WorksEconomics, Econometrics and Finance
When Better Artificial Intelligence Widens the Adoption Gap in Health Care
Abstract
Artificial intelligence (AI) can assist healthcare providers in clinical practice or replace services for which they are paid. We model how provider influence shapes two policy choices. The first is adoption, the share of cases in which coverage is switched to care using AI. The second is the review requirement, the number of paid reviews that must accompany AI when it is used in clinical care. Assistance can lower providers' costs and increase their profits while leaving paid reviews unaffected. However, increasingly capable AI may support autonomous use that can replace reviews and reduce providers' profits. Provider gains from adoption first rise and subsequently fall as AI capability improves. Hence, providers oppose sufficiently autonomous AI if their lost income from reviews exceeds savings from better assistance. Opposition to AI adoption arises earlier for AI that displaces more reviews. When providers can influence adoption but not the review requirement, their opposition causes under-adoption. The patient value forgone at the margin grows as AI improves. At interior policies, provider influence over both decisions widens adoption and raises review requirements above the patient optimum.
Introduction
The adoption of artificial intelligence (AI) in health care reflects two institutional features. First, clinical expertise gives providers a role in coverage, payment, and regulatory decisions, making them partial gatekeepers to technologies that may assist or replace their work. Second, patients largely finance care collectively through taxes and insurance premiums. Payers make coverage and payment decisions on their behalf, so the costs are borne collectively rather than necessarily at the point of care. We model how provider influence shapes AI adoption and review requirements.
AI technologies are increasingly being adopted in health care. For example, US Medicare pays for an autonomous AI examination that issues a diagnostic report for diabetic retinopathy without physician review [1]. Yet in their 2020 comments to the US Food and Drug Administration (FDA), radiology’s professional bodies supported assistive AI but opposed autonomous interpretation without physician confirmation, including systems that identify scans showing no abnormality and would allow them to bypass a radiologist’s review [6].1 Under fee-for-service payment, this distinction matters because payment attaches to the professional’s interpretation and written report [30]. Image reconstruction and automated measurements can lower the cost of an interpretation that the professional still performs and bills, whereas autonomous interpretation replaces the billed service. These effects give provider organisations a stake in adoption rules. Specialties closer to Medicare’s fee-advisory committee obtain higher prices [11], and physician groups’ political spending influences state scope-of-practice laws [28].
Our main contribution is a task-based theoretical microfoundation for the change from support for AI adoption to opposition. Earlier models explain resistance through threatened political power or incumbent rents [4, 14], and some find that resistance first rises and then falls with market size [15]. In our model, incumbents can initially be winners and subsequently losers. Using a task-based approach [3], we derive the sign of providers’ gain from the order in which AI reaches clinical tasks. This gain first rises with AI capability but falls when marginal savings per displaced review decline below the review margin. Opposition arises when eventual income losses from displaced reviews exceed retained savings. Provider influence therefore tilts adoption towards AI that assists paid work rather than replaces it, relative to its value to patients.
Following Grossman and Helpman [22], a policymaker weighs providers’ lobbying contributions against patient value when setting adoption and review requirements. When a body insulated from provider influence sets reviews at the patient optimum, the influenced policymaker controls only adoption. Provider opposition then causes under-adoption, and the patient value forgone at the margin rises as AI improves. When the policymaker sets both decisions, at interior review policies, it requires reviews that patients would not buy. In the quadratic example, setting review requirements independently at the patient optimum reduces patient losses unless under-adoption is severe.
The results contribute to two strands of health economics. The first studies how reimbursement shapes providers’ technology adoption [2, 7, 8, 23, 33], including second-best payment rules that, without lump-sum payments, accept overtreatment to speed adoption [27]. We show that payment can also shape adoption rules through provider influence. The second studies physician supervision requirements. Relaxing restrictions on nurse practitioners has lowered prices [24] and increased routine care [35], while Smith [34] finds no evidence of a change in patient volume. For nurse anaesthetists, Dulisse and Cromwell [16] find no evidence of increased inpatient deaths or complications following removal of the supervision requirement. Review requirements for AI are a new form of supervision requirement, and we show how provider influence can keep them above what patients would choose.
Model
A case is one patient’s service in which AI could be used, such as a chest X-ray. A review is a professional’s paid interpretation and written report, and providers are the professionals paid for reviews and the practices that bill for them. Adoption is the share of cases, called adopted cases, in which the payer switches coverage to care using AI, supplied by participating providers. Policy can set this share by indication, patient group, or care setting. For these cases, providers cannot instead bill for conventional care without AI. We assume that participation remains worthwhile at the practice level, so providers supply this care even when AI reduces income relative to existing care. Regulatory clearance is taken as given. The review requirement is the expected number of reviews per adopted case, common to adopted cases because it is set for a type of use rather than case by case. We impose , consistent with Medicare generally paying for only one interpretation of an emergency-room X-ray [30]. Thus can represent the share of AI-read X-rays requiring a radiologist’s report. AI capability indexes performance relative to typical task difficulty, so denotes capability below that benchmark. Information is complete and fees remain fixed as capability changes.
