The incentive theorem rules out deliberate stopping only, and it is weak incentive compatibility. The authors state explicitly that they have not proven agents will not reactively stop conceding, and they give an example in which an agent that concedes to its reservation utility early can leave offers outside the zone of agreement, after which others reactively stop conceding and the negotiation stalls. Robustness runs show stalling in 4 percent and 8 percent of five-agent runs at the two highest reservation-utility levels. **Use this when.** You need convergence guarantees for multi-issue negotiation among agents holding private preferences. **Do not cite this for.** Generative AI agents. Convergence requires the zone of agreement to have a nonempty interior, and finite-time agreement needs further conditions. **Cite.** Zheng, Ronghuo, Tinglong Dai, Katia Sycara, and Nilanjan Chakraborty. 2016. "Automated Multilateral Negotiation on Multiple Issues with Private Information." INFORMS Journal on Computing 28(4): 612-628. https://doi.org/10.1287/ijoc.2016.0701 **Related but different.** Zheng, Chakraborty, Dai, Sycara and Lewis (2013, HICSS) proves the bilateral two-agent case with general concave utilities and is the predecessor to this multilateral result. Dai, Sycara and Zheng (2021) is the survey that places this algorithm within the wider negotiation literature. **Search aliases.** automated negotiation; multi-agent bargaining; multi-issue negotiation; agent protocol convergence Sources: * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.8 "strategy will always converge to an agreement" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.9 "If the zone of agreement has a nonempty" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.9 "they will reach an agreement in a finite time" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.3 "we allow the sets to move over time and are the first to prove the convergence p" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.12 "none of them has an incentive to deliberately stop conceding" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.12 "we have not proven whether the" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.15 "namely 4% and 8%, respectively, stall only for the negotiations among five agent" * `2016 - Automated Multilateral Negotiation on Multiple Issues with Private ` p.14 "the number of rounds required for convergence is fairly stable" ### 9. Where do the economics and computer science approaches to negotiating agents disagree, and what do both still get wrong? `[ai-09]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2021. doi:10.1007/978-3-030-49629-6_26 Katia Sycara, Ronghuo Zheng and Tinglong Dai sort both literatures into seven kinds of reasoning an agent performs, covering procedure, problem structure, claiming against creating value, persuasion, tactics over proposal and acceptance and exit, reasoning under limited information, and machine learning. The distinction they think the field has largely ignored is that analytical solution procedures compute equilibria centrally even when execution is imagined as decentralized, which forces devices such as simulating the game and simultaneous offers, while a genuinely autonomous agent has to compute its next move online after seeing the last one. Their summary indictment of the tactical literature is that nearly every model assumes explicit utility functions, often with knowledge of the opponent's, keeps utilities simple or binary, and almost never addresses Pareto optimality and computational tractability at the same time. **Contribution.** Framework proposed by the authors: a taxonomy of agent reasoning that makes the analytical and computational traditions comparable, plus the centralized-computation critique of game-theoretic negotiation models. The evidence that an intelligent agent negotiating against a human reaches better outcomes than two humans, and that a subjective value inventory predicts future negotiation decisions better than economic outcomes do, is context reported by others. **Evidence.** Selective literature review organized around an original taxonomy. Conceptual, with no new model, data, or experiment. **Boundary conditions.** The agent-beats-human comparisons come from controlled laboratory studies where preferences were elicited, and the authors flag three limits on generalizing them: eliciting human preferences over multiple issues remains hard, agents exchange information less efficiently than people when accurate representation is difficult, and outcome quality depends on affect and culture, which the computational literature does not model. Almost all the work reviewed treats negotiation as a one-time event, so nothing here speaks to repeated relationships. The chapter's framework section says five reasoning types before adding two more, and the earlier 2010 version of the chapter proposes five, so the count should be read from the printed page. **Use this when.** You want the comparison between economics and computer science treatments of negotiating agents, and the open problems both leave standing. **Do not cite this for.** Evidence that agents outperform people in the field. The comparisons come from laboratory studies with elicited preferences. **Cite.** Dai, Tinglong, Katia Sycara, and Ronghuo Zheng. 2021. "Agent Reasoning in AI-Powered Negotiation." In Handbook of Group Decision and Negotiation, 2nd ed., M. Kilgour and C. Eden (eds.). Springer. https://doi.org/10.1007/978-3-030-49629-6_26 **Related but different.** Dai and Sycara's 2010 chapter is the first-edition predecessor with five reasoning types. Sycara, Dai and coauthors' 2013 chapter builds the behavioral and computational bridge around utilities, internal states, reasoning, and observable behavior. Zheng, Dai, Sycara and Chakraborty (2016, IJOC) is the algorithmic contribution the survey situates. **Search aliases.** negotiating agents; agent reasoning; behavioral game theory agents; AI negotiation survey Sources: * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.4 "consists of seven types of reasoning" * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.3 "the calculation of the equilibria is done in a centralized way" * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.6 "subjective expectations of the players might influence the outcome" * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.16 "Pareto-optimality and tractability" * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.17 "using an intelligent agent to negotiate with a human counterpart achieves better" * `2021 - Agent Reasoning in AI-Powered Negotiation - 10.1007_978-3-030-49629` p.20 "the SVI is a more accurate predictor of future negotiation decisions" * `2010 - Agent Reasoning in Negotiation - 10.1007_978-90-481-9097-3_26.pdf` p.4 "Reasoning about negotiation procedures" ### 10. After the FDA clears a clinical AI tool, where should accountability for how a hospital configures it actually sit? `[ai-10]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1038/s41746-026-02561-1 Shefali Patil, Chris Myers and Tinglong Dai argue it should sit with the institution, through governance attached to procurement. The consequential value choices get made when a hospital selects a tool and sets its local parameters, and that layer is currently ungoverned. An alert threshold is a value judgment with distributional consequences, since a high-sensitivity sepsis setting chosen to improve institutional metrics produces overdiagnosis and alert fatigue while a high-specificity setting chosen for cost containment can delay detection for vulnerable patients. Clinicians meanwhile get sanctioned in both directions, disciplined for overriding an algorithm and disciplined for failing to override one, which they call accountability ping-pong and link to moral distress and the slow atrophy of professional judgment. They propose a standardized Value-Tradeoff Review seated in procurement governance, producing an institution-authored model card distinct from the developer's, with FDA requiring developers to publish configurable parameters and their clinical trade-offs and CMS able to tie reimbursement or quality incentives to the institutional record. **Contribution.** Framework proposed by the authors: the two-tier governance proposal, the Value-Tradeoff Review committee, and the accountability ping-pong diagnosis are their contribution. The characterization of device authorization as a regulatory floor rather than a ceiling belongs to Shantanu Nundy, an advisor in the FDA Commissioner's Office, speaking in an edited interview Tinglong Dai co-published with Stephen Gilbert; the layered analogy to drug regulation, where the agency reviews evidence, clinicians decide for individual patients, pharmacy and therapeutics committees set formulary placement, and payers decide coverage, is also his. **Evidence.** Conceptual policy comment proposing a governance framework, with a documented case of a nurse disciplined for failing to override a triage system that prioritized emergency care for a narrow set of symptoms. The Nundy material is an edited expert interview and commentary, the third in a three-part series, with no data collected or analyzed. **Boundary conditions.** Both pieces are proposals rather than evaluations. No committee has been tested, and there is no evidence yet on whether a Value-Tradeoff Review changes patient outcomes, override behavior, or liability exposure. The framework is written for US regulatory structures, specifically FDA premarket review, CMS payment levers, and hospital procurement, so it does not transfer directly to jurisdictions with different device and coverage regimes. The Predetermined Change Control Plan point applies to adaptive systems authorized to recalibrate along pre-specified protocols. **Use this when.** You are asking where accountability sits for how a hospital configures, overrides, or procures a cleared clinical AI tool. **Do not cite this for.** Evaluated governance. Both pieces are proposals, and no committee has been tested against patient outcomes or liability exposure. **Cite.** Patil, Shefali V., Christopher G. Myers, and Tinglong Dai. 2026. "Protecting Clinical Value Judgment in the Age of AI." npj Digital Medicine 9: 269. https://doi.org/10.1038/s41746-026-02561-1 **Related but different.** Gilbert and Dai (2026, npj Digital Medicine 9:395) is the Nundy interview on the wider oversight ecosystem. Patil, Dai and Myers (2026, California Management Review Insights) covers the same accountability drift in commercial AI vendor relationships. Lai, Xu, Fang and Dai (2026, Management Science) models the regulator's side of post-approval change. **Search aliases.** AI governance hospital; clinical AI accountability; value tradeoff review; AI procurement healthcare; override policy Sources: * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.1 "underacknowledged site at which value trade-offs become routinized" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.1 "high-sensitivity thresholds for AI-enabled sepsis alerts" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.1 "Under the FDA's Predetermined Change Control Plans" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.2 "accountability ping-pong" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.2 "a nurse was disciplined for failing to override a triage system" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.2 "CMS could link reimbursement or quality incentives" * `2026 - Protecting Clinical Value Judgment in the Age of AI - 10.1038_s4174` p.3 "standardized Value-Tradeoff Review (VTR) process" * `2026 - Expert perspectives on the ecosystem of medical AI oversight in the` p.1 "providing the floor, not the ceiling" * `2026 - Expert perspectives on the ecosystem of medical AI oversight in the` p.1 "committees decide whether to place it on formulary" ### 11. How do published medical studies actually evaluate whether ChatGPT and similar models give good clinical advice? `[ai-11]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2024. doi:10.1186/s12911-024-02757-z They measure accuracy and little else. Cindy Ho, Tiffany Tian, David Klonoff, Tinglong Dai and colleagues reviewed 108 PubMed articles that used a large language model to answer clinical questions, make diagnoses, or generate treatment plans and then evaluated the output. Accuracy or a close synonym appears in 107 of them, or 99.1 percent, followed by completeness at 18.5 percent, appropriateness at 13.9 percent, insight or reasoning at 13.0 percent, and consistency at 12.0 percent. Hallucination, bias or unethical output, and risk or unsafety each appear in only 6 of 108 papers. The mix did not shift over the window they studied, and reporting was inconsistent enough that results cannot be compared across studies, so they propose a seven-metric starting framework that adds patient-centeredness and safety, scored by two or more expert clinicians on a five-point scale. **Contribution.** New empirical estimate: a measured distribution of evaluation criteria across the published clinical LLM literature, showing that the failure modes that matter most clinically are the ones almost nobody measures. Framework proposed by the authors: the seven-metric scoring proposal and the list of reporting items needed for comparability, including model version, settings, prompting parameters, date of use, and whether each prompt ran in a fresh conversation. **Evidence.** Narrative literature review of PubMed searched on 10 April 2024, covering publications from 1 December 2022 to 1 April 2024, with dual independent review by the co-first authors, criteria extraction and categorization, and a Cochran-Armitage trend test for change over time. **Boundary conditions.** One database and one search query, so arXiv preprints and non-indexed journals are excluded and the window begins in December 2022. The counts describe what researchers measured, not how well the models performed. Fourteen of the studies evaluated models on medical board exams, a format that makes accuracy the easiest thing to score and therefore inflates its apparent importance. The trend test was not statistically significant, so the stability claim is an absence of detected change over a 1.5-year window rather than a demonstrated constant. Open-source models are underrepresented relative to proprietary ones, so the picture is weighted toward GPT-3.5 and GPT-4. **Use this when.** You need a taxonomy of how the published literature evaluates large language models in clinical decision-making beyond raw accuracy. **Do not cite this for.** Model performance. The counts describe what researchers measured, and the search covers one database from December 2022 onward. **Cite.** Ho, Cindy N., Tiffany Tian, Alessandra T. Ayers, Rachel E. Aaron, Vidith Phillips, Risa M. Wolf, Nestoras Mathioudakis, Tinglong Dai, and David C. Klonoff. 2024. "Qualitative Metrics from the Biomedical Literature for Evaluating Large Language Models in Clinical Decision-Making." BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-024-02757-z **Related but different.** Gilbert and Dai's 2026 npj Digital Medicine series covers regulatory oversight of LLMs in healthcare rather than how researchers score them. Dai and Abràmoff (2023, INFORMS TutORials) deals with workflow and payment for cleared diagnostic AI rather than with evaluation practice. **Search aliases.** clinical LLM evaluation; medical LLM benchmark; qualitative metrics AI; ChatGPT clinical assessment; LLM evaluation framework Sources: * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.1 "We selected 108 relevant articles from PubMed for analysis" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.6 "Accuracy / Agreement / Classification / Correctness" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.6 "Hallucination" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.3 "The Cochran-Armitage test for trend was used" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.6 "the easiest metric by which to evaluate LLM performance" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.6 "the exact model that was used, any relevant settings" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.8 "We propose that the following seven metrics" * `2024 - Qualitative metrics from the biomedical literature for evaluating l` p.9 "the search was limited to only one database" ### 12. Are ambient AI scribes saving clinicians time, or are they mainly raising what hospitals can bill? `[ai-12]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1038/s41746-025-02272-z Joe Kvedar, Dan Polsky and Tinglong Dai argue that tools sold on burnout grounds are increasingly shaping coding intensity and revenue capture, which puts them on a collision course with payers. Their concern is the dynamic that follows. Payers can answer with automated downcoding and risk-score recalibration, and once both sides have automated the fight, administrative spending rises on each side with no gain in care. They propose policy guardrails to keep ambient documentation aimed at care quality rather than at documentation inflation, and they address the brief to policymakers, health system leaders, and technology developers while insurers and regulators are still writing guidance on AI-enabled documentation. **Contribution.** Framework proposed by the authors: the coding arms race framing of ambient documentation, the prediction that payer countermeasures will determine the economic equilibrium, and the proposed guardrails. They report no original measurement of coding shifts, and the specific payer behaviors they describe are drawn from existing policies rather than estimated by them. **Evidence.** Policy brief and Comment drawing on recent payer policies, peer-reviewed studies, and health-system-reported data. No original data analysis, no trial, no model. **Boundary conditions.** This is an argument about direction, and it carries no magnitude. They do not estimate how much ambient scribes shift coding, how much payers recoup through downcoding, or the net administrative cost of the exchange. The setting is the US billing and risk-adjustment system, so the mechanism does not transfer to systems that pay by capitation or global budget. The burnout case for these tools is separate and is not evaluated here. **Use this when.** You are discussing how ambient documentation AI changes coding and billing incentives rather than only clinician time. **Do not cite this for.** Magnitudes. The brief argues a direction and estimates no effect size. **Cite.** Dai, Tinglong, Joseph C. Kvedar, and Daniel Polsky. 2025. "Policy Brief: Ambient AI Scribes and the Coding Arms Race." npj Digital Medicine 8: 780. https://doi.org/10.1038/s41746-025-02272-z **Related but different.