Artificial Intelligence in Health Care

June 10, 2026

Key Findings

Survey respondents were largely unsure about whether adoption of agentic clinical A.I. would substantially improve patient access to basic primary care services, with more than 41% of panelists expressing uncertainty about the technology’s effects. Several panelists noted that primary care is fundamentally relationship-based and that A.I. is likely to be used for narrow clinical tasks as opposed to comprehensive care in the short-term. In the longer-term, panelists noted that A.I.’s impact will depend on regulatory policy and technological improvements.
Nearly half (46%) of experts agreed or strongly agreed that F.D.A. regulation of clinical A.I. algorithms should emphasize post-market surveillance over premarket clearance; 29% disagreed or strongly disagreed, and another 20% were uncertain. A number of respondents indicated that both pre- and post-market oversight were important, and that the emphasis on one versus the other may vary based on the specific algorithm. Some experts expressed concerns about the feasibility of post-market surveillance, given the large number of algorithms.
Most experts (83%) were either uncertain or disagreed that legislation that would make F.D.A.-authorized A.I. algorithms eligible to prescribe medications would safely improve patient access to needed drugs. Panelists noted that such a policy would need substantial safeguards and that it was premature, given lack of clinical evidence and real-world experience.

Read the full Health Affairs Forefront summary of results here

Survey Questions

Please note: our surveys will consistently use two modifiers to describe the size of an effect: “Substantial”: when an effect is large enough to meaningfully influence policy decisions, program implementation, or outcomes of interest “Measurable”: when the direction of an effect is clear, but the effect may not be sufficiently large to make much of a difference for a given policy, program, or outcome. Weighting by confidence level removes responses of “don’t know” (weight equals 0). Unweighted responses with responses of “don’t know” removed were qualitatively similar to weighted responses. Unweighted results including “don’t know” responses are presented in the exhibits below.

Response rate of 94% (59 out of 63 panelists responded)

The use of artificial intelligence (A.I.) in medicine is rapidly increasing, with applications in health care administration, clinical decision support, and patient engagement, among others. Some observers have argued that A.I. technologies will substantially improve the quality, availability, and efficiency of care, while others have raised concerns about potential issues related to safety, effectiveness, bias, and workforce disruption. Regulatory considerations around clinical A.I. are in flux as the technology and its clinical applications continue to evolve.

Question 1: Adoption of agentic clinical A.I. will substantially improve patient access to basic primary care services in regions with workforce shortages:

a. Strongly agree
b. Agree
c. Uncertain
d. Disagree
e. Strongly disagree
f. Don’t know

Question 2: F.D.A. regulation of clinical A.I. algorithms should emphasize post-market surveillance over premarket clearance.

a. Strongly agree
b. Agree
c. Uncertain
d. Disagree
e. Strongly disagree
f. Don’t know

Question 3: A Congressional bill (H.R. 238, Healthy Technology Act) would allow A.I. algorithms that have been authorized by the F.D.A. to be eligible for licensing to prescribe medications. If passed, such a policy would safely improve patient access to needed drugs.

a. Strongly agree
b. Agree
c. Uncertain
d. Disagree
e. Strongly disagree
f. Don’t know

Individual Survey Responses

Question One

Adoption of agentic clinical A.I. will substantially improve patient access to basic primary care services in regions with workforce shortages.