We evaluate policy from the perspective of (potential) patients, who largely fund care through taxes and premiums. Patient value is the monetary value of health gains less spending on the costs of care, including review fees. Payments to providers count as costs, as in health technology assessment (National Institute for Health and Care Excellence, 2022). Let be the review requirement that maximises patient value at adoption and for capability . With the number of cases normalised to one, the corresponding patient value relative to existing care, that is, care without AI, is . We assume and , so wider adoption reaches cases where AI adds progressively less value.
Provider gain: a task-based microfoundation
Each review earns providers a margin , the fee less the opportunity cost of the professional’s time. The providers we model are paid per service. AI can first assist with component tasks, such as image processing and measurement, lowering the provider’s cost while retaining the paid report. Removing that report requires AI to perform the complete interpretation autonomously, without professional review [1].
Let be the savings providers keep from AI-assisted tasks, such as preparing images, and the paid reviews per case lost to autonomous AI. In the main model, the organised provider group retains these savings even after AI replaces its paid report, as an integrated practice operating the AI service might. This return comes from the AI service, not from billing for an unperformed review. The provider gain per adopted case is , the savings kept less the margin lost on displaced reviews. Providers support adoption when and oppose it when . Expected provider gain is the same across adopted case types, although patient value and review needs may differ. The key object is the providers’ saving per displaced review at the capability margin, .
Assumption 1 (Declining savings per displaced review). and are continuously differentiable with and , , and finite limits and . The ratio is strictly decreasing, with .
Proposition 1 rests on this assumption. A common lag from assistance to autonomy and strictly log-concave task difficulty generate the declining ratio, which must also cross the review margin (Supplementary Section S2.3). Assistance precedes autonomy because professionals can catch errors when AI assists, whereas autonomy requires accuracy high enough to dispense with review. In mammography, computer-aided detection has assisted radiologists since FDA approval in 1998 [21], whereas AI-supported single reading for lower-risk mammograms was evaluated in the MASAI trial [25].
Consider reports whose difficulty is standard normal. AI can assist with a report’s component tasks once , saving , but can replace the report only once , with . The expected number of paid reviews displaced per case approaches as capability rises. The distribution of difficulty is the same within every adopted case type, so the following are expected gains and displaced reviews per case. Savings persist after autonomy. Writing for the standard normal distribution,
so falls from infinity to zero and Assumption 1 holds for all positive , , , and (Supplementary § 8.2), with and . With one existing report per case, . A single report with would give the same pattern in steps. The provider gain jumps to once AI reaches sufficient capability to assist and falls to when it becomes autonomous.
Differences in difficulty across reports smooth these steps, and the normal shape places most reports at moderate difficulty. Supplementary § 8.4 derives the same rise and fall in provider gain when replacing a report requires several component tasks to be autonomous. Figure 1 illustrates the example with , , , and .
Figure 1. Provider gain and review-adjusted return, AI adoption, and patient losses. Notes: Example (1) with quadratic patient value and the parameters in Supplementary § 10.3. Capability is measured relative to mean task difficulty, so is below that benchmark. Dotted lines mark the onset of under-adoption under separate () and joint () decisions. Panel (a) shows provider gain and the review-adjusted return , where . Provider gain under joint decisions is ; the return deducts the weighted patient cost of excess reviews. The kink in panel (c) occurs where the one-report cap ceases to bind, at .
Political choice
In our model, there is one organised provider group that offers contributions to the policymaker conditional on the policy chosen. The policymaker can be a public payer, such as Medicare, or a private health insurer, and can regulate both adoption and review rules. Contributions represent any form of influence, including official advisory roles, lobbying and public communications. Following Grossman and Helpman [22], the policymaker maximises total provider gain plus times patient value, both summed over adopted cases:
Providers earn for each adopted case plus the margin on excess reviews , the reviews required beyond the patient optimum . The margin on reviews up to is already reflected in . Reviews go first to the cases where they matter most, so patient value falls as reviews depart from , at a rate set by . Too few reviews forgo health gains worth more than their fees, and too many buy gains worth less. The policy that maximises is the one selected in the single-group menu-auction equilibrium [10]. The weight measures how the policymaker weighs benefits from providers and patients, giving up a dollar of patient value in return for at least dollars of provider gain. As grows, policy approaches the patient optimum, our benchmark.