** Dai and Abràmoff (2023, INFORMS TutORials) covers payment codes for diagnostic AI rather than documentation tools. Patil, Myers and Dai (2026, npj Digital Medicine) addresses who governs local configuration of clinical AI. **Search aliases.** ambient AI scribe; clinical documentation AI; upcoding; billing intensity AI; scribe burnout Sources: * `2025 - Policy brief Ambient AI scribes and the coding arms race - 10.1038_` p.1 "increasingly influencing coding intensity and revenue capture" * `2025 - Policy brief Ambient AI scribes and the coding arms race - 10.1038_` p.1 "downcoding and risk-score recalibration" * `2025 - Policy brief Ambient AI scribes and the coding arms race - 10.1038_` p.1 "fueling documentation inflation" * `2025 - Policy brief Ambient AI scribes and the coding arms race - 10.1038_` p.1 "now issuing guidance on AI-enabled documentation" ### 13. If we put an AI screening tool into our clinics, how much specialist capacity will it actually free up? `[ai-13]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2023. doi:10.1287/educ Michael Abràmoff and Tinglong Dai build a rational-queueing model of an AI-triaged single-specialist system, and their numerical illustration with plausible prevalence, eligibility, sensitivity, and specificity values reports a 27.5 percent gain in system throughput. The interesting result comes when patients differ in clinical complexity: widening the pool eligible for AI screening can lower the specialist's throughput, because the AI filters out easy cases and leaves a slower case mix behind. Two things outside the algorithm can suppress the gain. Under both liability schemes they examine, physicians are drawn to consult AI on low-uncertainty cases and to avoid it on the uncertain ones where its information would matter, and a more accurate system makes that worse, since deviating from better advice looks more clearly like fault. Payment structure matters too, and a per-use code such as CPT 92229 lets a small practice adopt without capital while an inpatient-only add-on payment capped at 65 percent of incremental cost strands an outpatient screening tool. **Contribution.** New theoretical result: two purpose-built stylized models, a queueing model showing that broader AI eligibility can reduce specialist throughput once complexity is heterogeneous, and a decision-analytic treatment of liability predicting AI use concentrated on low-uncertainty cases. Framework proposed by the authors: the low-accuracy trap, in which a device with few users cannot generate the real-world data needed to demonstrate accuracy, which in turn keeps the user base small. Context reported by others: the field evidence that point-of-care AI raised screening adherence from 49 percent to 95 percent in a primary-care study and that a randomized trial found 100 percent completion in the AI-offered group against 22 percent for referral, and the finding that roughly 75 percent of the 521 AI-enabled devices cleared as of July 2022 sit in radiology. **Evidence.** INFORMS TutORials review built around two stylized analytical models rather than field data, illustrated with the LumineticsCore autonomous diabetic retinopathy system. The throughput number is a model illustration under assumed parameters, not a measured outcome. **Boundary conditions.** The 27.5 percent figure comes from one parameterization of a single-specialist queue and should not be read as an expected gain for a real clinic. The complexity reversal requires heterogeneous patient complexity and enough easy cases for the AI to absorb. The liability predictions hold under the two schemes modeled and assume the physician weighs patient welfare alongside revenue and legal exposure. The reimbursement discussion is specific to US coding and payment rules. The extraction of this article flags an internal inconsistency in the numerical example and in the stated direction of the disease-prevalence effect, so the mechanism behind the throughput number should be read off the printed page before it is cited. **Use this when.** You need the queueing logic for how much specialist capacity an AI screening tool actually frees, including the complexity reversal. **Do not cite this for.** An expected gain for a real clinic. The headline percentage comes from one parameterization of a single-specialist queue. **Cite.** Dai, Tinglong, and Michael D. Abràmoff. 2023. "Incorporating Artificial Intelligence into Healthcare Workflows: Models and Insights." INFORMS TutORials in Operations Research, 133-155. https://doi.org/10.1287/educ.2023.0257 **Related but different.** Abràmoff, Dai and coauthors (2023, npj Digital Medicine) is the cluster-randomized trial measuring a 40 percent productivity gain in a real clinic, which is the empirical counterpart to this model. Dai and Singh (2025, POM) asks where in the workflow the AI should sit, before or after the clinician. Dai and Singh (2025, JMR) develops the liability and reimbursement model in full. **Search aliases.** AI workflow integration; specialist capacity; queueing model clinical AI; screening throughput; capacity freed by AI Sources: * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.8 "indicating a 27.5% increase in system throughput" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.10 "a greater proportion of high-complexity patients being seen by the specialist" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.12 "use AI for low uncertainty cases and avoid using AI" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.12 "as the precision of AI improves, the physician's tendency to avoid" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.15 "a low-demand AI device may fall into a low-accuracy trap" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.16 "the CPT code 92229, specifically created for autonomous AI" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.17 "maximum add-on payment is limited to 65% of the incremental cost" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.6 "had a screening rate of 100%, whereas the group referred" * `2023 - Incorporating Artificial Intelligence into Healthcare Workflows Mod` p.13 "approximately 75% of these devices are focused on a single medical specialty" --- ### 14. Can a medical AI model be made to forget a patient's data, and what does machine unlearning mean for privacy regulation and trust? `[ai-14]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1377/forefront Deleting a patient's records from a hospital database does nothing about what a trained AI model already learned from them. Tinglong Dai, Risa M. Wolf, and Haiyang Yang argue that this gap between data deletion and model memory has become a governance problem for medical AI. Machine unlearning, the targeted removal of a data point's influence from a trained model, exists mostly on paper. Retraining from scratch guarantees removal, yet for large language models it can take weeks and cost millions of dollars, which puts it out of reach for most clinical settings. Approximate methods trade one hazard for another: aggressive weight adjustment can erase useful knowledge, a failure known as catastrophic forgetting, while gentler adjustment can leave residual traces that violate privacy law. The authors then describe a cross-jurisdiction bind. The GDPR gives European patients an erasure right that can reach into a model's parameters. The FDA, which has cleared more than 1,000 AI-enabled medical devices, treats a cleared model as fixed and has never said whether unlearning counts as a major change. A US vendor serving European providers may satisfy one regulator only by drifting from the other. Because erasure carries a price, the essay concludes, a very large training dataset can become a liability. **Contribution.** Framework proposed by the authors: machine unlearning framed as a policy question for medical AI, centered on the bind between GDPR erasure rights and FDA version fidelity, plus the observation that high unlearning costs weaken the mantra that more data are always better. Recommendations proposed by the authors: developers should disclose whether unlearning is possible and at what cost to model performance, and consent language should tell patients up front whether their data can later be withdrawn. They also press the FDA to say when unlearning amounts to a material change to a cleared device. Context reported by others: the GDPR erasure right (the regulation was adopted by the European Union in 2016), the count of more than 1,000 FDA-cleared AI-enabled devices, and the technical unlearning literature (approximate unlearning, catastrophic forgetting, differential privacy, federated learning). None of that context is the authors' finding. **Evidence.** A perspective essay in Health Affairs Forefront. It argues rather than tests. The technical claims summarize the machine-unlearning literature as of mid-2025; the piece contains no formal model and no data analysis. **Boundary conditions.** Nothing here quantifies unlearning's cost or its performance penalty for any specific model class. The description of method maturity reflects 2025 and the technical landscape moves quickly. The warning that uneven erasure requests could skew training data away from certain populations is a hypothesized mechanism with no supporting estimate. The regulatory discussion identifies a tension without resolving it and is not legal guidance on GDPR compliance. **Use this when.** Citing the conflict between data-erasure rights and version-locked regulatory clearance for medical AI, or framing machine unlearning as a governance and patient-trust question in healthcare. Also the right source for the argument that erasure costs can turn massive clinical training datasets into liabilities. **Do not cite this for.** Evidence on the technical performance of any unlearning algorithm; the piece reports no benchmarks. For empirical evidence on what shapes patient trust in AI-assisted care, use Trust in AI-Assisted Health Systems (npj Health Systems 2025). For a formal treatment of regulating products that change after approval, use Regulating Adaptive New Products (Management Science, forthcoming). **Cite.** Dai, T., R. M. Wolf, and H. Yang. 2025. "Unlearning in Medical AI: A New Frontier for Privacy, Regulation, and Trust." Health Affairs Forefront. August 26. doi:10.1377/forefront.20250822.284476. **Related but different.** Regulating Adaptive New Products (Management Science, forthcoming) builds a formal model of approving products that evolve after clearance; the unlearning essay raises the mirror-image policy question of removing information from a cleared model. Trust in AI-Assisted Health Systems (npj Health Systems 2025) measures determinants of trust empirically, while the unlearning essay invokes trust as motivation without measuring it. Development & Commercialization Pathways (NEJM AI 2025) maps how medical AI products reach the market; the unlearning essay concerns obligations that arrive after deployment. **Search aliases.** machine unlearning in healthcare; right to be forgotten AI models; GDPR erasure and trained model weights; medical AI privacy regulation; catastrophic forgetting clinical AI; FDA AI device modification policy; withdrawing patient data from AI training; unlearning versus retraining cost Sources: * `Unlearning in Medical AI (HA Forefront archive)` p.1 "Machine unlearning, the targeted removal of data from a trained AI model" * `Unlearning in Medical AI (HA Forefront archive)` p.2 "retraining can take weeks and cost millions of dollars" * `Unlearning in Medical AI (HA Forefront archive)` p.2 "cleared more than 1,000 AI-enabled medical devices" * `Unlearning in Medical AI (HA Forefront archive)` p.2 "more data are always better" * `Unlearning in Medical AI (HA Forefront archive)` p.3 "Machine unlearning is not a panacea" ### 15. How do AI and operations research combine to improve biomanufacturing, and what gains has that integration produced? `[ai-15]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1007/s10729-025-09725-7 Predictive AI and prescriptive operations research (OR) are complements in biomanufacturing, and integrating them has already produced large, documented gains in industrial practice. That is the argument Tugce Martagan and Tinglong Dai advance in this Current Opinion piece. Biologic drugs are grown in living cells; a typical biopharmaceutical molecule contains more than 10,000 atoms where aspirin has 21, which makes processes acutely sensitive to raw materials and operating conditions and drives high batch-to-batch variability. AI supplies the predictive layer. OR converts predictions into structured, interpretable decisions and works with comparatively little data, a real advantage in data-poor bioprocess settings. The evidence the authors assemble comes from university-industry partnerships: AI forecasts of batch yield and failure risk, paired with stochastic-control timing of bleed-feed interventions, raised yield per bioreactor setup by 82 percent; Bayesian learning lifted bioreactor yield by 50 percent and cut batch-to-batch variability by 20 percent; together these projects generated 200 million dollars of additional drug output with no added raw materials or facility space, while using 40 percent less energy. The paper closes with a three-pillar roadmap spanning manufacturing, regulation, and technology, aimed at the industry's shift from batch to continuous production. **Contribution.** Framework proposed by the authors: a three-pillar roadmap (manufacturing, regulations, technology) that matches AI-OR capabilities to the challenges of continuous biomanufacturing, including real-time release testing under Quality by Design and digital twins for regulator-industry dialogue. Argument advanced by the authors: AI's predictive strength and OR's prescriptive strength complement each other, OR's low data requirements offset the data scarcity that limits AI in bioprocesses, and the field should move past human-in-the-loop arrangements toward collaborative human-AI decision-making supported by interdisciplinary talent. Context reported by others: every quantitative gain cited (the 82 percent bleed-feed result, the 50 percent yield and 20 percent variability improvements, the 200 million dollars of added output, the energy savings) comes from earlier peer-reviewed studies of industrial deployments by Martagan and colleagues, work the paper notes was recognized by the INFORMS Franz Edelman Award; Herbert Simon's 1987 call for AI-OR collaboration frames the piece. **Evidence.** A commentary in Health Care Management Science's Current Opinion format. It argues rather than tests. No new data were generated or analyzed; the numbers it reports are drawn from prior published studies of bleed-feed control and Bayesian process optimization at partner biomanufacturing firms. **Boundary conditions.** The scope is biopharmaceutical production processes; the authors explicitly leave drug-discovery R&D and supply chain management for future work. The reported gains are demonstrations from specific industrial partnerships, so the 82 percent and 50 percent figures should not be read as expected returns for a typical facility. Regulatory guidance for AI in drug manufacturing (FDA and EMA documents) was still exploratory at the time of writing, and the authors list adoption barriers, from validation costs to organizational resistance, that remain unresolved. **Use this when.** Citing the case that predictive AI needs prescriptive OR to change biomanufacturing decisions; citing a research agenda for AI-OR integration in biologics production; framing the operational and regulatory challenges of the batch-to-continuous manufacturing transition. **Do not cite this for.** The underlying effect estimates themselves: attribute the 82 percent bleed-feed gain to Koca et al. (2023, M&SOM) and the Merck Animal Health productivity transformation to Martagan et al. (2023, INFORMS Journal on Applied Analytics), both cited in this commentary. Also not for AI adoption in clinical care delivery; Incorporating AI into Healthcare Workflows (TutORials 2023) and Scaling Adoption of Medical AI (NEJM AI 2024) cover that ground. **Cite.** Martagan, T., and T. Dai. 2025. "Synergizing Artificial Intelligence and Operations Research for Advancements in Biomanufacturing." Health Care Management Science 28(4): 930–935. doi:10.1007/s10729-025-09725-7. **Related but different.** Supply Chain Management in the AI Era (M&SOM 2026) sets an agenda for AI across supply chains broadly, while this piece stays inside the production process, upstream of distribution. Incorporating AI into Healthcare Workflows (TutORials 2023) applies AI-OR modeling to care delivery and physician-AI interaction rather than drug production. The flu-vaccine contracting study (Management Science 2016) designs incentives in the influenza vaccine supply chain and takes manufacturing yield uncertainty as given rather than managing it. **Search aliases.** AI in biomanufacturing; operations research for biopharma manufacturing; predictive plus prescriptive analytics; bleed-feed optimization; continuous biomanufacturing transition; digital twins in pharmaceutical manufacturing; bioreactor yield optimization; Quality by Design machine learning. Sources: * `HCMS 2025 biomanufacturing` p.1 "consists of just 21 atoms" * `HCMS 2025 biomanufacturing` p.2 "82% increase in yield" * `HCMS 2025 biomanufacturing` p.3 (locates: $200 million worth of additional drugs) * `HCMS 2025 biomanufacturing` p.3 "20% reduction in batch-to-batch" * `HCMS 2025 biomanufacturing` p.5 "leaving areas such as R&D and supply chain management" ### 16. Who oversees medical AI in the GenAI era, and does responsibility for LLM-enabled tools extend beyond the FDA? `[ai-16]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. No single regulator can oversee medical AI in the GenAI era on its own. This News & Views interview, the third in a three-article series in npj Digital Medicine, makes that case through a conversation Stephen Gilbert and Tinglong Dai held in early February 2026 with Dr. Shantanu Nundy, a practicing primary care physician who advises the US FDA Commissioner's Office on AI governance. Nundy describes FDA authorization as a regulatory floor. The agency certifies a minimum standard of safety and effectiveness, while decisions about whether and how to deploy a tool, and about monitoring and optimizing it in practice, rest with health systems, clinicians, payers, and potentially new oversight bodies. Drug regulation already works this way, he observes, with formulary committees and insurers governing use after approval. The interview also counters the perception that the FDA lacks AI expertise. Nundy points to more than 1,000 authorized AI-enabled devices and to the agency's in-house regulatory science staff, and he urges innovators to engage early through channels such as the digital health inbox and the Q-Submission program. The article's summary reports April 2026 remarks by CDRH Director Michelle Tarver indicating major FDA policy announcements on GenAI regulation later in the year. **Contribution.