Name	Vote	Confidence	Comments
Margarita Alegria	Agree	7
David Asch	Strongly agree	8
John Ayanian	Disagree	7
Peter Bach	No response
Laurence Baker	Uncertain	5
David Blumenthal	Uncertain	7	As someone who studies this subject, and is a primary care physician, I think it is premature to judge what impact AI will have on the availability, quality and equity of primary care services.
Erin Fuse Brown	Disagree	6	Disagree that the impact of agentic AI will substantially improve patient access. At best it’s uncertain, but if it’s like other clinical technology innovations (EHRs, telehealth) it may not be able to close the gap caused by practitioner shortages.
Melinda Buntin	Agree	6
Michael F. Cannon	Strongly agree	6	In the long run.
Lawrence Casalino	Uncertain	5
Amitabh Chandra	Don’t know	0
Lanhee J. Chen	Uncertain	6
Michael Chernew	Agree	8
Janet Currie	Uncertain	6	I’m not sure AI can be trusted to deliver primary care without substantial oversight, and the understaffed places might be the least able to provide that oversight.
Lesley Curtis	Agree	6
David Cutler	Agree	6
Julie Donohue	Disagree	8	There is no evidence that companies marketing AI clinical services will prioritize primary care services, or primary care shortage areas. In fact, financial incentives may run counter to marketing primary care services.
Joseph Doyle	Agree	8
David Dranove	Agree	8
Stacie Dusetzina	Strongly disagree	5
Jose Esarce	Disagree	7
Elliott Fisher	Uncertain	10
Richard Frank	Uncertain	2
Craig Garthwaite	Uncertain	6	The answer here depends on how regulation permits the use of AI to leverage non-physician providers in offering medical services. AI-augmented mid-level providers could expand access but would require some regulatory relief.
Darrell Gaskin	Strongly disagree	10	Primary care is about human contact. A.I. will not be able to build trust with patients.
Martin Gaynor	Disagree	7	AI will initially be used to augment or substitute in familiar tasks and where it’s profitable. Regrettably, I don’t see expanding access fitting those criteria.
Sherry Glied	Disagree	7
David Grabowski	Disagree	7	I think “substantially improve” is too strong. It might improve certain aspects of care but it will not be uniform.
Jonathan Gruber	Strongly agree	7	Depends critically on evolution of scope of practice laws.
Vivian Ho	Uncertain	10	It depends on what regulations are adopted. Incumbents could introduce barriers to entry.
Jason Hockenberry	Uncertain	9
Haiden Huskamp	Agree	6
Benedic Ippolito	Uncertain	5
Anupam Jena	Uncertain	7
Nancy Keating	Disagree	3
Aaron Kesselheim	Uncertain	10
Jonathan Kolstad	Uncertain	9	Much of clinical medicine is not easy accessible to language-based AI. It is likely a new round of AI that is more native to the health care setting, not just agentic AI as it exists.
R Tamara Konetzka	Uncertain	7	I’m confident we are uncertain. It could help but will require substantial oversight.
Rick Kronick	Uncertain	1
Valerie Lewis	Disagree	5	I have seen no data to this effect, and even the notion that agentic AI can supplant basic primary care sources suggests introducing a two tiered system (those that get human care and those that get AI care, even if with some human oversight) that requires significant ethical interrogation. We have a long history of racing to deploy scientific advances before reckoning with their ethics and the human cost of getting that sequence wrong should give us serious pause before we make underserved regions the proving ground for agentic clinical AI’s role in primary care.
Nicole Maestas	Uncertain	7
Tom McGuire	Strongly agree	8
Ellen Meara	Uncertain	5	I think it will measurably increase access, but countervailing forces, like use of AI in prior authorization, will hinder access.
Ateev Mehrotra	Uncertain	6	Our hope with many new technologies such as AI is that they will differentially help those who live in underserved communities. However, the research generally shows the opposite. These technologies are differentially used by those who live in wealthier, urban communities. Having all said that, it is possible that AI is different. Also, it may be possible that it improves access in communities with workforce shortage but still worsens existing disparities.
David Meltzer	Strongly disagree	9
Joseph Newhouse	Uncertain	8	I’m not sure if there is a difference between uncertain and don’t know.
Sean Nicholson	Strongly agree	8
Steve Parente	No response
Stephen Patrick	Uncertain	8	Access to basic primary care is less clear. There are clear use cases throughout medicine, but access to primary care may or may not be one. At present, clinical AI is most useful as an assistant rather than a replacement. I’m skeptical that will change for some areas on medicine, especially primary care.
Harold Pollack	No response
Daniel Polsky	Uncertain	8	I’ve very confident in ‘uncertain’. The potential is there, but there is no way to know how this will play out. If I were to guess, the services may be hard to reimburse which would push against access.
Ninez Ponce	Agree	7
Thomas Rice	Disagree	3
Meredith Rosenthal	Disagree	7	It is hard to predict what AI will do in 5+ years but in the short run it is poised to be a complement rather than a substitute for human labor
Joseph Ross	Disagree	8	I am disagreeing because I don’t believe that agentic AI provides primary care services. Might it automatically refill medication orders or order screening tests, yes? But will it engage patients in counseling and management decisions, not in the way that I define primary care.
Brendan Saloner	Uncertain	5	It depends on how much patients feel that they can trust agents, how accurate and vetted the agent’s information is, and how strong the regulatory oversights are.