Results
Support, then opposition
The provider gain rises while savings on newly assisted tasks exceed the income lost from displaced reviews, . Providers can save time preparing images or taking measurements while keeping paid interpretation. Once , further capability improvements displace more review income, , than they add in savings, , so total gain falls although savings keep accumulating. Opposition requires eventual lost review income, , to exceed full savings, (in example (1), ). This is plausible in screening, where most examinations show no abnormality. Autonomous reporting could remove these reviews, whereas assistance only shortens them. Figure 1(a) shows the gain rising, peaking at , and turning negative at .
Proposition 1 (Provider support). Under Assumption 1, the provider gain satisfies for and for , where . A finite zero exists if and only if . Then and .
AI technologies differ in these parameters, i.e., the savings , the reviews displaced , and the lag between assistance and autonomy. AI that takes measurements on an X-ray mainly saves time, a high , whereas AI that reports normal X-rays removes reviews, a high . Corollary 1 shows how these parameters move the opposition threshold .
Corollary 1 (Selection of AI technologies). In example (1), if , the threshold rises with and and falls with and .
Provider influence therefore affects which type of AI is adopted earlier or to a greater extent. Technologies that displace more reviews per report (higher ) meet opposition at lower capability , whereas assistive technologies that generate large savings to providers (higher ) or have a long lag before autonomy (higher ) keep provider support to higher . Holding patient value fixed, provider influence thus tilts adoption toward assistive AI.
Under-adoption
Superscripts mark the decision, with for the patient optimum, for separate decisions, and for joint decisions. When adoption and reviews are regulated separately, the review requirement is set at the patient optimum by a body that providers cannot influence, implying . Patient-optimal adoption solves , whereas the policymaker chooses to maximise . The adoption gap is the patient value of the marginal adopted case. It is zero at , positive under under-adoption, when extending AI to further cases would still benefit patients, and negative under over-adoption.
Corollary 2 (Separate decisions). At interior policies,
Hence if and only if and , and for such the adoption gap widens as rises.
By Corollary 2, adoption stops where the patient value of the marginal case, weighted by , offsets the provider’s loss. Once under-adoption begins, each improvement in capability displaces more review income than it adds in savings, so more capable AI widens the adoption gap. The gap is also larger for AI that displaces more reviews, as reflected by a higher in example (1). When the provider gain is positive, as for , providers push adoption beyond the patient optimum, leading to over-adoption, as in Figure 1(b).
Excess reviews
When and are regulated jointly, the policymaker sets both adoption, , and the review requirement, (Proposition 2). Each excess review earns providers the margin , at a rising cost to patients.
Proposition 2 (Joint decisions). At interior policies, excess reviews and adoption satisfy
Hence , and if and only if and , where solves .
Excess reviews compensate providers for the review income they lose to AI. A policy may, for example, pay for AI to read an X-ray but still require a professional report, even where that report adds less health benefit than it costs. Patients bear this cost in every adopted case.
Excess reviews also increase adoption. At an interior review policy, each adopted case now brings the policymaker the provider gain plus . This is the margin on excess reviews, , less their weighted cost to patients, . Under-adoption therefore begins only once the gain falls below , and never if exceeds providers’ eventual loss from adoption, . In Figure 1(a), the review-adjusted return crosses zero later, at . In Figure 1, the cap affects this gain only at lower capability, where most cases still need a review and little room remains for excess reviews. More generally, the cap can remain binding at high capability; the interior adoption and threshold results then need not apply (Supplementary Section S3.2). Because at interior review policies, excess reviews also increase over-adoption while the provider gain is positive. Patients pay for this wider adoption, and their aggregate loss is
In (6), is the loss from adoption differing from the patient optimum. The remaining term is the cost of excess reviews across adopted cases, which persists even at patient-optimal adoption.
To compare the two decisions, take an example in which patient value is quadratic in adoption, with . At interior policies,
The adoption gap is then . Under separate decisions it rises whenever falls, and under joint decisions it does so when the cap does not bind. Adoption itself falls only if in that region. Figure 1(b) shows adoption rising, falling, and recovering while the adoption gap keeps widening after the cap ceases to bind. Let be the patient loss under separate decisions. In this example, whenever because the extra adoption due to excess reviews is only valuable when the shortfall is large. Figure 1(c) shows separate decisions losing less at every capability .
Discussion
Our results imply that review requirements for AI deserve the same scrutiny as adoption, because broad adoption can come bundled with reviews patients would not buy. How providers are paid matters through the margin on each review, which raises both the income lost on displaced reviews and the value to providers of excess reviews. Under a fixed payment per case, an additional review brings no additional fee but still costs the professional’s time, so the margin is negative. Provider influence therefore pushes the review requirement below the patient optimum (Supplementary Section S5.3), matching the skimping concern of prospective payment [17, 18]. Compensating providers through excess reviews can cost patients less than paying them directly per AI use, and may be less visible to voters (Supplementary Section S5.1). Levaggi et al. [27] obtain a similar trade-off for a welfare-maximising payer without lump-sum payments, whereas here it arises from provider influence.