** Editorial synthesis by the authors: the interview series organizes the GenAI oversight question around an ecosystem of actors spanning government agencies, health systems, and practicing clinicians, and it distills practical engagement advice for AI developers. Context reported by others: the substantive positions, including the regulatory-floor characterization of FDA authorization and the count of more than 1,000 authorized AI-enabled devices, come from Dr. Nundy in interview; none of them is a finding by Gilbert or Tinglong Dai. **Evidence.** An edited expert interview in the News & Views format, conducted in early February 2026 and annotated with references. It reports an expert's views and offers commentary; nothing in it is tested empirically. **Boundary conditions.** US regulatory setting in early 2026, a period of rapid policy movement; the summary itself notes April 2026 remarks by CDRH Director Michelle Tarver indicating further FDA announcements on GenAI regulation within the year, so specifics may date quickly. Dr. Nundy advises the FDA as a contractor, and the article states that the views expressed are his own and do not represent official FDA policy. Do not treat the piece as agency guidance or as a legal reading of device statute. **Use this when.** Citing the position that oversight of GenAI-enabled and LLM-enabled medical devices spans multiple actors beyond the FDA; describing the split between premarket authorization and downstream governance by health systems, clinicians, insurers, and other bodies; or directing developers to FDA engagement channels such as Q-Submissions and the digital health inbox. **Do not cite this for.** Official FDA policy on generative AI; the piece explicitly disclaims that status. Claims about the TEMPO pilot's design belong to the companion article "Expert Perspectives on Recent US Digital Medicine Regulation Policy Changes and the TEMPO Pilot," and the wellness-versus-device boundary for LLM tools is covered in "Expert Perspectives on US Regulatory Approaches to Large Language Models in Healthcare." Readers who need quantitative evidence on how AI devices actually reach the US market should cite Development and Commercialization Pathways (NEJM AI 2025) instead. **Cite.** Gilbert, Stephen, and Tinglong Dai. 2026. “Expert Perspectives on the Ecosystem of Medical AI Oversight in the GenAI Era.” npj Digital Medicine, 9: 395. https://www.nature.com/articles/s41746-026-02785-1. **Related but different.** "Expert Perspectives on Recent US Digital Medicine Regulation Policy Changes and the TEMPO Pilot" (Gilbert and Dai, npj Digital Medicine 2026) opens the same series and examines the FDA TEMPO pilot, which ties device oversight to the CMS ACCESS outcomes-based payment model and randomizes participants on a 90 to 10 intervention-to-control split. "Expert Perspectives on US Regulatory Approaches to Large Language Models in Healthcare" (Gilbert and Dai, npj Digital Medicine 2026) covers the wellness-device boundary and Nundy's two-dimensional lens, clinical severity crossed with tool autonomy, for judging which LLM applications warrant scrutiny; the present article takes up the multi-actor governance question instead. Development and Commercialization Pathways (Lee et al., NEJM AI 2025) supplies empirical analysis of AI device routes to market, where this interview offers qualitative perspective. **Search aliases.** who regulates medical AI in the US; FDA generative AI oversight; LLM medical device regulation; medical AI governance ecosystem; regulatory floor for AI devices; FDA Q-Submission digital health engagement; Shantanu Nundy FDA interview; health system responsibility for AI monitoring. Sources: * `Ecosystem GenAI oversight` p.1 "extends beyond the remit of any single regulator" * `Ecosystem GenAI oversight` p.1 (locates: authorized over 1000 AI-enabled devices) * `Ecosystem GenAI oversight` p.2 "sets a regulatory floor for safety" * `Ecosystem GenAI oversight` p.2 "major policy announcements from the FDA on GenAI" * `TEMPO pilot` p.2 (locates: 90 to 10 intervention to control split) ### 17. What do INFORMS Fellows think about the integration of AI and operations research? `[ai-17]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1287/ijds Senior operations researchers mostly welcome AI and want the field's mathematical rigor and identity to survive the boom. That is the picture from a 2024 survey of INFORMS Fellows analyzed by Holly Wiberg, Tinglong Dai, Henry Lam, and Radhika Kulkarni, who read the results against Herbert Simon's 1987 call to blend AI with operations research and management science (OR/MS). Forty-five Fellows responded, a 21.3% response rate from the 211 on the distribution list; 91.11% reported more than two decades in the profession. Most (69.2%) see the two fields as adjacent disciplines, and 60.0% say AI is extensively or moderately integrated into their own work. Confidence runs in both directions: 87.8% agree that OR/MS methodologies are fundamental to enhancing AI algorithms, and 71.4% expect AI advances to drive future innovation in OR/MS methodology. The worries are institutional. Student recruiting feels harder to 43.9% of respondents, and academics rate AI's role as positive far less often than practitioners (57.6% versus 88.9%). Only 11.4% call AI a threat to their subdiscipline, yet 88.4% want stricter ethical guidelines. The paper closes with an agenda: diversify funding beyond federal sources and build AI-OR crosstraining into doctoral education, using joint sessions at existing societies as the near-term venue. **Contribution.** New empirical estimate: the first systematic survey evidence on how INFORMS Fellows perceive the AI-OR/MS relationship, including the adjacency, integration, and attitude figures above. Framework proposed by the authors: an updated reading of Simon's 1987 agenda for the generative-AI era, with recommendations for professional societies and doctoral education. Context reported by others: Simon's original Interfaces article and the cited AI-for-optimization examples (machine learning for combinatorial optimization, LLM-based model formulation) are prior work by others, never findings of this survey. **Evidence.** A survey administered to INFORMS Fellows between September and October 2024, yielding 45 responses (21.3% of the 211 fellows with contact information), reported as descriptive percentages alongside quoted free-text answers. This is a reference and agenda piece: it interprets the survey and argues for integration rather than testing hypotheses. **Boundary conditions.** Respondents are Fellows, the field's most senior cohort; 34.88% are retired, and the authors state the sample is unrepresentative of the more than 12,000-member INFORMS community. The survey also predates the 2025 shifts in the U.S. federal funding environment, a caveat the authors themselves attach to the funding results. Percentages describe this small elite sample at one moment; do not read them as prevalence estimates for early-career researchers or for the field at large. **Use this when.** You need survey evidence on how the OR/MS community, especially its senior members, views AI; you want a citable bridge between Simon's 1987 vision and current AI-OR integration debates; or you are writing about INFORMS-level institutional strategy for the AI era, from conferences to education programs. **Do not cite this for.** Evidence that combining AI with optimization improves decisions in any deployed system; the survey measures perceptions. For AI agendas in specific operations domains, Supply Chain Management in the AI Era (M&SOM 2026) covers supply chains and Incorporating AI into Healthcare Workflows (TutORials 2023) covers healthcare delivery. Nor does it support claims about clinician attitudes toward medical AI, where Peer Perceptions (npj Digital Medicine 2025) is the right source. **Cite.** Wiberg, Holly, Tinglong Dai, Henry Lam, and Radhika Kulkarni. 2025. “Synergizing Artificial Intelligence and Operations Research: Perspectives from INFORMS Fellows on the Next Frontier.” INFORMS Journal on Data Science, ePub ahead of print. https://doi.org/10.1287/ijds.2025.0077. **Related but different.** AI and Operations (POM 2026) maps how AI changes operations management research questions, while this paper documents community attitudes with survey data. Supply Chain Management in the AI Era (M&SOM 2026) sets a domain-specific agenda for supply chains rather than a field-level one. Scaling Adoption of Medical AI (NEJM AI 2024) addresses governance of AI deployment in healthcare, a policy question the Fellows survey reaches only through general ethics attitudes. **Search aliases.** AI and operations research integration; INFORMS Fellows survey on AI; Herbert Simon two heads are better than one; OR/MS and machine learning synergy; future of operations research in the AI era; AI for optimization and optimization for AI; attracting students to operations research amid the AI boom. Sources: * `IJDS 2025.0077` p.3 "received responses from 45 fellows" * `IJDS 2025.0077` p.4 "adjacent disciplines (27 respondents; 69.2%)" * `IJDS 2025.0077` p.5 "(36; 87.8%) agree that OR/MS methodologies" * `IJDS 2025.0077` p.6 "18 respondents (43.9%) reporting greater" * `IJDS 2025.0077` p.7 "(38; 88.4%) calls for stricter ethical guidelines" # Global Supply Chains Tinglong Dai came to supply chains through inventory theory and stayed for the incentives. Tinglong Dai's earliest work in this area took up how a manufacturer should split orders across capacity-limited suppliers over a short season, and how badly the textbook infinite-horizon answer misleads when the horizon is one year. The question kept getting bigger. Zhaolin Li, Daewon Sun and Tinglong Dai studied what happens when a retailer's manager is paid in stock and investors read inventory as a signal of demand. A buyback contract that coordinates perfectly under symmetric information becomes a conduit for that manager's short-termism, and the supplier absorbs part of the loss. Give the firm a menu of contracts and it can signal by choosing terms rather than by buying inventory nobody needs. Soo-Haeng Cho, Fuqiang Zhang and Tinglong Dai looked at the US influenza vaccine chain, where shortages happen in seasons of plenty because doses show up late. The manufacturer carries the design risk of producing before the strain is fixed while the provider captures most of the value of an on-time shipment, so timeliness gets underfunded, orders shrink, and the loop tightens. A partial buyback paired with a late-delivery rebate breaks it. Measurement is the second thread. Christopher Tang and Tinglong Dai have argued for years that ESG scores grade a boundary the firm gets to choose. Push the dirty work onto a supplier and the number improves while the world does not. Hau Lee joined them to propose extending the Scope 1/2/3 construct past carbon and aggregating ESG across a network by position rather than by weighted sum. The third thread is geopolitics, which arrived whether the field wanted it or not. Tang and Tinglong Dai read de-risking through four flows: material, information, financial, and human. Money and people moved long before goods did. Mariana Socal, Maqbool Dada, colleagues at Hopkins and Tinglong Dai have been tracing where American drug supply actually comes from, and the answer sits one step upstream of where policy looks. Finished antibiotics arrive from many countries. The chemistry inside them comes mostly from China. Tariffs aimed at foreign drugmakers would tax American ones. Most recently Tinglong Dai helped organize forty-odd operations researchers and practitioners around a statement on AI and supply chains. They laid out five layers, and the one that binds is infrastructure. Data centers wait years for grid access. The models are not the constraint. ## Questions ### 1. How is AI actually changing supply chain management, and where is the hype outrunning the evidence? `[supply-chains-01]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1287/msom They organize AI's effect on supply chains into five layers: intelligence, execution, strategy, human, and infrastructure. AI adds the most when it works inside the structure operations management already supplies, since better prediction does not convert proportionally into better decisions when small forecast errors flip discrete choices. They are deliberately deflationary about forecasting, because most supply chain cost sits in the labor and energy of storing and moving goods, and a warehouse shipping in cases of 24 cannot act on the difference between 63.5 and 65.3 predicted units. What they expect to bind hardest is the infrastructure layer, where data centers wait two to four years for grid access and specialized tooling suppliers can run past twelve months. **Contribution.** Framework proposed by the authors: the five-layer architecture of an AI-enabled supply chain (intelligence, execution, strategy, human, infrastructure), the argument that AI creates value inside operations-management structure rather than by displacing it, and the diagnosis that paired technical debt and organizational debt strand pilots. Context reported by others: the benchmarking evidence that machine-learning forecasters do not consistently beat classical statistical methods, the experimental result that managers blame agents for correct algorithm overrides, and the retail workforce-management evidence on store-manager overrides are findings from work the article cites, not their own estimates. **Evidence.** OM Forum vision statement in Manufacturing & Service Operations Management: a structured synthesis by 42 operations-management researchers, industry practitioners, and technology leaders, organized into five layers and illustrated with sector case studies. No original model, dataset, or estimation, so every quantitative statement in it is sourced from elsewhere. **Boundary conditions.** This is a consensus position piece rather than a test of anything. The deflationary claim about forecasting applies to high-volume physical distribution where holding and movement costs dominate and shipment granularity is coarse; it says nothing about settings where forecast error itself is the dominant cost. The claim that AI has not yet produced a structural shift comparable to paved roads or standardized containers is a statement about the record as of writing, and they say the potential exists and may arrive in ways nobody anticipates. **Use this when.** You need a community research agenda for AI in supply chain management, including where the forecasting claims outrun the evidence. **Do not cite this for.** A test of anything. It is a consensus position piece, and the deflationary forecasting claim is scoped to high-volume physical distribution. **Cite.** Cohen, Maxime C., Tinglong Dai, Georgia Perakis, et al. 2026. "OM Forum: Supply Chain Management in the AI Era: A Vision Statement from the Operations Management Community." Manufacturing & Service Operations Management 28 (3): 687-705. https://doi.org/10.1287/msom.2025.1065 **Related but different.** Wuest, Kusiak, Dai and Tayur (2020), "Impact of COVID-19 on Manufacturing and Supply Networks": a much shorter early-pandemic position piece arguing against procyclical cuts to AI and digital spending, with no layered framework. **Search aliases.** AI supply chain; generative AI supply chain; agentic supply chain; digital supply chain research agenda; supply chain forecasting AI Tinglong Dai and coauthors (2026), "Assured Autonomy: How Operations Research Powers and Orchestrates Generative AI Systems": about governing autonomous generative-AI systems across four domains, with supply chains as one case, rather than about supply chain management as a whole. Sources: * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.2 "five layers through which AI interacts" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.3 "included 42 OM" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.4 "Improvements in prediction accuracy" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.7 "shipped in cases of 24 units" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.7 "Paved roads and standardized shipping containers" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.10 "grid access delays of two to four years" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.11 "do not consistently outperform classical statistical approaches" * `2026 - Supply Chain Management in the AI Era A Vision Statement from the O` p.15 "technical debt and organizational debt" ### 2. Is US-China supply chain de-risking actually working, and what should I watch besides trade in goods? `[supply-chains-02]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2024. doi:10.1353/asp Christopher Tang and Tinglong Dai read de-risking through four flows: material, information, financial, and human. Goods are the slowest of the four, so anyone judging the policy from import statistics alone will see the least informative signal. Published data show the money and the people moved first: net foreign direct investment into China fell to $33 billion in 2023, and the number of Chinese students at US universities dropped from 370,000 in 2019 to roughly 290,000 in 2024. They also insist that this is a two-sided interaction, since China's domestic-content push predates recent Western measures, and diversification carries its own risk when a 2021 McKinsey survey found only 2 percent of companies could see past their second-tier suppliers. **Contribution.** Framework proposed by the authors: the four-flow lens (material, information, financial, human) applied to US-China de-risking, and the argument that de-risking is an interactive, two-sided shift rather than a Western policy applied to a passive China. Context reported by others: the FDI series for China, Mexico, Vietnam and India, the product-level import-share shifts, the 3.4 percent change in US manufacturing employment from 2018 to 2023, the student and expatriate counts, and the McKinsey supplier-visibility figure are all published third-party statistics the essay assembles, not their estimates. **Evidence.** Policy essay in Asia Policy. A qualitative four-flow framework applied to published trade, FDI, tariff, and migration statistics together with government and industry sources. No formal model and no original estimation, so none of the numbers here is a causal estimate of what de-risking policy caused. **Boundary conditions.** Scoped to US-China de-risking as of mid-2024, and the FDI, trade-share, and student figures are point-in-time snapshots that later data can overturn. The four-flow structure is an organizing device they propose, not a tested theory with parameters. Product-level reshuffling is described as partly cosmetic, since the same firms relocating assembly can move an import share without changing who owns the capability. **Use this when.