Kosali Simon	No response
Jon Skinner	Agree	8
Ben Sommers	Uncertain	5	Does an AI chatbot providing guidance on whether someone needs to see a physical provider count as access?
Neeraj Sood	Strongly agree	6
David Stevenson	Disagree	3
Kevin Volpp	Strongly agree	8	Huge gaps in care could likely be addressed if agentic AI is combined with opt out type defaults
Rachel Werner	Don’t know	0		7
			Strongly agree	8
			Disagree	7
			No response
			Uncertain	5
			Uncertain	7	As someone who studies this subject, and is a primary care physician, I think it is premature to judge what impact AI will have on the availability, quality and equity of primary care services.
			Disagree	6	Disagree that the impact of agentic AI will substantially improve patient access. At best it’s uncertain, but if it’s like other clinical technology innovations (EHRs, telehealth) it may not be able to close the gap caused by practitioner shortages.
			Agree	6
			Strongly agree	6	In the long run.
			Uncertain	5
			Don’t know	0
			Uncertain	6
			Agree	8
			Uncertain	6	I’m not sure AI can be trusted to deliver primary care without substantial oversight, and the understaffed places might be the least able to provide that oversight.
			Agree	6
			Agree	6
			Disagree	8	There is no evidence that companies marketing AI clinical services will prioritize primary care services, or primary care shortage areas. In fact, financial incentives may run counter to marketing primary care services.
			Agree	8
			Agree	8
			Strongly disagree	5
			Disagree	7
			Uncertain	10
			Uncertain	2
			Uncertain	6	The answer here depends on how regulation permits the use of AI to leverage non-physician providers in offering medical services. AI-augmented mid-level providers could expand access but would require some regulatory relief.
			Strongly disagree	10	Primary care is about human contact. A.I. will not be able to build trust with patients.
			Disagree	7	AI will initially be used to augment or substitute in familiar tasks and where it’s profitable. Regrettably, I don’t see expanding access fitting those criteria.
			Disagree	7
			Disagree	7	I think “substantially improve” is too strong. It might improve certain aspects of care but it will not be uniform.
			Strongly agree	7	Depends critically on evolution of scope of practice laws.
			Uncertain	10	It depends on what regulations are adopted. Incumbents could introduce barriers to entry.
			Uncertain	9
			Agree	6
			Uncertain	5
			Uncertain	7
			Disagree	3
			Uncertain	10
			Uncertain	9	Much of clinical medicine is not easy accessible to language-based AI. It is likely a new round of AI that is more native to the health care setting, not just agentic AI as it exists.
			Uncertain	7	I’m confident we are uncertain. It could help but will require substantial oversight.
			Uncertain	1
			Disagree	5	I have seen no data to this effect, and even the notion that agentic AI can supplant basic primary care sources suggests introducing a two tiered system (those that get human care and those that get AI care, even if with some human oversight) that requires significant ethical interrogation. We have a long history of racing to deploy scientific advances before reckoning with their ethics and the human cost of getting that sequence wrong should give us serious pause before we make underserved regions the proving ground for agentic clinical AI’s role in primary care.
			Uncertain	7
			Strongly agree	8
			Uncertain	5	I think it will measurably increase access, but countervailing forces, like use of AI in prior authorization, will hinder access.
			Uncertain	6	Our hope with many new technologies such as AI is that they will differentially help those who live in underserved communities. However, the research generally shows the opposite. These technologies are differentially used by those who live in wealthier, urban communities. Having all said that, it is possible that AI is different. Also, it may be possible that it improves access in communities with workforce shortage but still worsens existing disparities.
			Strongly disagree	9
			Uncertain	8	I’m not sure if there is a difference between uncertain and don’t know.
			Strongly agree	8
			No response
			Uncertain	8	Access to basic primary care is less clear. There are clear use cases throughout medicine, but access to primary care may or may not be one. At present, clinical AI is most useful as an assistant rather than a replacement. I’m skeptical that will change for some areas on medicine, especially primary care.
			No response
			Uncertain	8	I’ve very confident in ‘uncertain’. The potential is there, but there is no way to know how this will play out. If I were to guess, the services may be hard to reimburse which would push against access.
			Agree	7
			Disagree	3
			Disagree	7	It is hard to predict what AI will do in 5+ years but in the short run it is poised to be a complement rather than a substitute for human labor
			Disagree	8	I am disagreeing because I don’t believe that agentic AI provides primary care services. Might it automatically refill medication orders or order screening tests, yes? But will it engage patients in counseling and management decisions, not in the way that I define primary care.
			Uncertain	5	It depends on how much patients feel that they can trust agents, how accurate and vetted the agent’s information is, and how strong the regulatory oversights are.
			No response
			Agree	8
			Uncertain	5	Does an AI chatbot providing guidance on whether someone needs to see a physical provider count as access?
			Strongly agree	6
			Disagree	3
			Strongly agree	8	Huge gaps in care could likely be addressed if agentic AI is combined with opt out type defaults
			Don’t know	0