Applications
Three applications illustrate the payment margins rather than identify provider motives. Computer-aided detection in mammography assists radiologists while retaining their interpretation. Its use spread after Medicare began paying for it in 2001 [20], although large studies found no gain in diagnostic accuracy [21, 26]. Autonomous retinal screening can expand access beyond existing eye-care provision [1]. If it mainly reaches patients who would otherwise go unscreened, displaced review income, , is small. Autonomous reporting of normal chest X-rays was evaluated retrospectively in Danish public hospitals for its potential to reduce radiologist workload [31]. Hospital doctors there are salaried, with pay based on hours rather than clinical output (European Observatory on Health Systems and Policies, 2024). Replacing a reading therefore does not directly remove a per-reading fee from the radiologist.
Other explanations
Other explanations also predict resistance to autonomous AI. Performance may be uncertain or vary across sites, as the radiology bodies argued [6]. The precautionary value of review then belongs in , and influence adds reviews beyond it. Radiologists may also have difficulties in combining their judgement with AI predictions [5], and liability law tends to favour following the standard of care over following an AI recommendation [32]. None of these explanations links review requirements to how providers are paid, and our model provides that link.
Separating professionals and billing organisations
We model an integrated provider. If hospitals or other organisations that bill for services keep the savings from greater AI adoption while only professionals are organised, opposition starts earlier, because the professionals lose review income without keeping the offsetting savings. A paid-report requirement then also becomes a more valuable concession to those professionals, because it restores both their review margin and the savings from assisted care (Supplementary Section S5.2). The ranking in Figure 1(c) can then reverse, for about at the figure’s parameters if professionals keep none of the return from AI-only care.
Conclusion
We make three qualitative predictions. First, for given capability and savings , opposition to autonomous AI should be stronger when providers are paid per review than when they are salaried or paid per case. Second, under fee-for-service payment and at interior policies, review requirements attached to AI should exceed the patient optimum, by more where provider organisations have greater influence over the payer. Third, provider influence should favour assistive over autonomous AI relative to their value to patients, even when autonomous AI benefits patients. Overall, AI that first assists providers and later replaces their paid reviews can turn them from supporters into opponents of adoption. Requiring paid reviews alongside AI can keep their support for longer, but patients pay for reviews they would not buy. Better AI therefore may not bring policy closer to what patients would choose.
Data availability
No new data were created or analysed in this study. The figure is computed directly from the parameter values reported in the Supplementary Material, and the code that generates it is provided with the Supplementary Material.
Supplementary Material
Section S1 derives the aggregate model from payment accounting. Section S2 proves Proposition 1 and Corollary 1, and extends the example to bundled component tasks. Section S3 gives the political equilibrium and proves Corollary 2 and Proposition 2. Section S4 develops the quadratic example and Figure 1. Section S5 treats direct payments, ownership of the AI-only return, nonpositive margins, and case-specific provider gains. Numbers without an S prefix refer to the main text. All comparisons hold the specified primitives fixed.
Foundations and aggregate representation
Payment accounting
Coverage is switched to care using AI for adopted cases, which participating providers supply. Conventional care without AI is not reimbursable for those cases. Participation constraints at the practice level, including the cost of contributions, do not bind for the policies considered, so a negative provider gain below is a loss relative to existing care, not an exit. The integrated practice continues to supply the AI service and retains its return even when no professional report is billed.
To derive the main text’s aggregate functions, index case types by complexity . Each type is a continuum of cases. Write for reviews per case in existing care, for the type-specific patient optimum for reviews, and for the optimal common requirement over adopted cases. Each review pays and costs the provider , giving margin .
With AI, let be the payment for the other services in a case and their provider cost, both measured relative to existing care, so the provider’s savings are . Let be the monetary value of the change in patient outcomes and access, net of resource costs borne outside the provider. Patient value per adopted case of type is and the provider’s gain per case is . Hence
Writing and gives , the type-level counterpart of the provider term in (2). Fixed costs and patient benefits of assisted tasks are included in and , and the corresponding payment is included in .
Within a type, reviews go first to the cases in which they are most valuable, so the marginal health value of additional reviews declines as the review share rises. Suppose is twice continuously differentiable and on the relevant review range. The interior patient optimum solves . Completing the square around this point yields exactly
where . Comparative statics in vary at fixed , which leaves , , , and patient value unchanged.
From a complexity cutoff to adoption
Let be the case distribution, with continuous density . Adoption covers cases with complexity , so . With a common requirement, the patient optimum is the average type-level optimum among adopted cases,
For , write . Because is this average, and , where is the dispersion of review needs among adopted cases. Integrating the type-level patient value and provider gain, with , therefore gives exactly (2), with
Deviations of type-level review needs from cancel in the provider term and enter patient value only through the mismatch cost . The marginal value equals the value of the marginal type net of its mismatch cost, so the main text’s means that this value falls along the complexity order. When displaced reviews are common to all cases, , so is the average requirement in existing care less . The normal example takes the same distribution of reporting difficulty within each type, so and are expectations and . If provider gains instead vary by type, the separate-decision condition is , and Section S5.4 gives a condition for the gap to keep widening.