** You need the four-flow framing of de-risking across material, information, financial and human flows. **Do not cite this for.** Current data. The investment, trade-share and student figures are mid-2024 snapshots, and the four-flow structure is an organizing device rather than a tested theory. **Cite.** Dai, Tinglong, and Christopher S. Tang. 2024. "De-Risking Global Supply Chains: Looking beyond Material Flows." Asia Policy 19 (4): 153-176. https://doi.org/10.1353/asp.2024.a942841 **Related but different.** Socal, Sun, Ballreich, Lambert, Dai and Dada (2025), JAMA Health Forum: puts measured import data behind one sector's China dependence, where this essay works across sectors qualitatively. **Search aliases.** de-risking; decoupling; friendshoring; China plus one; economic security supply chain; reshoring policy Tinglong Dai and Tang (2022), Service Science: the same coauthor pair on supply chain measurement, aimed at ESG ratings rather than geopolitics. Sources: * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.4 "increase domestic content in eleven key sectors to 30% by 2020" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.10 "pressure the ride-sharing company Didi to delist from the New York" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.13 "FDI in China fell to $33 billion on a net basis" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.17 "universities fell from 370,000 in 2019 to about 290,000 in 2024" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.20 "manufacturing employment only increased by a modest 3.4%" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.22 "number of direct links between the West and China is decreasing" * `2024 - De-risking Global Supply Chains Looking Beyond Material Flows - 10.` p.23 "only 2% of companies reported visibility beyond their second-tier suppliers" ### 3. Why can a company improve its ESG rating without changing anything except which supplier does the polluting? `[supply-chains-03]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1287/serv Because nearly every rating grades the operations a firm controls directly, which makes the measurement boundary something the firm chooses. Christopher Tang and Tinglong Dai named five obstacles that keep ESG measures from crossing that boundary, starting with supply chain opacity and ending with inconsistent law and enforcement. Henn (2016) documented the mechanism concretely at ExxonMobil, which reported cutting greenhouse gas emissions while its actual emissions rose, achieved by moving dirty operations onto supply chain partners. The social pillar is the worst of the three, and practitioners agree: in the 2019 BNP Paribas ESG Global Survey, 46 percent of respondents called it the hardest element to incorporate into investment analysis. **Contribution.** Framework proposed by the authors: the diagnosis that ESG measurement stops at the firm boundary and therefore rewards outsourcing, the five named obstacles to supply-chain-aware ESG, and a research agenda for the social pillar built on three COVID-era supply chain cases. Context reported by others: the ExxonMobil emissions-shifting case is documented by Henn (2016); the contradictory Chevron ratings from Refinitiv and Sustainalytics come from Eaglesham and Shifflett (2021); the Global Health Security Index ranking, the BNP Paribas survey figure, and the ESG asset totals are third-party sources they cite. **Evidence.** Conceptual review and research agenda in Service Science, organized around three COVID-era cases (online platforms, public health supply chains, and vaccine development and distribution), plus a structured search of five OM journals for papers mentioning ESG. No original data and no estimation. **Boundary conditions.** The argument is about how ratings are constructed, so it holds wherever a score grades the reporting entity rather than the network. Where suppliers already fall inside a firm's consolidated reporting, the outsourcing channel closes. The journal search covers five OM journals through 2021 and says nothing about the finance or accounting literatures, and the case selection is pandemic-era and deliberately illustrative rather than representative. **Use this when.** You are arguing that firm-level ESG scores reward outsourcing emissions because they grade the reporting entity rather than the network. **Do not cite this for.** The Exxon example, the eleven-fold supply-chain emissions figure, or the rating-disagreement examples, all of which are cited from other sources. **Cite.** Dai, Tinglong, and Christopher S. Tang. 2022. "Integrating ESG Measures and Supply Chain Management: Research Opportunities in the Post-Pandemic Era." Service Science 14 (1): 1-12. https://doi.org/10.1287/serv.2021.0295 **Related but different.** Dai, Lee and Tang (2024), "Toward Supply-Chain-Aware ESG Measures": the follow-on chapter that builds the measurement and aggregation machinery this paper calls for. **Search aliases.** Scope 3 emissions; supply chain ESG; ESG rating disagreement; sustainable sourcing; value chain emissions Tinglong Dai and Tang (2022), "Unifying ESG and Supply Chain Thinking" (Asia Global Institute): a longer companion treatment of the same argument with fuller case material. Tinglong Dai and Tang (2024), "Natural Hazards and Supply Chain": turns the same lens on disaster risk and the emissions feedback into hazards. Sources: * `2022 - Integrating ESG measures and supply chain management Research oppor` p.1 "we identified 15 such papers with the majority published" * `2022 - Integrating ESG measures and supply chain management Research oppor` p.2 "(1) supply chain opacity, (2) ambiguous relationship" * `2022 - Integrating ESG measures and supply chain management Research oppor` p.3 "shifting dirty operations to its supply chain partners" * `2022 - Integrating ESG measures and supply chain management Research oppor` p.5 "ranked the first among 195 countries in the Global Health Security" * `2022 - Integrating ESG measures and supply chain management Research oppor` p.9 "Chevron Corp. earned the best rating from Refinitiv" * `2022 - Integrating ESG measures and supply chain management Research oppor` p.9 "Nearly half (46%) of respondents feel that this is the case" * `2022 - Unifying ESG and Supply Chain Thinking An Urgent Call to Action in ` p.4 "being held responsible for their entire supply chain on social issues" ### 4. How could you build an ESG score that does not reward a company for outsourcing its emissions? `[supply-chains-04]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2024. doi:10.1007/978-3-031-60867-4_15 Hau Lee, Christopher Tang, and Tinglong Dai propose two moves. Extend the Scope 1/2/3 construct past carbon so that social and governance impacts are also traced into the supplier network, and use tiered metrics so a firm cannot pool its own workforce with its suppliers' to flatter a diversity number, which one company they visited did to report a workforce over 70 percent female while leadership stayed predominantly male. Then aggregate by network position instead of by weighted sum. Their illustrative PageRank-style computation on six firms leaves the upstream chip makers at their original scores while Best Buy, farthest downstream, falls from 90 to roughly 60.7 and Dell from 80 to roughly 50.7. **Contribution.** Framework proposed by the authors: extending the Scope 1/2/3 construct across all three ESG pillars, tiered social metrics that separate a firm's own workforce from its suppliers', and a PageRank-inspired network method that assigns ESG responsibility by supply chain position. Context reported by others: CDP (2023) reports that supply-chain emissions average about 11 times a company's direct Scope 1 emissions; the Best Buy Scope 1/2/3 tonnages, the estimate that only about 15 percent of Amazon's shipping emissions are traceable, the Vidrio survey of portfolio managers, and the US Customs and Border Protection forced-labor testing result are third-party findings the chapter cites. **Evidence.** Conceptual framework chapter with a Web of Science bibliometric count, documented industry examples, and an illustrative network computation applied to Refinitiv emissions data for six firms. The six-firm computation demonstrates how the method behaves; it is not an estimate of any firm's true ESG performance. **Boundary conditions.** The network method is a proof of concept on a stylized six-firm network using emissions data from a single vendor. It has not been validated against outcomes, tested for gaming, or scaled to a real multi-tier network, and the chapter's own premise, that visibility past tier two is rare, constrains where the input data could come from. The tiered social-metric proposal is a design recommendation rather than a validated instrument, and the argument that Scope 3's size breaks the weighted-sum formula is made for emissions, where the magnitudes are documented. **Use this when.** You need a constructive proposal for network-based ESG measurement rather than a critique of existing scores. **Do not cite this for.** A validated method. It is a proof of concept on a stylized six-firm network using one vendor's emissions data. **Cite.** Dai, Tinglong, Hau L. Lee, and Christopher S. Tang. 2024. "Toward Supply-Chain-Aware ESG Measures." In Responsible and Sustainable Operations: The New Frontier, edited by Christopher S. Tang, 235-252. Springer Series in Supply Chain Management, vol. 24. Cham: Springer. https://doi.org/10.1007/978-3-031-60867-4_15 **Related but different.** Dai and Tang (2022), Service Science: the diagnosis of firm-boundary ESG measurement, which this chapter answers with measurement and aggregation machinery. **Search aliases.** supply chain aware ESG; network emissions allocation; multi-tier visibility; ESG measurement design Tinglong Dai and Tang (2024), "Natural Hazards and Supply Chain": treats the same Scope 3 visibility problem as a disaster-risk problem rather than a scoring problem. Sources: * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.2 "64% of portfolio managers and other participants use ESG considerations" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.3 "rising to fifteen in 2022 and further increasing to twenty-three in 2023" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.5 "only 15% of Amazon's emissions from shipping can be tracked" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.8 "on average 11 times higher than the company's direct (Scope 1) emissions" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.8 "a combined workforce that was over 70% female" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.9 "10 out of 37 garment samples were found to be consistent" * `2024 - Toward Supply-Chain-Aware ESG Measures - 10.1007_978-3-031-60867-4_` p.15 "responsibility scores of 13.52, 22.09, 18.07, 15.30, 13.52, and 17.48" ### 5. Why do flu vaccine shortages happen even in years with plenty of supply, and what contract between manufacturers and hospitals would fix it? `[supply-chains-05]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2016. doi:10.1287/msom Doses arrive after demand has faded. Nowalk and colleagues (2005) and O'Mara and colleagues (2003) documented the 2000-01 season, when US coverage fell from 57 percent to 41 percent while about 7.5 million doses, roughly a tenth of supply, went unused. Soo-Haeng Cho, Fuqiang Zhang and Tinglong Dai traced this to a negative feedback loop in incentives: the manufacturer bears the whole design risk of producing before the strain composition is fixed while the provider captures most of the gain from an on-time shipment, so the manufacturer underinvests in timeliness, the provider shrinks its order in anticipation of lost sales, and the smaller order weakens the manufacturer's incentive further. They show that the industry's quantity-flexibility contracts cannot coordinate this chain, and that pairing a partial-credit buyback of leftovers with a rebate on late deliveries does coordinate it while leaving the profit split fully negotiable. **Contribution.** New theoretical result: the negative incentive feedback loop between at-risk early production and provider ordering; the impossibility of coordination under an ordinary quantity-flexibility contract in this setting; a buyback-plus-late-rebate contract that coordinates the chain with an arbitrary profit split; and the welfare finding that a per-unit subsidy paid only to the manufacturer cannot reach the social optimum when a provider orders and then distributes. Context reported by others: the 2000-01 coverage decline and the 7.5 million unused doses are figures from Nowalk et al. (2005) and O'Mara et al. (2003), which they cite as motivation. **Evidence.** Game-theoretic supply chain contracting model in which the manufacturer chooses between an at-risk early production mode and a regular mode under design, delivery, and demand uncertainty. Disciplined by actual contracts collected from two influenza vaccine manufacturers and two academic medical centers, with numerical evaluation calibrated to industry prices, plus extensions to social welfare, trivalent vaccines, and random yield. **Boundary conditions.** A single manufacturer selling to a single provider, with risk-neutral parties and a newsvendor demand structure. Coordination requires the contract to be more generous on late-delivered doses than on on-time ones, which is exactly what the standard quantity-flexibility form cannot do. The late-rebate contract coordinates only in a knife-edge case that gives the retailer the entire chain profit and the full design risk, which requires dominating retailer bargaining power and is unrealistic against a large manufacturer. The welfare result about manufacturer-only subsidies is specific to chains where a provider orders and then distributes, and departs from prior results derived for direct-to-consumer settings. **Use this when.** You need the negative incentive feedback in vaccine delivery timing and the contract form that coordinates it. **Do not cite this for.** Multi-manufacturer settings. It is a single manufacturer selling to a single provider under newsvendor demand. **Cite.** Dai, Tinglong, Soo-Haeng Cho, and Fuqiang Zhang. 2016. "Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supply Chain." Manufacturing & Service Operations Management 18 (3): 332-346. https://doi.org/10.1287/msom.2015.0574 **Related but different.** Liljenquist, Dai and Bai (2021), Population Health Management: same industry and same lateness problem, answered with a nonprofit public utility manufacturer and a split FDA strain-selection calendar rather than a contract redesign. **Search aliases.** influenza vaccine supply; on-time delivery contract; vaccine shortage; quantity flexibility contract; delivery timing incentives Tinglong Dai and Song (2021), Health Care Management Science: vaccination delivery operations downstream of the manufacturer, covering cold chain, syringes, and appointment matching. Tinglong Dai, Li and Sun (2012), Decision Sciences: also a buyback-based coordination problem, but the friction is signaling to investors rather than delivery timing. Sources: * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.1 "7.5 million vaccine doses, or 10.6% of the total supply, remained" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.2 "doses shipped before October 15:" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.3 "negative feedback loop in the firms’ incentives" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.7 "implies that the QF contract can never coordinate the" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.9 "in which case the retailer fully bears the risk associated" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.10 "The BLR contract coordinates the sup-" * `2016 - Contracting for On-Time Delivery in the U.S. Influenza Vaccine Supp` p.12 "it is not possible to achieve the social optimum by" ### 6. How do stock-based executive incentives distort a retailer's inventory decisions, and can supply chain contracts be redesigned so the signaling does not hurt the supplier? `[supply-chains-06]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2012. doi:10.1111/j When a manager's pay is tied to the interim share price and investors read order quantity as evidence of demand strength, the manager of a strong firm buys more inventory than cash flow alone would justify, purely to separate the firm from a weak one. Zhaolin Li, Daewon Sun and Tinglong Dai show that a single coordinating buyback contract makes this everyone's problem, because the two parties split chain profit in fixed proportions and the deadweight loss of the signal travels upstream to the supplier. A plain wholesale-price contract can beat the coordinating buyback in this world, since over-ordering partly cancels the understocking caused by double marginalization. The clean fix is a menu of two coordinating buyback contracts, which lets the firm signal by choosing terms and restores the first best for any level of equity-based pay. **Contribution.** New theoretical result: characterization of the separating equilibrium in which equity-based pay drives a strong retailer to over-order; the demonstration that a single coordinating buyback contract transfers that deadweight loss upstream and can be worse than a plain wholesale-price contract; and a menu of two coordinating buyback contracts that attains the first best at any level of equity-based incentive. Context reported by others: the rise in equity-based pay from 37 percent to 55 percent of top-five executive compensation at S&P 500 firms between 1993 and 2003 is a published statistic they cite as motivation. **Evidence.** Game-theoretic signaling model: a three-period inventory signaling game with asymmetric demand information, separating-equilibrium characterization refined by the intuitive criterion, multidimensional screening for the menu design, and a numerical illustration. Theory throughout, with no data and no estimation. **Boundary conditions.** The over-ordering result holds in the separating equilibrium and only once the manager's weight on the interim share price passes a threshold; below that the distortion does not appear, and the weak firm orders its unconstrained optimum in any case. The wholesale-price contract beats the coordinating buyback only when the manager weights share price heavily enough for over-ordering to offset double marginalization. The menu result requires the fixed charge to sit in a range that deters the weak type while still attracting the strong one, and under the profit-maximizing menu the supplier earns strictly positive surplus, which is the firm's signaling cost and a pure wealth transfer rather than an efficiency loss. **Use this when.