Question Two

F.D.A. regulation of clinical A.I. algorithms should emphasize post-market surveillance over premarket clearance.

Name	Vote	Confidence	Comments
Margarita Alegria	Agree	8
David Asch	Agree	7
John Ayanian	Agree	5
Peter Bach	No response
Laurence Baker	Disagree	7
David Blumenthal	Disagree	8	AI is multiple technologies. Premarket clearance works for some, but not others. Post-market surveillance is important for all.
Erin Fuse Brown	Agree	4	FDA regulation should do both initial premarket clearance and robust post-market surveillance. Unlike static devices, premarket clearance alone is insufficient– however, this doesn’t mean that FDA should de-emphasize premarket review, however, it should engage in robust post-market surveillance (pharmacovigilence, safety monitoring, learning systems oversight) to track how the dynamic nature of AI algorithms is operating in practice. However, pre-market clearance for every new iteration of the technology is likely unworkable.
Melinda Buntin	Disagree	4
Michael F. Cannon	Strongly agree	10	If we’re in the realm of “should,” the FDA should neither impose barriers to AI entering markets nor conduct postmarket surveillance of AI. Even though I oppose both, I answered “strongly agree” because premarket barriers are far worse.
Lawrence Casalino	Uncertain	5	Algorithms will change so rapidly that hard to see how FDA can approve.
Amitabh Chandra	Uncertain	5
Lanhee J. Chen	Agree	7
Michael Chernew	Strongly agree	7
Janet Currie	Agree	8	There is so much we don’t understand about how AI will work in practice that it seems foolish to believe that premarket clearance will identify all possible problems.
Lesley Curtis	Agree	7
David Cutler	Uncertain	6
Julie Donohue	Strongly disagree	8	We need a completely new regulatory regime to address AI clinical services that includes both pre- and post-marketing components.
Joseph Doyle	Agree	8	Human decision makers will be adopting tools to improve quality and lower costs. We need to learn what works best, and that takes experimentation.
David Dranove	Strongly agree	9
Stacie Dusetzina	Disagree	7	Both should be emphasized.
Jose Esarce	Disagree	7
Elliott Fisher	Disagree	3
Richard Frank	Disagree	7
Craig Garthwaite	Uncertain	6
Darrell Gaskin	Uncertain	5
Martin Gaynor	Don’t know	0
Sherry Glied	Agree	6
David Grabowski	Agree	7
Jonathan Gruber	Uncertain	5	We should shift towards post-market, but not sure it should be emphasized.
Vivian Ho	Uncertain	10	I don’t know enough about clinical algorithms to make an informed judgement on this question. And I’m wondering whether the FDA has sufficient expertise either.
Jason Hockenberry	Strongly disagree	7	It is still unclear whether current models can distinguish between quality of evidence.
Haiden Huskamp	Agree	7
Benedic Ippolito	Agree	4
Anupam Jena	Uncertain	7
Nancy Keating	Uncertain	3
Aaron Kesselheim	Disagree	9	While post-market surveillance is critical, we should not abandon pre-approval testing and feedback
Jonathan Kolstad	Strongly agree	10
R Tamara Konetzka	Agree	8
Rick Kronick	Agree	8