Task mechanism and foundations
Proof of Proposition 1
Under Assumption 1, write , so that with . Since is continuous and strictly decreasing, with , there is a unique with , and the provider gain strictly increases before and strictly decreases thereafter. Since , , so the maximum is positive, and . On the decreasing branch every finite value is strictly above this limit. A negative limit therefore gives exactly one zero , followed by a negative provider gain. A nonnegative limit gives a positive provider gain at every finite , with convergence to zero from above at equality.
More generally, let be a constant added to the provider gain. The derivative and turning point of are unchanged, its maximum is positive, and its limit is . A finite zero therefore exists if and only if . At equality there is no finite crossing. Where it exists, , , and
Taking , with excess reviews under joint decisions from Proposition 2, proves the threshold under joint decisions and its existence condition .
The paired normal example and proof of Corollary 1
In the paired example, and . With ,
which is continuous and strictly decreasing from infinity to zero. Assumption 1 therefore holds for all positive , with , , and . The opposition condition of Proposition 1 becomes , and the condition under joint decisions becomes .
Proof of Corollary 1. Let be any of . Since and , implicit differentiation gives , which has the sign of . From (1),
Hence rises with and and falls with and . The same argument applies to for fixed , holding the other primitives fixed. Since the adoption gap under separate decisions is by (3) and , the gap is larger for AI that displaces more reviews.
Beyond the normal example
Assumption 1 does not depend on the normal example. Let tasks become feasible at capability requirement , with saving and displaced reviews , and let and be the densities of assisted and autonomous tasks over . Then , , and , which is strictly decreasing if and are non-increasing and at least one strictly decreases, that is, if later tasks bring less saving per displaced review or assisted tasks are concentrated at lower requirements. Neither channel requires every assisted task to precede every autonomous one. The autonomy requirements here are those needed to replace a report, rather than to automate an individual component. For the paired reports of the main text with a general difficulty density , , which decreases for every lag if is strictly log-concave. Assumption 1 also needs the unscaled density ratio to straddle . The normal density satisfies this for all positive . Other strictly log-concave densities, such as the logistic, may require parameter restrictions. Assistance preceding autonomy alone guarantees only .
Bundled component tasks
Suppose each report requires component tasks, with independent standard normal difficulties and the same distribution within each case type. AI assists with component once , saving , and performs it autonomously once . Savings persist after autonomy, which adds no saving beyond assistance. Let be the expected number of paid reviews per case that AI can displace. This occurs only when all components are autonomous. Linearity of expectation gives savings , while independence gives report-replacement probability . Hence
Writing , differentiation yields
The ratio falls from infinity to zero: it is the ratio in (S2.1), divided by . Also, and . Thus Assumption 1 holds, Proposition 1 applies, and eventual opposition again occurs if and only if . The case is exactly (1); with , component assistance can lower costs before the complete report can be replaced. Independence and equal savings per component are assumptions of this example. The thresholds generally differ with ; Figure 1 retains .
Political equilibrium and policy results
Menu-auction equilibrium
Let . Write for patient value and for total provider gain, so that in (2). The policymaker maximises contributions plus and can always choose the maximiser of . An implemented policy therefore requires , so the provider group’s net gain is at most , which is maximised at the maximiser of . The truthful schedule with attains this bound, because it makes a best response for the policymaker, which breaks ties in favour of the provider group [10, 22]. The same argument applies when policy is restricted to separate decisions, .
Proof of Corollary 2 and Proposition 2
For the interior results, take with slack review bounds and review optima in the range where patient value is quadratic. Write . The aggregate objective is
Under separate decisions , and strict concavity in adoption gives . Since decreases in and is zero at , if and only if . By Proposition 1, this holds if and only if and . Differentiating the identity along equilibrium gives , which is positive for and proves the final statement of Corollary 2.
At any , strict concavity in yields excess reviews . Inserting gives the reduced objective
Its first-order condition is (5). At its derivative is , so . Comparison with gives iff . Section S2.1, with , shows that this holds if and only if and , where .
Because the change of variable absorbs the dependence of the common requirement on adoption, optimisation over is equivalent to optimisation over and adds no term to these first-order conditions when the review bounds are slack.
Only strict concavity in is needed for the signs. If excess reviews cost patients a , with strictly convex and , then solves , and the reduced objective adds . Hence , and Proposition 2 holds with in place of . The quadratic case gives the closed forms.