** You are analyzing how stock-based executive pay distorts inventory decisions and how buy-back contracts respond. **Do not cite this for.** An unconditional over-ordering claim. It holds in the separating equilibrium once the manager's weight on the interim share price passes a threshold. **Cite.** Dai, Tinglong, Zhaolin Li, and Daewon Sun. 2012. "Equity-Based Incentives and Supply Chain Buy-Back Contracts." Decision Sciences 43 (4): 661-685. https://doi.org/10.1111/j.1540-5915.2012.00363.x **Related but different.** Dai, Cho and Zhang (2016), M&SOM: another buyback-based coordinating contract, where the friction is at-risk early production and delivery timing rather than signaling to investors. **Search aliases.** equity incentives inventory; signaling supply chain; buy-back contract; managerial short-termism ordering Sources: * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.2 "increased to 55% by 2003" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.3 "exceeds a threshold a strong firm's manager may over-order" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.11 "the pooling equilibrium is not stable" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.12 "can transfer the deadweight loss from downstream to upstream" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.14 "which is 13.9% higher" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.16 "Hence, the first-best can be achieved" * `2012 - Equity-Based Incentives and Supply Chain Buy-Back Contracts - 10.11` p.17 "would leave some money on the table to the supplier" ### 7. How dependent is the United States on China for its antibiotic supply? `[supply-chains-07]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1001/jamahealthforum The dependence sits one step upstream of where most people look. Mariana Socal, Tinglong Dai and colleagues analyzed 33 years of US customs records, and finished antibiotic doses arrive from a healthily diversified set of countries, led by India at 31.9 percent of volume over 2020 to 2024. Active pharmaceutical ingredients look nothing like that: China supplied 62.6 percent of imported API volume across those five years and 70.1 percent in 2024 alone, with API import concentration crossing the high-concentration threshold in 2008 and climbing past a Herfindahl-Hirschman Index of 5000 by 2024. The visible diversification downstream may itself be illusory, since the countries shipping them finished antibiotics may be buying their chemistry from China, which trade data cannot see. **Contribution.** New empirical estimate: country-level concentration of US antibiotic imports measured separately for finished dosage forms and active pharmaceutical ingredients from 1992 through 2024, including China's 62.6 percent share of API volume for 2020-2024 and 70.1 percent in 2024, Europe's fall from 75.7 percent of API imports in 1992 to 13.1 percent in 2024, and the sharp divergence between volume shares and cost shares by country. Context reported by others: the estimate that antibiotics are 42 percent more likely than other drug classes to go into shortage, and the 2004 closure of the last US penicillin API plant, are prior findings the paper cites. **Evidence.** Cross-sectional analysis of US Customs importation records from USA Trade Online, January 1992 through December 2024, measuring volume in metric tons, inflation-adjusted cost, and country-level market concentration via the Herfindahl-Hirschman Index. Descriptive and associational; the study estimates no causal effect. **Boundary conditions.** Antibiotics only, and import records only. Trade data cannot see the origin of APIs embedded in finished doses arriving from India or Italy, so measured diversification of finished-form sourcing may overstate genuine diversification. Volume shares and cost shares diverge sharply by country, so the answer depends on which is used, and import prices are not US acquisition prices, since the paper finds no matching decline in the National Average Drug Acquisition Cost. The paper's warning that tariffs could worsen shortages is a policy inference, not an estimate. **Use this when.** You need US import-data estimates separating dependence on Chinese active ingredients from dependence on Chinese finished doses. **Do not cite this for.** The origin of ingredients embedded in finished doses arriving from India or Italy, which trade data cannot observe. **Cite.** Socal, Mariana P., Yunxiang Sun, Jeromie M. Ballreich, Jennifer Dailey Lambert, Tinglong Dai, and Maqbool Dada. 2025. "US Antibiotic Importation and Supply Chain Vulnerabilities." JAMA Health Forum 6 (10): e253871. https://doi.org/10.1001/jamahealthforum.2025.3871 **Related but different.** Socal, Sun, Ballreich, Acha, Yazdi, Dai and Dada (2026), Health Affairs Scholar: uses similar import data to project tariff-driven price changes rather than to measure concentration. **Search aliases.** antibiotic supply chain; API dependence China; pharmaceutical import; drug shortage security; generic antibiotic sourcing Socal, Acha, Yang, Sun, Dada, Tinglong Dai, Anderson and Ballreich (2025) on oncology drug shortages: traces shortage causes drug by drug, where this paper measures upstream country concentration. Tinglong Dai and Tang (2024), Asia Policy: the same dependence question at the level of whole economies rather than one drug class. Sources: * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.2 "antibiotics are 42% more likely to experience shortages" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.5 "indicating a highly concentrated market since 2008" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.5 "India was the leading originating country for antibiotic FDFs" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.5 "China accounted for 70.1% of API imports" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.5 "In 1992, Europe supplied 75.7% of API imports" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.6 "Italy was the originating country for 2.6% of the total imported volume" * `2025 - US Antibiotic Importation and Supply Chain Vulnerabilities - 10.100` p.7 "may be an underestimation of the US dependency on China" ### 8. Would tariffs on imported pharmaceutical ingredients raise the price of American-made generic drugs? `[supply-chains-08]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1093/haschl/qxaf247 Yes, and that is the part most tariff arguments miss. Mariana Socal, Maqbool Dada, Tinglong Dai and colleagues modeled 17 generic APIs against six years of US import records linked to 2024 Medicaid prices. Under base-case assumptions a flat 100 percent worldwide API tariff raises finished-drug prices by an average of 30 percent, about $21.15 per prescription, while a blended tariff matching announced country-specific rates raises them roughly 10 percent, about $6.22. The burden falls unevenly by molecule because sourcing is concentrated by molecule: hydrocortisone and tetracyclines, largely Chinese-sourced, would rise 15.0 and 13.1 percent under a China-only tariff and under 1 percent under a Europe-only tariff. **Contribution.** New empirical estimate: scenario estimates of how API tariffs would move the price of generic drugs manufactured in the United States, broken out by tariff design and by molecule, from linked import and Medicaid price data, together with the finding that the API's share of the finished drug price drives the result more than the pass-through rate does. Framework proposed by the authors: the tariff impact calculator mapping API import value and country of origin into finished-drug price changes. Context reported by others: the shares of US prescriptions and spending accounted for by generics, and the sourcing statistics for individual molecules, come from published sources the paper cites. **Evidence.** Scenario modeling with sensitivity analysis. Six years of US Census Bureau import records (2019-2024) linked to 2024 Medicaid State Drug Utilization prices for the 17 generic APIs meeting inclusion criteria. These are model-based projections under stated assumptions, not observed price changes. **Boundary conditions.** The population is generic drugs manufactured in the United States using imported APIs. The estimates do not generalize to drugs made with US-produced APIs, to branded drugs, or to imported finished dosage forms. Base-case results assume full pass-through; if supply chain participants absorbed even a quarter of the added cost, the increases fall by roughly half. Only 17 APIs met inclusion criteria, so molecule-level results are not a census of the generic market, and the paper flags Chinese export restrictions as a separate risk that tariff modeling does not capture. **Use this when.** You need modeled price effects of ingredient tariffs on generic drugs manufactured in the United States from imported active ingredients. **Do not cite this for.** Drugs made with US-produced ingredients, branded drugs, or imported finished dosage forms. **Cite.** Socal, Mariana P., Yunxiang Sun, Jeromie Ballreich, Joy Acha, Mohammad Ali Yazdi, Tinglong Dai, and Maqbool Dada. 2026. "Potential Impact of Tariffs on Active Pharmaceutical Ingredients on the Price of U.S.-Made Generic Drugs." Health Affairs Scholar 4 (2): qxaf247. https://doi.org/10.1093/haschl/qxaf247 **Related but different.** Dada, Dai, Sun and Socal (2025), "Tariffs as a Hidden Tax" (working paper): models absorption stage by stage through a four-stage domestic chain and finds far smaller consumer price effects, under different assumptions about pass-through and about which stage is tariffed. **Search aliases.** pharmaceutical tariff; API tariff; generic drug price; drug manufacturing policy Socal, Dada and Dai (2025), Health Affairs Scholar: whether tariffs relocate manufacturing, rather than what they do to prices. Socal, Sun, Ballreich, Lambert, Tinglong Dai and Dada (2025), JAMA Health Forum: measures the underlying import dependence this scenario model prices out. Sources: * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.1 "average price increase of 30% (additional $21.15 per prescription)" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.3 "China accounted for 44.4% of the total imported value" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.4 "80% under the world 100% tariff if 80% of the price" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.4 "$96.10 higher price for penicillin G" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.6 "Hydrocortisone and tetracyclines, whose imported" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.6 "price increases would be about" * `2026 - Potential Impact of Tariffs on Active Pharmaceutical Ingredients on` p.7 "may inadvertently undermine Made-in-America" ### 9. How much of a tariff on imported drug ingredients actually reaches the pharmacy counter, and does the government come out ahead? `[supply-chains-09]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.2139/ssrn Very little reaches the counter, because each domestic stage absorbs part of the shock. In their base case, with a 25 percent markup at each of three domestic stages, half the cost increase passed on at each stage, and an active ingredient worth about $4 inside a $20 drug, a 25 percent tariff on that ingredient raises the final consumer price by about 0.32 percent, so a patient with 20 percent coinsurance sees a copayment move from $4.00 to about $4.02. The fiscal side is where the surprise sits. That 25 percent tariff generates $1.00 of gross revenue per unit, and after roughly $0.20 of lost corporate tax at a 20 percent rate and about $0.06 of higher public procurement when government programs buy half the volume, the government nets about $0.75 per unit, roughly 75 percent of the headline amount. **Contribution.** New theoretical result: a condition under which government net receipts from a tariff turn negative, which they show depends neither on the tariff rate nor on the imported good's cost. New empirical estimate: calibrated pass-through and net-revenue figures for US pharmaceutical supply chains across four tariff scenarios. Framework proposed by the authors: the four-stage supply chain model with stage-level markups and partial pass-through, and the demonstration that percentage versus fixed-dollar markup conventions determine the entire inflationary footprint of a trade shock. Context reported by others: the car floor mat and wine retail-pricing episodes are documentary cases they draw on rather than original data. **Evidence.** Analytical model of a four-stage supply chain with stage-level markups and partial pass-through, calibrated with 2019-2024 US pharmaceutical import data and run across four tariff scenarios, supported by documentary case evidence. Working paper, not yet peer reviewed. This is calibration, not measurement of realized prices. **Boundary conditions.** The headline numbers are calibrated to a specific illustrative case: an active ingredient worth roughly $4 inside a $20 drug, a 25 percent markup at each of three domestic stages, and half of each cost increase passed on at each stage. The muted price effect depends on the ingredient being a small share of the finished price and on intermediaries holding percentage markups; under fixed dollar markups the retail arithmetic changes completely. The fiscal accounting assumes a 20 percent corporate tax rate and government programs buying half the volume. Tariffing finished dosage forms removes one absorbing domestic stage and raises the retail effect several fold, to between 0.04 and 1.59 percent for a 25 percent tariff. **Use this when.** You need the multi-stage pass-through logic explaining why a large ingredient tariff can reach the pharmacy counter as a small price change. **Do not cite this for.** Observed pass-through. The headline numbers are calibrated to a roughly four-dollar ingredient inside a twenty-dollar drug with stated markup and pass-through assumptions. **Cite.** Dada, Maqbool, Tinglong Dai, Yunxiang Sun, and Mariana Socal. 2025. "Tariffs as a Hidden Tax: Price Pass-through in Multi-Stage Supply Chains." Working paper. https://doi.org/10.2139/ssrn.5237643 **Related but different.** Socal, Sun, Ballreich, Acha, Yazdi, Dai and Dada (2026), Health Affairs Scholar: the peer-reviewed scenario model of API tariffs, which assumes full pass-through for its base case and therefore reports much larger price effects for US-made generics. **Search aliases.** tariff pass-through; multi-stage supply chain pricing; hidden tax; input tariff incidence; drug pricing policy Socal, Dada and Dai (2025), Health Affairs Scholar: the relocation question rather than the pass-through question. Sources: * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.4 "China (98.72%)" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.6 "a much more modest increase of 17% despite a 125% tariff" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.9 "a 0.32% increase in the final price" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.11 "increases by only about 0.01% to 0.39%" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.11 "the government nets approximately $0.75 per unit" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.12 "rise between 0.04% and 1.59%" * `2025 - Tariffs as a Hidden Tax Price Pass-Through in Multi-stage Supply Ch` p.13 "independent of both the tariff rate and the cost of the imported finished good" ### 10. Would tariffs on imported drugs actually bring pharmaceutical manufacturing back to the United States? `[supply-chains-10]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2025. doi:10.1093/haschl/qxaf122 Branded and generic makers face opposite arithmetic, so a single tariff pushes them in opposite directions. Branded firms have margin and a legal opening: converting an API into a finished dose often fails the substantial-transformation test that establishes country of origin, so a US-made branded API can go abroad for cheap final processing and return as a finished drug without incurring a tariff. Generic production is labor-intensive with thin margins and no pricing cushion, so with Mariana Socal and Maqbool Dada Tinglong Dai concluded that reliance on foreign suppliers is likely to continue unless tariff-driven global price increases outweigh the amortized costs of relocation and higher domestic production costs. If the goal is domestic capacity, grants, low-interest loans, tax credits, and advance purchase commitments do the work that tariffs cannot. **Contribution.** Framework proposed by the authors: the split-market analysis of tariff incidence and relocation incentives, showing why branded and generic manufacturers respond in opposite directions, and the identification of the substantial-transformation rule as the specific mechanism giving branded firms an onshoring incentive for APIs. Context reported by others: the finding that more than 92 percent of facilities making generic APIs for the US market are overseas, the record of 323 active drug shortages in early 2024, the generic shares of US prescriptions and spending, and the estimate that APIs run 20 to 30 percent of drug manufacturing cost are all published statistics they cite. **Evidence.** Economic policy analysis in Health Affairs Scholar of tariff incidence and relocation incentives across two market structures, reasoning over published industry data and trade-law rules. No model estimation and no original dataset, so the predictions are conditional reasoning rather than measured effects. **Boundary conditions.** Stated for US pharmaceutical markets under the tariff proposals in force in 2025. The generic prediction is explicitly conditional: foreign sourcing continues unless tariff-driven global price increases outweigh the amortized costs of relocation plus higher domestic production costs. The branded API incentive depends on the substantial-transformation test continuing to treat final dose conversion as insufficient to confer origin. Short-run pass-through is muted by distributor and group purchasing organization contracts that can shield buyers for one to three years, so near-term and long-run price paths differ, and because the tariffs came by executive order rather than congressional mandate, the resulting policy uncertainty itself weakens any relocation incentive. **Use this when.** You are asking whether pharmaceutical tariffs would relocate manufacturing to the United States. **Do not cite this for.** A prediction independent of policy. The generic-drug result is conditional on tariff-driven global price increases outweighing relocation and higher domestic production costs. **Cite.** Socal, Mariana, Maqbool Dada, and Tinglong Dai. 2025. "Prescription for Made in America? Tariffs and U.S. Drug Manufacturing." Health Affairs Scholar 3 (7): qxaf122. https://doi.org/10.1093/haschl/qxaf122 **Related but different.