With the one-report cap and , the conditional review choice at is
The reduced objective is . At an interior adoption choice its first-order condition is
Where the cap binds, ; where it is slack, . Thus a binding cap and changing case composition need not preserve . If is independent of , as in Figure 1, and strict concavity gives , strictly when and adoption is interior. Joint under-adoption then occurs exactly when .
In the normal example with and , let . As ,
If , the cap binds at every capability and this limit is . The interior threshold condition cannot be used in this case. Figure 1 has , and the cap is slack at both under-adoption thresholds.
Patient losses and total surplus
At , the patient-value term of (2) gives . Subtracting this from proves (6). The adoption component equals
Strict concavity makes it positive whenever , including over-adoption. Under separate decisions the loss is .
At fixed , total surplus changes with by . The total-surplus (efficient) optimum is
Thus, at interior review policies, reviews under joint decisions exceed the total-surplus optimum at the same adoption exactly when .
Quadratic example and figure
Adoption formulas
Let , . At interior policies,
More generally, differentiating an interior first-order condition, at or at , gives
because does not vary with . Since , adoption falls if . In the quadratic example , so both decisions have and the same turning points.
Comparing separate and joint decisions
Write , so . The cost of excess reviews per adopted case is . Let be the patient loss under separate decisions.
Corollary S1 (Separate or joint decisions). At interior policies, if and only if
In particular, whenever .
Proof. The adoption formulas are (S4.1) and (S4.2). With all adoption choices interior, the two losses are
Their difference is
which proves the equivalence. If , the left side of (S4.4) is nonpositive, so the inequality fails.
Joint decisions reduce patient losses only when adoption under separate decisions is below half the patient optimum, and by enough to compensate for the cost of excess reviews.
Figure parameters and numerical values
For the illustration in Figure 1, take the paired reports of (1), , existing care , and
These values imply the unconstrained excess review and . With the cap ,
The cap binds for , where and . For , and , as in the interior formulas above. The joint patient loss is . The figure is produced by the Python script figures/make_figure1_revision.py, provided with this Supplementary Material.
Outside the capped range, adoption falls between the zeros of , at and , and rises elsewhere. The thresholds are and . The shortfall keeps widening after adoption recovers and tends to , because decreases for and is bounded. For , , so and separate decisions lose less. The script checks numerically that all three adoption policies are interior, that joint adoption rises for , and that at every plotted .
Extensions
Direct payments for AI use
Payers pay for services listed in their fee schedules, so direct compensation for displaced reviews takes the form of a payment per AI use, such as Medicare’s payment for the autonomous retinal examination [1]. Each dollar of such a payment costs patients a dollar, collectively through premiums and taxes. Suppose the policymaker can also pay providers per adopted case, independently of reviews. The payment lowers and raises by , adding to (S3.1), so the review rule at fixed adoption is unchanged. The equilibrium requirement can change if the payment changes adoption and hence . The chosen payment is zero if , and if it must be capped by a budget or coding constraint.
At the main model’s interior joint policy, excess reviews raise provider income by per adopted case at a patient cost of . A direct payment of equal provider value costs patients . At this interior policy, reviews are therefore the cheaper transfer iff . With the cap , the chosen excess review is , and its patient cost per dollar of provider gain is . Reviews are then cheaper than an equal direct payment iff , so is sufficient but not necessary. Their clinical value offsets part of their fee, since the first excess reviews at the patient optimum have zero net cost to first order. This comparison holds adoption and the payment system fixed. Compensation through reviews can also reflect lower visibility [13] or interest groups’ preference for indirect instruments [22], despite pressure toward efficient transfers [9].
Who retains the return from AI-only care
The main model assumes that the organised provider group keeps the saving even after AI replaces its paid report. Ownership of complementary services can change which providers capture such returns [12]. To isolate this incidence, keep patient outcomes, patient-facing payments, , and the technology fixed, and let the organised group receive only of the return on a case that AI handles alone, with going to an unorganised provider. This is not a change in the payer’s fee. Take the paired normal example with one existing report per case (), so .
An assisted case with a paid report gives the group relative to existing care, and an AI-only case gives . An excess paid report turns an AI-only case back into assisted care, so it is worth to the group rather than . The provider gain at the patient optimum and the joint review policy are
where the upper bound is the cap . For this is the main model. Since changes sign once and , support turns to opposition at a finite if and only if . Excess reviews add to the return on adoption, so joint decisions still increase adoption at every capability, and joint under-adoption, where it occurs, starts later than .
At the Figure 1 parameters, moving from to moves the separate-decision opposition threshold from 0.686 to 0.068 and the joint threshold from 1.014 to 0.684. The cap then binds for . The larger review incentive does not fully offset the earlier loss of AI-only income. With the same patient-value primitives, joint decisions yield lower patient loss only for approximately . Thus the main figure’s patient-loss ranking is an illustration, not a general consequence of separating decisions. If instead changes total reimbursement or resource costs, must also change and these formulas do not apply.