** Socal, Sun, Ballreich, Acha, Yazdi, Dai and Dada (2026), Health Affairs Scholar: quantifies what API tariffs would do to the price of US-made generics, where this piece treats the relocation decision. **Search aliases.** pharmaceutical reshoring; drug manufacturing tariffs; onshoring generics; supply chain industrial policy Dada, Tinglong Dai, Sun and Socal (2025), "Tariffs as a Hidden Tax" (working paper): the pass-through and fiscal accounting behind the same policy. Socal, Sun, Ballreich, Lambert, Tinglong Dai and Dada (2025), JAMA Health Forum: measures where the import dependence sits by drug class. Sources: * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.1 "yet account for more than 80% of pharmaceutical spending" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.1 "Over 92% of the facilities manufacturing generic APIs" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.1 "reaching a record high of 323 active instances in early 2024" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.2 "Without much pricing cushion, generic drug manufacturers are unlikely to relocat" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.2 "required to establish the country of origin of a drug for tariff purposes" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.2 "protect buyers from price increases for 1-3 years" * `2025 - Prescription for made in America Tariffs and U.S. drug manufacturin` p.2 "grants, low-interest loans, tax credits" ### 11. Why couldn't the US government find out where its N95 masks were made during COVID-19, and what rules would fix that? `[supply-chains-11]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2020. doi:10.1007/s11606-020-05987-9 Because nobody was required to say. Ge Bai, Gerard Anderson and Tinglong Dai searched five years of financial disclosures from the three largest US PPE manufacturers together with more than 1,700 media reports published between 14 January and 26 April 2020, and they found no basic supply chain data, not even the split between domestic and foreign N95 capacity. Their two proposals attach to instruments that already exist: require publicly traded PPE manufacturers to disclose production capacity and the percentage imported from each foreign country in their Form 10-K, and stress test manufacturers holding more than 10 percent US market share the way banks have been tested since 2008, with stockpile procurement keyed to the results. They also warn that firm-level testing alone will not be enough, since the nonwoven fibers that N95 masks are made from are manufactured exclusively in China, a vulnerability no single manufacturer controls. **Contribution.** New empirical estimate: an original document search of five years of SEC filings from the three largest US PPE manufacturers plus more than 1,700 media reports from January to April 2020, establishing that no domestic-versus-foreign N95 capacity data was disclosed anywhere. Framework proposed by the authors: mandatory Form 10-K disclosure of PPE production capacity and country-level import shares, plus bank-style stress testing of manufacturers above a 10 percent US market share with stockpile procurement keyed to the results, and classification of N95 masks as strategic items under the Kraljic scheme. Context reported by others: the estimate that about 90 percent of N95 masks used in the US are imported, the estimate that clinicians accounted for nearly 20 percent of US COVID-19 cases, the February 2020 congressional testimony on stockpile levels, and the Taiwanese consortium's 15 million daily masks are third-party reports they cite. **Evidence.** Policy commentary in the Journal of General Internal Medicine with original document analysis of manufacturer disclosures and media reports. The core finding is a documented absence of information, established by search rather than by estimation, and the piece makes no causal claim. **Boundary conditions.** Scoped to US personal protective equipment and to what was publicly disclosed as of spring 2020, so the finding is about disclosure rather than about what firms internally knew. The stress-test proposal is written for manufacturers above a 10 percent US market share, and they note that manufacturer-level testing cannot reach inputs outside any single firm's control. The proposals were never implemented, so nothing in the paper evaluates their effect, and the comparative country evidence is illustrative rather than a matched design. **Use this when.** You need the case for supply chain data transparency and stress testing, grounded in what was publicly knowable about protective-equipment origin in 2020. **Do not cite this for.** What firms internally knew. The finding concerns public disclosure, and the stress-test proposal targets manufacturers above a ten percent US market share. **Cite.** Dai, Tinglong, Ge Bai, and Gerard Anderson. 2020. "PPE Supply Chain Needs Data Transparency and Stress Testing." Journal of General Internal Medicine 35 (9): 2748-2749. https://doi.org/10.1007/s11606-020-05987-9 **Related but different.** Dai, Zaman, Padula and Davidson (2021), Journal of Clinical Nursing: argues supply chain capability belongs alongside surveillance and vaccines as a named pillar of global health preparedness, broader in scope and without the disclosure mechanics. **Search aliases.** PPE supply chain; supply chain transparency; stress testing supply chains; medical supply resilience; N95 sourcing Tinglong Dai and Tang (2022), Service Science: uses the same PPE failure as evidence that ESG scoring stops at the firm boundary. Tinglong Dai and Tang (2024), "Natural Hazards and Supply Chain": places the PPE episode in a general disruption-management framework. Sources: * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.1 "more than 1,700 media reports about the PPE supply chain" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.1 "an estimated 90% of N95 masks are imported" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.1 "clinicians account for nearly 20% of the COVID-19-infected cases" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.1 "at least 300 million N95 masks were required" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.1 "such as N95 masks, are classified as strategic items" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.2 "Such disclosure can be included in Form 10-K" * `2020 - PPE Supply Chain Needs Data Transparency and Stress Testing - 10.10` p.2 "nonwoven fibers, the key raw material for N95 masks, are exclusively manufacture" --- ### 12. Did the war in Ukraine reshape global supply chains more durably than COVID-19 did? `[supply-chains-12]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. That was Tinglong Dai's argument at the time, and the reasoning was about kind rather than degree. The pandemic hit capacity and logistics, and capacity recovers. The war forced a political sorting that does not. Russia is under 2 percent of global GDP and Ukraine is 0.14 percent, so the direct exposure is narrow and concentrated: energy, where Russia supplied close to 40 percent of Europe's natural gas and 65 percent of Germany's while ranking third among oil exporters, and food, where Russia and Ukraine together accounted for roughly a third of global wheat exports, with countries including Kazakhstan and Tanzania drawing more than 90 percent of their wheat from Russia. The durable change is elsewhere. Tinglong Dai argued the world was moving toward a supply chain Iron Curtain, with Russia and its allies on one side and the West on the other. Belt and Road would have to shrink, since Russia and Ukraine are both critical links in it. Decoupling from China would accelerate for critical goods such as semiconductor materials, medical supplies and electric batteries. ESG pressure would push firms to avoid geopolitical risk in advance, including by exiting an entire economy. Thomas Friedman's 1996 proposition that economically integrated countries do not fight each other did not survive. The line Tinglong Dai would still defend is that companies can no longer separate business from geopolitics. **Contribution.** Framework proposed by the author: the argument that a geopolitical rupture reshapes supply chains more permanently than a demand or capacity shock, the supply chain Iron Curtain framing, and the four reshaping mechanisms are his. Context reported by others: every quantitative figure here, the energy shares, the oil ranking, the wheat export shares, and the crude price move, comes from public reporting cited in the piece rather than from any analysis of Tinglong Dai's own. **Evidence.** Opinion essay for a general audience. No model, no data collection, no new estimates. Written three weeks into the invasion. **Boundary conditions.** Written in March 2022 and explicitly forward-looking, so it should be read as a forecast made at that moment rather than as a description of what followed. The quantitative shares are point-in-time and have since moved, particularly European gas sourcing. Treat the mechanisms as hypotheses that later work can test, which is what the 2024 de-risking paper begins to do. **Use this when.** You want Tinglong Dai's early public position that the war marked a structural rather than cyclical break in global supply chains, or the supply chain Iron Curtain framing. **Do not cite this for.** Any empirical estimate. It is an op-ed, and its numbers belong to the outlets it cites. For a research treatment of the same question, cite Dai and Tang (2024, Asia Policy). **Cite.** Dai, Tinglong. 2022. "Russia's War with Ukraine Could Permanently Reshape the Global Supply Chain." Fast Company, March 15, 2022. Originally published in The Conversation, March 11, 2022, as "Ukraine War and Anti-Russia Sanctions on Top of COVID-19 Mean Even Worse Trouble Lies Ahead for Global Supply Chains." **Where this framing first appeared.** The Iron Curtain formulation reached print a week before this essay. The Associated Press quoted him on it on March 4, 2022, in "Russia's invasion of Ukraine leaves global trade in tatters," where Tinglong Dai argued that the bedrock of global trade since the Cold War had been a separation of geopolitics from business, together with an assumption that rational decision-making would prevail, and that a new sort of Iron Curtain was forming. The essay below develops the argument; the AP story is where it was first attributed to him. The Conference Board's Committee for Economic Development took the framing up directly: on May 10, 2022 its Trade and Economic Globalization Committee held a Trustee briefing titled "War in Ukraine: Is a New Economic Iron Curtain Next?", which Tinglong Dai gave. **Related but different.** Dai and Tang (2024, Asia Policy) develops the four-flow framework for de-risking, which is the research version of this argument. Dai and Tang (2022, Newsweek) makes a companion case a month later. Cohen, Dai, Perakis and coauthors (2026, M&SOM) takes up what AI changes about all of this. **Search aliases.** Ukraine war supply chain; Russia sanctions supply chain; supply chain Iron Curtain; geopolitics and business; Belt and Road disruption; geopolitical risk sourcing; decoupling; wheat and energy exports Sources: * Fast Company, Mar 15 2022, republished from The Conversation, Mar 11 2022 "the end of the end of history" framing, Fukuyama * The Conversation, Mar 11 2022 "a new type of supply chain Iron Curtain" * The Conversation, Mar 11 2022 Belt and Road "will almost certainly need to scale back in size and scope" * The Conversation, Mar 11 2022 "Global supply chains, like the rest of the world, will never be the same again" * Associated Press, Mar 4 2022 quoted on "a new sort of 'Iron Curtain'", identified as a Johns Hopkins business professor who studies supply chains * The Conference Board / CED, Apr 27 2022 program document Trustee briefing scheduled for May 10, 2022, "War in Ukraine: Is a New Economic Iron Curtain Next?" ### 13. Why does reshoring to the United States keep stalling when investors, customers and regulators all say they want it? `[supply-chains-13]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. Christopher Tang and Tinglong Dai set out four obstacles, and none of them is solved by wanting reshoring more. First, the arithmetic often fails outside subsidized sectors. Personal protective equipment is the cautionary case: hospital executives told them throughout the pandemic that they valued resilience and preferred American-made supply, domestic mask makers invested on the strength of that, and the demand evaporated once inexpensive imported masks returned to the market, leaving layoffs behind. Stated preference and purchasing behavior parted company. Their suggestion was to stop competing on commodity output and build products that justify a premium, such as masks that are reusable, better fitted or biodegradable. Second, the business model does not follow the factory home. A reshored manufacturer should sell responsiveness rather than unit cost, making just-in-time credible again for domestic clients and even guaranteeing it by subscription, should automate hard while the technological advantage over China still exists, and should treat the domestic plant as an export base rather than an inward-looking one. Third, nobody can see far enough down their own supply chain to prove domestic content, which undercuts any price premium and any serious risk or ESG assessment. Fourth, Wall Street's preference for asset-light businesses penalizes the balance sheet of any firm that builds a factory, which is a large part of why semiconductors were invented in the United States and fabricated elsewhere. Their recommendation there was to make company valuation reflect supply chain risk and the speed of product development that proximity buys. **Contribution.** New empirical estimate: the count of corporate earnings calls mentioning reshoring, 106 in the first eight months of 2022 against six in the same period of 2019, is their own analysis of Capital IQ transcript data and appears first in this piece. Framework proposed by the authors: the four obstacles and the corresponding remedies. Context reported by others: the Reshoring Initiative figures on companies intending to reshore and jobs created, the Kearney 2021 Reshoring Index sentiment numbers, and the finding that only 2 percent of firms report visibility beyond their second-tier suppliers, which is a 2021 McKinsey survey. **Evidence.** Opinion essay for a general audience, drawing on the authors' research, on conversations with hospital executives during the pandemic, and on one original count of earnings-call mentions. No model and no formal identification. **Boundary conditions.** Written in October 2022, before the full effect of the Inflation Reduction Act and CHIPS Act incentives could be observed, and before the 2025 tariff regime. The earnings-call count measures how often executives talk about reshoring, which is not a measure of how much reshoring occurred. The PPE episode is a single well-documented case rather than a representative sample. **Use this when.** You need a structured account of why reshoring intent does not convert into reshoring, particularly the gap between stated preference for domestic sourcing and actual purchasing, or the argument that asset-light valuation norms penalize manufacturing investment. **Do not cite this for.** Estimates of how much manufacturing actually returned, or for tariff effects, which are Socal, Dada and Dai (2025) and the related pharmaceutical work. **Cite.** Dai, Tinglong, and Christopher S. Tang. 2022. "Everybody Talks About Made in America. But It Isn't That Simple." Wall Street Journal, October 23, 2022. **Related but different.** Socal, Dada and Dai (2025, Health Affairs Scholar) asks the same question for pharmaceuticals with data rather than argument. Dai, Bai and Anderson (2020) is the underlying transparency case, and the visibility obstacle here is its practical consequence. Dai and Tang (2024, Asia Policy) reframes the whole question as de-risking rather than reshoring. **Search aliases.** reshoring; Made in America; nearshoring; manufacturing onshoring; supply chain visibility; asset-light valuation; just-in-time versus just-in-case; domestic manufacturing policy Sources: * Wall Street Journal, Oct 23 2022 authors' own count of earnings-call mentions, 106 in 2022 against six in 2019, from Capital IQ transcripts * Wall Street Journal, Oct 23 2022 PPE reshoring case, domestic mask investment followed by demand collapse * Wall Street Journal, Oct 23 2022 2 percent of firms report visibility beyond tier-two suppliers, 2021 McKinsey survey * Wall Street Journal, Oct 23 2022 asset-light valuation norms as a barrier to reshoring investment # Markets, Platforms and Marketing-Operations Most of what Tinglong Dai does in this area starts from a constraint that sales-contract models usually assume away. You cannot sell what you do not have. Once inventory can bind, the compensation plan and the stocking decision stop being separate problems, and the answers change in ways Tinglong Dai did not anticipate. Kinshuk Jerath and Tinglong Dai took this up first in 2013. A firm may stock above the first-best level, because inventory that never runs out keeps high demand from being censored, and a firm that can see demand learns whether the salesperson worked. The extra units buy information. In a short Operations Research note they then showed that limited inventory rescues the quota-bonus contract from a nonexistence result the theory had carried for years, since the optimal quota turns out to be the stock itself. Later they let the firm hide its own supply-side actions, and the optimal commission plan went concave. Practitioners keep reporting regressive plans, and this gives them a rationale with no risk aversion in it. Rongzhu Ke, Christopher Ryan and Tinglong Dai handed both jobs to one person. The optimal plan takes a shape they call a mast and a sail. It is monotone in sales and can still withhold a bonus from a manager who delivered better inventory outcomes. It also pays, in effect, for throwing stock away. Capturing even a noisy trace of turned-away customers, a waiting list will do, repairs that. Then timing. Ruiting Zuo, Jussi Keppo and Tinglong Dai put price and commission on the same clock. Under dynamic pricing a fixed commission gives up little. Freeze both levers and the loss is large. The ranking reverses when price sits at a ceiling, whether regulation put it there or the firm's own overbooking exposure did. Two other strands live here. Yifu Li, Xiangtong Qi and Tinglong Dai worked out why the best part of an experience belongs in the middle when it is the part customers remember, and why a low point should sit right beside it. The OM Forum Tinglong Dai wrote with Ying-Ju Chen, Gizem Korpeoglu, Ersin Körpeoğlu, Ozge Sahin, Christopher Tang and Shihong Xiao sorts online platforms into five business models and names how they fail. One failure mode is growth itself, since a larger user base can add search friction faster than it adds matches. Health insurance is a price mechanism, so that work sits here too. A per-visit copayment and a percentage coinsurance rate push testing intensity in opposite directions, which makes patient cost sharing a clumsy instrument for anyone hoping to curb unnecessary imaging. ## Questions ### 1. How should a retailer pay a store manager who both drives demand and keeps shelves stocked, when the sales lost to stockouts are never observed? `[markets-01]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2021. doi:10.1287/mnsc Rongzhu Ke, Christopher Ryan and Tinglong Dai study a manager who splits effort between creating demand and keeping inventory usable, while the firm sees only sales, capped by whatever sat on the shelf. They find the optimal plan pays a base salary plus a bonus over a two-part region they call a mast and a sail: the bonus arrives when inventory clears and the stock level passes a threshold, or when inventory fails to clear and sales beat a target that itself falls as realized inventory rises. That plan is monotone in sales and can fail to be monotone in inventory, so a manager can post better numbers on both dimensions and still lose the bonus. It also opens an ex post hole, since writing off or hiding stock to fake a stockout can earn the bonus, and a firm that can gauge unsatisfied demand even noisily, through a waiting list for instance, gets back a monotone plan where that incentive disappears. **Contribution.** New theoretical result: the mast-and-sail characterization of the optimal multitask plan, its failure of joint monotonicity, the resulting incentive to dispose of inventory, and the monotone repair under a noisy signal of unmet demand are all results of this paper. They derive them with a bang-bang optimal-control approach for a multitask moral-hazard problem with a censored output signal. **Evidence.** Game-theoretic principal-agent model, no data. A risk-neutral, limited-liability agent allocates effort across marketing and operations; output is censored because demand above available inventory is never observed. Solved over a finite action set using information-trigger contracts, supplemented by numerical experiments that compare candidate contract forms. **Boundary conditions.** This is a multitasking moral-hazard setting with limited liability, risk neutrality, a finite action set, and a monotone likelihood ratio structure on the output distributions. Nonmonotonicity in inventory arises when the sail threshold sits above the mast threshold, and not otherwise. The monotone repair requires the firm to observe some signal of turned-away customers, however noisy. The result that the firm overstocks relative to the classical newsvendor level belongs to the extension where the firm also chooses the initial stocking quantity, and the mechanism there is that more inventory lowers the probability of clearing and so lowers the expected bonus payout. **Use this when.** You need the optimal incentive structure for an agent who both drives demand and manages inventory when sales lost to stockouts go unobserved. **Do not cite this for.** Settings without the monotone likelihood ratio structure. Nonmonotonicity in inventory arises only when the sail threshold sits above the mast threshold. **Cite.** Dai, Tinglong, Rongzhu Ke, and Christopher Thomas Ryan. 2021. "Incentive Design for Operations-Marketing Multitasking." Management Science 67 (4): 2211-2230. https://doi.org/10.1287/mnsc.2020.3651. **Related but different.** Dai and Jerath (2013, Management Science) has a single task and studies the bonus and the stocking level jointly. Dai and Jerath (2016, Operations Research) concerns whether a quota-bonus contract exists at all. Dai and Jerath (2019, Marketing Science) adds hidden action on the firm's side instead of a second task for the agent. **Search aliases.** multitasking incentives; operations marketing interface; store manager compensation; unobserved lost sales; moral hazard inventory Sources: * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.3 "resembles a "mast" and "sail"" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.11 "neither monotone in i nor jointly monotone" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.11 "rewarded for disposing of inventory" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.17 "ex post moral hazard hiding of inventory that afflicted" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.19 "overinvests in inventory as compared with the clas-" ### 2. Is a bonus that requires hitting both a sales quota and an inventory quota ever the optimal way to pay a store manager? `[markets-02]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2021. doi:10.1287/mnsc In the multitasking setting Rongzhu Ke, Christopher Ryan and Tinglong Dai study, a plan of that form, which they call a corner plan, cannot be optimal once the output distributions satisfy the strict monotone likelihood ratio property and the agent earns positive rents. How much the shortfall costs depends on the agent's cost structure. In representative numerical scenarios in the paper, corner plans came within roughly one percent of the optimum when the marketing and operational tasks were highly complementary in that cost structure, and fell short by as much as eighteen percent when they were not. They also show the efficiency loss grows linearly in the unit revenue rate, so the worst-case gap has no bound. **Contribution.** New theoretical result: the impossibility of corner plans in a multitask model with a censored sales signal, proved under strict MLRP with strictly positive agent rents. The one percent and eighteen percent figures are theirs as well, and they come from representative numerical scenarios reported in the paper. They are not general bounds and should not be quoted as such. **Evidence.** Game-theoretic principal-agent model plus numerical experiments, no data. Corner plans are compared against the optimal plan across cost structures that vary in how complementary the two tasks are. A separate numerical comparison evaluates a monotone repair that removes the mast and turns the sail into a weighted-sum threshold on sales and inventory; it does about as badly as corner plans in the unfavorable cases. **Boundary conditions.** The impossibility argument needs the strict monotone likelihood ratio property on the output distributions and strictly positive agent rents. Drop either and it does not go through. The percentages describe particular numerical scenarios, and the near-optimal case requires strong complementarity between the two tasks in the agent's cost function. The unbounded worst case is a property of the model as the unit revenue rate grows. **Use this when.** You need the result that a bonus contingent on hitting both a sales quota and an inventory quota is never optimal in this multitasking setting. **Do not cite this for.** Quota-bonus contracts in single-task salesforce settings, which is Dai and Jerath (2016, Operations Research). The impossibility needs strict MLRP and strictly positive agent rents. **Cite.** Dai, Tinglong, Rongzhu Ke, and Christopher Thomas Ryan. 2021. "Incentive Design for Operations-Marketing Multitasking." Management Science 67 (4): 2211-2230. https://doi.org/10.1287/mnsc.2020.3651. **Related but different.** The mast-and-sail question above comes from the same paper and covers the optimal plan's shape. Dai and Jerath (2013, Management Science) is the contrast worth knowing: with a single selling task and censored demand, a quota-bonus contract is optimal. **Search aliases.** dual quota bonus; corner plan impossibility; multitasking contract; joint quota compensation Sources: * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.12 "an optimal compensation plan cannot be a corner compensation plan" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.13 "close to optimal (within 1%)" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.14 "worst-case loss in performance of corner com-" * `2021 - Incentive Design for Operations-Marketing Multitasking - 10.1287_mn` p.15 "in bad cases (losses of up to 18%" ### 3. Why would a company pay its salespeople declining commission rates at higher sales levels? `[markets-03]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2019. doi:10.1287/mksc Kinshuk Jerath and Tinglong Dai trace it to moral hazard on the firm's side. Hold the firm's inventory-related action fixed and the optimal contract is extreme and convex, paying the salesperson only when sales hit the top outcome, even when a bad inventory draw caused the shortfall. Make that action endogenous and hidden from the salesperson, and the optimum starts paying a positive bonus at the intermediate outcome too, because the rep needs some assurance the firm will not quietly cut corners on inventory to shrink its expected payout. When revenue per unit is low enough the plan turns concave, with the top outcome worth barely more than the middle one, and no risk aversion enters the story. **Contribution.** New theoretical result: the concavity of the optimal commission plan under double moral hazard, and the finding that making the firm's supply-side action observable restores the extreme top-outcome-only bonus. The literature's standard explanation for regressive plans is agent risk aversion, and they identify a different force. Context reported by others: the figure that U.S. firms spend roughly three times as much on salesforce compensation as on advertising motivates the problem and is not their estimate. **Evidence.** Stylized principal-agent model, no data. Risk-neutral salesperson with limited liability, three-point demand and inventory distributions satisfying a monotone likelihood ratio property, hidden effort on the agent side and a hidden inventory action on the firm side. The paper also compares contracting before yield is realized against contracting after. **Boundary conditions.** Concavity arises when revenue per unit is low enough; at high unit revenue the plan stays convex. Everything rests on the three-outcome distributional structure with MLRP and a risk-neutral, limited-liability agent. If the firm's inventory-related action is observable to the salesperson, the firm always chooses the more effective action and the simple top-outcome-only bonus returns. The early-versus-late contracting comparison has two-sided costs, so its comparative statics move with the probability of the best inventory outcome, with how effective the salesperson is, and with unit revenue. **Use this when.** You need the explanation for concave or declining commission rates arising from hidden firm-side supply actions rather than agent risk aversion. **Do not cite this for.** High unit-revenue settings, where the plan stays convex. **Cite.** Dai, Tinglong, and Kinshuk Jerath. 2019. "Salesforce Contracting under Uncertain Demand and Supply: Double Moral Hazard and Optimality of Smooth Contracts." Marketing Science 38 (5): 852-870. https://doi.org/10.1287/mksc.2019.1171. **Related but different.** Dai and Jerath (2013, Management Science) keeps the firm's action observable and asks how the stocking level itself changes. Dai and Jerath (2016, Operations Research) is about existence of the quota-bonus form. Zuo, Dai and Keppo (2026, M&SOM) moves the same question into continuous time with a price lever. **Search aliases.** salesforce commission; double moral hazard; concave compensation; declining commission rate; supply uncertainty contracting Sources: * `2019 - Salesforce Contracting Under Uncertain Demand and Supply Double Mor` p.8 "optimal for the firm to use an extreme, convex reward structure" * `2019 - Salesforce Contracting Under Uncertain Demand and Supply Double Mor` p.9 "positive bonus for medium sales outcome is there to protect the salesperson" * `2019 - Salesforce Contracting Under Uncertain Demand and Supply Double Mor` p.10 "the optimal compensation plan may be concave when the per-unit revenue" * `2019 - Salesforce Contracting Under Uncertain Demand and Supply Double Mor` p.11 "we identify a force different from risk aversion" ### 4. My company already uses dynamic pricing. Is it also worth making sales commissions change over time, or is a fixed commission good enough? `[markets-04]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2026. doi:10.1287/msom Ruiting Zuo, Jussi Keppo and Tinglong Dai model a firm selling perishable inventory through a commissioned agent whose selling effort it cannot observe, with both price and commission adjustable continuously over a finite horizon. Once the price moves dynamically, holding the commission fixed costs the firm little, while freezing both levers costs a lot. In the paper's numerical example, the profit gap between the fully dynamic strategy and dynamic pricing paired with an optimized fixed commission stayed below thirteen percent. In a separate calibration to airline demand estimates, dynamic pricing with a fixed commission captured roughly ninety-nine percent of the fully dynamic strategy's value and beat dynamic contracting on its own, which matches the industry pattern of sophisticated fare management alongside commissions that barely move. **Contribution.** New theoretical result: the joint continuous-time design of price and agent incentive, including the finding that the dynamic contract attains the first-best outcome because the firm can revise incentives continuously as inventory and time change. New numerical estimate: the below-thirteen-percent gap in the generic example and the roughly ninety-nine-percent capture rate in the airline calibration are both theirs, and they come from two different exercises. Context reported by others: travel agencies accounted for 44 percent of airline ticket sales as recently as 2019, which motivates the setting. **Evidence.** Continuous-time principal-agent model with stochastic optimal control over a finite horizon, solved numerically by a tree-grid method. Two distinct numerical exercises: a generic parameterized example, and a calibration to airline-industry demand estimates. No econometric estimation and no causal identification. **Boundary conditions.** The ranking of the two levers is conditional. When the price effect on demand is weak, dynamic contracting with a static price beats a fixed commission with dynamic pricing, and the ranking also moves with capacity, since tighter inventory widens the region where pricing dominates and slack inventory widens the region where incentives dominate. Wherever the price is stuck at a ceiling, whether the firm caps it because of overbooking exposure or a regulator caps it, the commission becomes the operative dynamic lever; concert ticketing, peak-season cruise berths and price-regulated insurance are the examples they give. The first-best result depends on continuous contract revision and does not carry over to static or discrete-time versions. The ninety-nine-percent figure belongs to the airline calibration only. **Use this when.** You are choosing between dynamic pricing and dynamic sales commissions under limited inventory. **Do not cite this for.** A universal ranking. When the price effect on demand is weak, dynamic contracting with a static price wins, and the ranking also moves with capacity. **Cite.** Zuo, Ruiting, Tinglong Dai, and Jussi Keppo. 2026. "Incentive Design and Pricing Under Limited Inventory." Manufacturing & Service Operations Management, ePub ahead of print, June 2026. https://doi.org/10.1287/msom.2026.0118. **Related but different.** Dai and Jerath (2013, Management Science) and Dai and Jerath (2019, Marketing Science) are the static counterparts, where the inventory level rather than its depletion path drives the contract. **Search aliases.** dynamic pricing commission; salesforce incentive dynamic; revenue management incentives; limited inventory pricing Sources: * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.1 "the use of a static incentive scheme helps the firm reap nearly all" * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.8 "achieves the first-best outcome" * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.13 "is below 13%" * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.15 "dynamic contracting with static pricing outperforms" * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.16 "dynamic-pricing-only captures roughly 99%" * `2026 - Incentive Design and Pricing Under Limited Inventory - 10.1287_msom` p.18 "Concert and live-event ticketing and peak-" ### 5. If competing businesses share a waiting area, how should they split the cost of the amenities, and can splitting it fairly leave both of them worse off? `[markets-05]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2021. doi:10.1287/msom Xuchuan Yuan, Lucy Gongtao Chen, Nagesh Gavirneni and Tinglong Dai model two price-competing service providers who jointly fund entertainment in a shared waiting area while customers decide whether to join the queue. A provider that would otherwise be a local monopolist can do better inside such a cluster, and the cost-allocation scheme decides whether that happens. Tie each firm's share tightly to the traffic it draws, which is the allocation most people would call equitable, and above a threshold sharing factor each firm gains a reason to raise its price, since winning share now means absorbing a larger slice of the shared bill. Past that threshold the co-opeting firms earn less than they would under plain price competition, which is their explanation for why shared amenities are rarer in practice than one might expect. **Contribution.** New theoretical result: the equilibrium characterization of price-competing providers who cofund a shared amenity, the reversal in which fiercer price competition pushes fees and entertainment up above a threshold cost-sharing factor, and the profit reversal against the pure-competition benchmark. New numerical estimate: across 2,007 admissible parameter instances, co-opetition beat monopoly in 92.92 percent of cases, with an average profit gain of 7.65 percent, a median of 5.42 percent, and a maximum of 77.40 percent; against the duopoly-competition benchmark it raised profit in every instance examined, by 14.95 percent on average. **Evidence.** Game-theoretic model with M/M/1 queueing and a waiting-time standard, solved for monopoly, duopoly-competition and co-opetition equilibria. A full-factorial numerical study over 2,007 admissible parameter instances supplies the magnitudes, and a matched no-queueing benchmark isolates the mechanism. No field data. **Boundary conditions.