Nonpositive margins
The main results assume . With no marginal review income, , the provider gain is and there are no excess reviews. Under a fixed payment per case, the incremental fee for an additional review is zero, so . Displacing reviews then saves the provider resources, and .
If the quadratic patient value remains valid over the relevant range, the constrained requirement under joint decisions at positive adoption is
Thus influence favours fewer reviews. The cap in Figure 1 does not alter this nonpositive-margin solution when .
Case-specific provider gains
Retain adoption by a complexity cutoff and take uniformly distributed case types , so adopted types satisfy . Let , with and the main model’s provider gain: more complex cases require greater AI capability. Take , , with differentiable . Under separate decisions the policymaker maximises . At interior patient-optimal and political adoption, write . The first-order and strict second-order conditions are
Along this interior solution, implicit differentiation gives . Thus the marginal adoption gap satisfies
Equivalently, since by (S5.2), , where is the change in the effective capability of the marginal case. Under-adoption occurs when , which also implies by Proposition 1. Given , the gap then widens exactly when , is locally unchanged at equality, and narrows when the inequality reverses. Since , the widening condition is . If adoption expands into more complex cases sufficiently quickly, AI can become less autonomous for the marginal case despite increasing capability. This composition effect can reverse the widening result; the common-gain model excludes it.
References
- [1]Abràmoff, M. D., Roehrenbeck, C., Trujillo, S., Goldstein, J., Graves, A. S., Repka, M. X., & Silva, E., III. (2022). A reimbursement framework for artificial intelligence in healthcare. npj Digital Medicine, 5(1), Article 72. https://doi.org/10.1038/s41746-022-00621-w
- [2]Acemoglu, D., & Finkelstein, A. (2008). Input and technology choices in regulated industries: Evidence from the health care sector. Journal of Political Economy, 116(5), 837–880. https://doi.org/10.1086/595014
- [3]Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30. https://doi.org/10.1257/jep.33.2.3
- [4]Acemoglu, D., & Robinson, J. A. (2000). Political losers as a barrier to economic development. American Economic Review, 90(2), 126–130. https://doi.org/10.1257/aer.90.2.126
- [5]Agarwal, N., Moehring, A., Rajpurkar, P., & Salz, T. (2023). Combining human expertise with artificial intelligence: Experimental evidence from radiology (NBER Working Paper No. 31422). National Bureau of Economic Research. https://doi.org/10.3386/w31422
- [6]American College of Radiology & Radiological Society of North America. (2020). Comments on the FDA public workshop “Evolving Role of Artificial Intelligence in Radiological Imaging” [Letter to the U.S. Food and Drug Administration, Docket No. FDA-2019-N-5592, 30 June 2020]. Retrieved October 28, 2026, from https://web.archive.org/web/20250102200744/https://www.acr.org/-/media/ACR/NOINDEX/Advocacy/acr_rsna_comments_fda-ai-evolvingrole-ws_6-30-2020.pdf
- [7]Baker, L. C. (2001). Managed care and technology adoption in health care: Evidence from magnetic resonance imaging. Journal of Health Economics, 20(3), 395–421. https://doi.org/10.1016/S0167-6296(01)00072-8
- [8]Baker, L. C., & Lamiraud, K. (2022). Adoption of hospital diagnosis-related group financing in Switzerland and the availability of computed tomography scanners. Health Economics, 31(12), 2537–2557. https://doi.org/10.1002/hec.4594
- [9]Becker, G. S. (1983). A theory of competition among pressure groups for political influence. Quarterly Journal of Economics, 98(3), 371–400. https://doi.org/10.2307/1886017
- [11]Chan, D. C., & Dickstein, M. J. (2019). Industry input in policy making: Evidence from Medicare. Quarterly Journal of Economics, 134(3), 1299–1342. https://doi.org/10.1093/qje/qjz005
- [12]Chen, B. K., Gertler, P. J., & Yang, C.-Y. (2016). Physician ownership of complementary medical services. Journal of Public Economics, 144, 27–39. https://doi.org/10.1016/j.jpubeco.2016.09.005
- [13]Coate, S., & Morris, S. (1995). On the form of transfers to special interests. Journal of Political Economy, 103(6), 1210–1235. https://doi.org/10.1086/601449
- [14]Comin, D., & Hobijn, B. (2009). Lobbies and technology diffusion. Review of Economics and Statistics, 91(2), 229–244. https://doi.org/10.1162/rest.91.2.229