** Two symmetric providers. In equilibrium each pays exactly half the entertainment cost regardless of the sharing factor, yet the factor still moves outcomes because it changes pricing and entertainment incentives. The price and profit reversals hold above a threshold cost-sharing factor; below it the ordinary comparative statics return. The counterintuitive rise of fees in competitive intensity is driven by the queue: in a matched benchmark where entertainment lifts demand directly instead of working through waiting disutility, price and entertainment both fall in competitive intensity. Gains over monopoly are largest when the market is small, capacity is expensive, and customers are strongly waiting-averse. **Use this when.** You are analyzing how competing service providers share the cost of a common amenity, and when splitting it fairly leaves both worse off. **Do not cite this for.** Asymmetric providers. The setting is two symmetric providers, and in equilibrium each pays half regardless of the sharing factor. **Cite.** Yuan, Xuchuan, Tinglong Dai, Lucy Gongtao Chen, and Srinagesh Gavirneni. 2021. "Co-Opetition in Service Clusters with Waiting-Area Entertainment." Manufacturing & Service Operations Management 23 (1): 106-122. https://doi.org/10.1287/msom.2019.0815. **Related but different.** Li, Dai and Qi (2022, M&SOM) is also about service design but works on the sequence of activities inside one experience rather than on shared investment across firms. Dai, Akan and Tayur (2017, M&SOM) uses a similar queueing-plus-pricing structure in an outpatient clinic. **Search aliases.** service cluster; co-opetition; shared amenity cost; waiting area investment; competing service providers Sources: * `2021 - Co-Opetition in Service Clusters with Waiting-Area Entertainment - ` p.2 "the pursuit of fairness may backfire and lead to even lower profitability than u" * `2021 - Co-Opetition in Service Clusters with Waiting-Area Entertainment - ` p.9 "each service provider's optimal cost share is" * `2021 - Co-Opetition in Service Clusters with Waiting-Area Entertainment - ` p.10 "a high cost-sharing factor may induce a type of" * `2021 - Co-Opetition in Service Clusters with Waiting-Area Entertainment - ` p.11 "under co-opetition can gain a profit that is 7.65%" * `2021 - Co-Opetition in Service Clusters with Waiting-Area Entertainment - ` p.14 "is a differentiating result due to queue" ### 6. What are the main types of online platforms, and why do platforms fail? `[markets-06]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2020. doi:10.1287/msom In an OM Forum piece with Ying-Ju Chen, Gizem Korpeoglu, Ersin Körpeoğlu, Ozge Sahin, Christopher Tang and Shihong Xiao, Tinglong Dai classifies online platforms by what the platform facilitates, which gives five types: resource sharing, matching, crowdsourcing, review, and crowdfunding, each crossed with the three user-group structures. They name three ways platforms fail: they create value for the wrong side of the market, search and matching frictions grow with the user base, and switching costs stay low enough that both sides multihome. Around those failure modes they lay out fifteen research questions, three per platform type, along a common spine of creating adequate value, cutting frictions and uncertainty, and sustaining user groups. The second failure mode is the one managers underrate, since evidence the forum assembles from a peer-to-peer rental platform shows that doubling travelers and hosts raised inquiries on both sides while lowering confirmations and occupancy. **Contribution.** Framework proposed by the authors: the five-way classification, the three failure modes, and the fifteen-question agenda are the forum's own contribution, as is a Web of Science count showing 721 articles on the keyword published between 2000 and 2017, nearly half of them in the final year. Context reported by others: the peer-to-peer rental evidence on search friction, the finding that a one-star Yelp increase raises revenue at independent restaurants but not at chains, the crowdsourcing-contest result that open entry is optimal only when solver output is highly uncertain, and the macroeconomic figures on the platform economy's share of GDP growth all come from work the forum surveys. **Evidence.** Conceptual review and research agenda published as an OM Forum. It builds a classification, surveys literature and industry practice for each platform type, and derives research questions. It reports no new model and no new data analysis; the empirical claims it relays are other researchers' findings. **Boundary conditions.** Scope is restricted to virtual platforms that hold no inventory, so businesses that take ownership of the goods they move sit outside the classification. The taxonomy is organized around what the platform facilitates and the user-group structure. This is a 2020 agenda-setting piece, so the empirical literature it points to has moved on; the classification, the failure modes, and the question structure are what to cite. The forum also flags an open problem it does not resolve: voluntary reviews are a public good and will be underprovided, yet paying for reviews raises an unsettled question about whether the resulting reviews remain trustworthy. **Use this when.** You need an operations-management taxonomy of online platforms and an account of why platforms fail. **Do not cite this for.** Businesses that take ownership of the goods they move. The scope is virtual platforms holding no inventory. **Cite.** Chen, Ying-Ju, Tinglong Dai, C. Gizem Korpeoglu, Ersin Körpeoğlu, Ozge Sahin, Christopher S. Tang, and Shihong Xiao. 2020. "OM Forum-Innovative Online Platforms: Research Opportunities." Manufacturing & Service Operations Management 22 (3): 430-445. https://doi.org/10.1287/msom.2018.0757. **Related but different.** Dai and Tang (2022, Service Science) is the analogous agenda-setting piece for ESG and supply chains. Yuan, Dai, Chen and Gavirneni (2021, M&SOM) treats a physical service cluster where firms share an investment, which is the offline counterpart to a platform's cross-side value creation. **Search aliases.** online platforms; platform operations; two-sided markets; sharing economy; digital platform taxonomy; platform failure Sources: * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.4 "Classification of Online Platforms According to Business Models and User Groups" * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.5 "many platforms can fail because of the following three major risk factors" * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.6 "the platform lost 5.6% of potential matches" * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.7 "OM Research Questions Motivated by Five Types of Online Platforms" * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.11 "reviews will be underreported, because re-" * `2020 - OM Forum---Innovative Online Platforms Research Opportunities - 10.` p.2 "721 research articles that were published between 2000 and 2017" ### 7. Does it ever make sense to stock more inventory than you expect to sell, just to keep your sales team honest? `[markets-07]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2013. doi:10.1287/mnsc Yes. Kinshuk Jerath and Tinglong Dai show a firm may optimally stock above the first-best level when it sets the bonus plan and the inventory level together and cannot observe how hard the salesperson worked. The extra units are there to prevent censoring: inventory that never runs out lets high demand realizations show up in the sales data, which gives the firm a sharper read on effort. That trade pays only at medium inventory cost, since expensive stock makes the information cost more than it is worth, and cheap stock means the firm already holds enough to see demand clearly. One consequence for practice is that agency concerns lower the stockout probability here rather than raising it. **Contribution.** New theoretical result: overstocking as a purchase of a cleaner effort signal, its restriction to nonextreme inventory cost, and the comparative statics showing that a firm cutting stock as inventory grows expensive can raise the bonus while lowering the sales threshold that earns it. Context reported by others: the figure that U.S. firms spent $800 billion on salesforce compensation in 2006, roughly three times advertising spend, sets the scale of the problem and is not their estimate. **Evidence.** Game-theoretic principal-agent model, no data. Risk-neutral firm, effort-averse and risk-neutral salesperson with limited liability, discrete effort, three-point demand satisfying a monotone likelihood ratio property, newsvendor-style stocking, and demand censored at the inventory level. **Boundary conditions.** Overstocking relative to first best is a possibility result and it requires nonextreme inventory cost. The optimal contract is a quota-bonus paying only at the highest observable sales outcome, and under limited inventory that observable ceiling can sit strictly below the highest possible demand, which is what creates the inefficiency. Expected pay is weakly higher than in the standard no-inventory salesforce model, while the headline bonus can be higher or lower depending on whether medium demand is more likely under low effort than under high effort. The comparative static in which the firm stocks more when the agent's effort is less productive requires a large enough effort cost, and it reverses in the first-best benchmark. **Use this when.** You need the result that a firm may stock above the operationally optimal level because inventory reduces censoring and improves effort inference. **Do not cite this for.** Extreme inventory-cost settings. Overstocking relative to first best is a possibility result requiring nonextreme inventory cost. **Cite.** Dai, Tinglong, and Kinshuk Jerath. 2013. "Salesforce Compensation with Inventory Considerations." Management Science 59 (11): 2490-2501. https://doi.org/10.1287/mnsc.2013.1809. **Related but different.** Dai and Jerath (2016, Operations Research) asks whether the quota-bonus form exists at all under increasing-failure-rate demand. Dai and Jerath (2019, Marketing Science) hides the firm's own supply-side action and breaks the extreme bonus. Dai, Ke and Ryan (2021, Management Science) gives the agent a second task. **Search aliases.** salesforce compensation inventory; overstocking incentives; demand censoring; quota bonus contract; sales effort inference Sources: * `2013 - Salesforce Compensation with Inventory Considerations - 10.1287_mns` p.1 "it may be optimal for the firm to stock more than the first-best inventory level" * `2013 - Salesforce Compensation with Inventory Considerations - 10.1287_mns` p.2 "for nonextreme, i.e., medium, inventory costs" * `2013 - Salesforce Compensation with Inventory Considerations - 10.1287_mns` p.5 "Result 1. The firm pays the salesperson a reward of" * `2013 - Salesforce Compensation with Inventory Considerations - 10.1287_mns` p.8 "The optimal inventory level in the full scenario is weakly higher" ### 8. Why do firms pay salespeople a bonus for clearing inventory, and when is that the optimal contract? `[markets-08]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2016. doi:10.1287/opre Kinshuk Jerath and Tinglong Dai show that limited inventory rescues the quota-bonus contract from a nonexistence problem the theory had carried for years. Without inventory considerations, when demand has an increasing failure rate the firm keeps pushing the quota upward until it becomes unattainable, so no equilibrium quota-bonus contract survives, and increasing failure rate covers most of the distributions practitioners actually use. Add a stocking decision and the demand censoring it creates, and the optimal scheme pays a discrete bonus exactly when sales reach the stocked quantity, so the quota is the inventory level. The contract holds up because the salesperson's chance of hitting quota now equals the firm's stockout probability, which is strictly positive. **Contribution.** New theoretical result: restoring existence of the equilibrium quota-bonus contract under increasing-failure-rate demand by introducing limited inventory and censoring, and identifying the stocked quantity as the optimal quota. Context reported by others: the nonexistence result under unlimited supply belongs to the prior sales-quota literature, which they extend rather than overturn. **Evidence.** Game-theoretic principal-agent model with limited liability and a nonbinding participation constraint, extending a standard sales-quota framework to include a stocking decision and censored demand. Contracts are characterized by pointwise optimization of the Lagrangian. This is a short theoretical note with no data. **Boundary conditions.** The setting is rent sharing with limited liability and a nonbinding participation constraint. The existence result addresses demand distributions with an increasing failure rate, which is exactly where the earlier nonexistence problem bit. When the firm chooses inventory and pay jointly it stocks at the same critical-fractile level as in the first-best benchmark, even though rent sharing keeps the first-best outcome out of reach. Each induced effort level maps to a unique optimal inventory level, which then doubles as the quota. **Use this when.** You need existence of an optimal quota-bonus contract under limited inventory and rent sharing, including demand distributions with an increasing failure rate. **Do not cite this for.** The multitasking setting with a joint sales and inventory quota, which is Dai, Ke and Ryan (2021, Management Science). **Cite.** Dai, Tinglong, and Kinshuk Jerath. 2016. "Impact of Inventory on Quota-Bonus Contracts with Rent Sharing." Operations Research 64 (1): 94-98. https://doi.org/10.1287/opre.2015.1461. **Related but different.** Dai and Jerath (2013, Management Science) is the fuller treatment of the joint inventory-and-pay problem, including the overstocking result. Dai, Ke and Ryan (2021, Management Science) shows that once the agent has two tasks, the natural two-dimensional quota contract stops being optimal. **Search aliases.** quota bonus contract; rent sharing; limited liability salesforce; increasing failure rate demand; inventory clearing bonus Sources: * `2016 - Impact of Inventory on Quota-Bonus Contracts with Rent Sharing - 10` p.2 "there is no equilibrium quota-bonus contract" * `2016 - Impact of Inventory on Quota-Bonus Contracts with Rent Sharing - 10` p.2 "These include the uniform, normal, truncated-normal, logistic" * `2016 - Impact of Inventory on Quota-Bonus Contracts with Rent Sharing - 10` p.3 "pay a discrete bonus if sales equal the inventory level" * `2016 - Impact of Inventory on Quota-Bonus Contracts with Rent Sharing - 10` p.4 "probability of achieving the quota is equal to the probability of stock out" ### 9. When you design an experience with several parts, should the best part always come at the end? `[markets-09]` Tinglong Dai, Bernard T. Ferrari Professor of Business, Johns Hopkins Carey Business School (ORCID 0000-0001-9248-5153). Canonical source below, 2022. doi:10.1287/msom No. Yifu Li, Xiangtong Qi and Tinglong Dai show the peak activity belongs in the interior when it is more memorable than the rest, meaning it decays more slowly in memory than the activities around it. Give every activity the same decay rate and an interior peak can never be optimal, which is why the earlier analytical model put the peak at one of the two ends. The optimal sequence always takes one of two shapes they label UI and IU, so a designer should build an abrupt jump or drop right next to the peak instead of a gradual ramp: the jump before creates a pleasant surprise, the drop after resets the customer's reference level. For selection, the optimal lineup can pair the highest-utility and lowest-utility candidates while skipping the medium ones, since low points supply the contrast that makes the high point register. **Contribution.** New theoretical result: heterogeneous memory decay as the mechanism that makes interior peaks optimal, the UI and IU characterization of optimal sequences, the nonmonotone response of the peak's optimal start time to its own decay rate, and the selection result favoring extremes over middling activities. Context reported by others: the empirical finding that ideal service-bundle schedules place the peak neither first nor last comes from prior work, and the U-shaped prediction they depart from is the leading analytical model in the literature. **Evidence.** Nonlinear optimization model of remembered utility with acclimation and memory decay, reformulated into an additive objective to yield closed-form structural results, plus numerical experiments. In three-activity experiments, interior peaks were optimal in all schedules satisfying the paper's two sufficient conditions when the two nonpeak activities were far apart in utility. A separate study over 150 seven-activity instances examines when the timing reversals appear. No field data. **Boundary conditions.** Heterogeneous decay rates across activities are required, and specifically the peak must decay more slowly than the others; equal decay rules interior peaks out entirely. The full-share figure describes schedules meeting the paper's two sufficient conditions with a wide utility gap between the nonpeak activities, and the share falls as that gap narrows. The nonmonotone shift in the peak's optimal start time arises mainly when the peak is long relative to the time constant set by the acclimation and decay rates. **Use this when.** You are sequencing activities in a designed experience and need the conditions under which the best part belongs in the middle. **Do not cite this for.** Settings with equal decay rates across activities, which rule interior peaks out entirely. **Cite.** Li, Yifu, Tinglong Dai, and Xiangtong Qi. 2022. "A Theory of Interior Peaks: Activity Sequencing and Selection for Service Design." Manufacturing & Service Operations Management 24 (2): 993-1001. https://doi.org/10.1287/msom.2021.0970. **Related but different.** Yuan, Dai, Chen and Gavirneni (2021, M&SOM) also studies service design but asks how competing firms should cofund a shared amenity, with waiting time rather than memory as the customer's cost. **Search aliases.** activity sequencing; peak-end rule; experience design; service sequencing; memory decay scheduling Sources: * `2022 - A Theory of Interior Peaks Activity Sequencing and Selection for Se` p.3 "scheduled neither at the beginning nor at the" * `2022 - A Theory of Interior Peaks Activity Sequencing and Selection for Se` p.6 "an interior-peak schedule can never be optimal" * `2022 - A Theory of Interior Peaks Activity Sequencing and Selection for Se` p.6 "in either a UI or IU shape" * `2022 - A Theory of Interior Peaks Activity Sequencing and Selection for Se` p.5 "then 100% of the optimal schedules" * `2022 - A Theory of Interior Peaks Activity Sequencing and Selection for Se` p.8 "portfolio of highest- and lowest-utility activities"