- [15]Desmet, K., & Parente, S. L. (2014). Resistance to technology adoption: The rise and decline of guilds. Review of Economic Dynamics, 17(3), 437–458. https://doi.org/10.1016/j.red.2013.09.005
- [16]Dulisse, B., & Cromwell, J. (2010). No harm found when nurse anesthetists work without supervision by physicians. Health Affairs, 29(8), 1469–1475. https://doi.org/10.1377/hlthaff.2008.0966
- [17]Ellis, R. P. (1998). Creaming, skimping and dumping: Provider competition on the intensive and extensive margins. Journal of Health Economics, 17(5), 537–555. https://doi.org/10.1016/S0167-6296(97)00042-8
- [18]Ellis, R. P., & McGuire, T. G. (1986). Provider behavior under prospective reimbursement: Cost sharing and supply. Journal of Health Economics, 5(2), 129–151. https://doi.org/10.1016/0167-6296(86)90002-0
- [20]Fenton, J. J., Lee, C. I., Xing, G., Baldwin, L.-M., & Elmore, J. G. (2014). Computer-aided detection in mammography: Downstream effect on diagnostic testing, ductal carcinoma in situ treatment, and costs. JAMA Internal Medicine, 174(12), 2032–2034. https://doi.org/10.1001/jamainternmed.2014.5410
- [21]Fenton, J. J., Taplin, S. H., Carney, P. A., Abraham, L., Sickles, E. A., D’Orsi, C., Berns, E. A., Cutter, G., Hendrick, R. E., Barlow, W. E., & Elmore, J. G. (2007). Influence of computer-aided detection on performance of screening mammography. New England Journal of Medicine, 356(14), 1399–1409. https://doi.org/10.1056/NEJMoa066099
- [22]Grossman, G. M., & Helpman, E. (1994). Protection for sale. American Economic Review, 84(4), 833–850.
- [23]Iversen, T., & Ma, C.-t. A. (2022). Technology adoption by primary care physicians. Health Economics, 31(3), 443–465. https://doi.org/10.1002/hec.4447
- [24]Kleiner, M. M., Marier, A., Park, K. W., & Wing, C. (2016). Relaxing occupational licensing requirements: Analyzing wages and prices for a medical service. Journal of Law and Economics, 59(2), 261–291. https://doi.org/10.1086/688093
- [25]Lång, K., Josefsson, V., Larsson, A.-M., Larsson, S., Högberg, C., Sartor, H., Hofvind, S., Andersson, I., & Rosso, A. (2023). Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): A clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. Lancet Oncology, 24(8), 936–944. https://doi.org/10.1016/S1470-2045(23)00298-X
- [26]Lehman, C. D., Wellman, R. D., Buist, D. S. M., Kerl ikowske, K., Tosteson, A. N. A., & Miglioretti, D. L. (2015). Diagnostic accuracy of digital screening mammography with and without computer-aided detection. JAMA Internal Medicine, 175(11), 1828–1837. https://doi.org/10.1001/jamainternmed.2015.5231
- [27]Levaggi, R., Moretto, M., & Pertile, P. (2012). Static and dynamic efficiency of irreversible health care investments under alternative payment rules. Journal of Health Economics, 31(1), 169–179. https://doi.org/10.1016/j.jhealeco.2011.09.005
- [28]McMichael, B. J. (2017). The demand for healthcare regulation: The effect of political spending on occupational licensing laws. Southern Economic Journal, 84(1), 297–316. https://doi.org/10.1002/soej.12211
- [30]Palmetto GBA. (2019). Billing and coding: Repeat X-ray or EKG interpretations by same or different physician (Article A53423) [Medicare Administrative Contractor billing article quoting the Medicare Claims Processing Manual, Pub. 100-04, ch. 13, § 100.1]. https://www.cms.gov/medicare-coverage-database/view/article.aspx?articleid=53423
- [31]Plesner, L. L., Müller, F. C., Nybing, J. D., Laustrup, L. C., Rasmussen, F., Nielsen, O. W., Boesen, M., & Andersen, M. B. (2023). Autonomous chest radiograph reporting using AI: Estimation of clinical impact. Radiology, 307(3), Article e222268. https://doi.org/10.1148/radiol.222268
- [32]Price, W. N., II, Gerke, S., & Cohen, I. G. (2019). Potential liability for physicians using artificial intelligence. JAMA, 322(18), 1765–1766. https://doi.org/10.1001/jama.2019.15064
- [33]Selder, A. (2005). Physician reimbursement and technology adoption. Journal of Health Economics, 24(5), 907–930. https://doi.org/10.1016/j.jhealeco.2005.03.004
- [34]Smith, L. B. (2022). The effect of nurse practitioner scope of practice laws on primary care delivery. Health Economics, 31(1), 21–41. https://doi.org/10.1002/hec.4438
- [35]Traczynski, J., & Udalova, V. (2018). Nurse practitioner independence, health care utilization, and health outcomes. Journal of Health Economics, 58, 90–109. https://doi.org/10.1016/j.jhealeco.2018.01.001