Valerie Lewis	Uncertain	10	I don’t know enough about the FDA regulatory process to weigh in confidently on this. In addition, this question phrases these as a zero sum trade off. We absolutely need post-market surveillance, but also not at the expense of premarket clearance.
Nicole Maestas	Don’t know	0
Tom McGuire	Agree	8
Ellen Meara	Agree	5
Ateev Mehrotra	Disagree	6	The idea of post-market clearance is great in theory, but may be impractical in many cases. I worry that it means that a lot of AI algorithms get released and widely adopted before we know they have harmed people. An underlying assumption with post-market surveillance is that algorithms are low risk. I worry that assumption may be incorrect.
David Meltzer	Strongly disagree	7
Joseph Newhouse	Uncertain	8	Obviously the FDA could do both; I don’t think enough is now known to emphasize one strategy over the other and the optimal strategy may well be an adaptive strategy, emphasizing pre-market for agents that do poorly on initial pre-market testing and conversely.
Sean Nicholson	Agree	7	The FDA cannot keep pace with AI evolution and, importantly, providers who implement these new services have incentives to carefully consider patient safety and unintended consequences.
Steve Parente	No response
Stephen Patrick	Agree	7
Harold Pollack	No response
Daniel Polsky	Agree	6
Ninez Ponce	Uncertain	5
Thomas Rice	Strongly disagree	6
Meredith Rosenthal	Agree	8
Joseph Ross	Agree	9	Agree, but this is mostly because algorithms change so frequently in response to new data and inputs. So to monitor for clinical validity and utility, and ensure that the algorithm is not drifting or embedding bias, postmarket surveillance is the only choice.
Brendan Saloner	Disagree	7	It’s very hard to put the genie back in the bottle — the FDA needs to pursue regulation on both tracks with vigor (and catch up, they are already behind!)
Kosali Simon	No response
Jon Skinner	Strongly agree	8
Ben Sommers	Disagree	5
Neeraj Sood	Disagree	6
David Stevenson	Disagree	3
Kevin Volpp	Agree	7
Rachel Werner	Don’t know	0

Question Three

A Congressional bill (H.R. 238, Healthy Technology Act) would allow A.I. algorithms that have been authorized by the F.D.A. to be eligible for licensing to prescribe medications. If passed, such a policy would safely improve patient access to needed drugs.

Name	Vote	Confidence	Comments
Margarita Alegria	Uncertain	5
David Asch	Disagree	5
John Ayanian	Disagree	5
Peter Bach	No response
Laurence Baker	Uncertain	5
David Blumenthal	Uncertain	9	Again, I know too much. We are just receiving un-peer-reviewed data from the first effort to allow AI to prescribe, Doctronics. Premature to reach conclusions.
Erin Fuse Brown	Disagree	6	FDA approval is not a substitute for professional clinical judgment of a prescriber. The policy proposed in HR 238 contains too few safeguards to ensure that drugs are safely prescribed.
Melinda Buntin	Strongly agree	6
Michael F. Cannon	Uncertain	10	Not familiar with the bill. All available evidence shows that prescription regulation fails to achieve its goals. Congress should eliminate it. Short of that, anything that reduces, avoids, or eliminates the barriers prescription regulation puts in the way of patients accessing the medicines they want is worth doing.
Lawrence Casalino	Uncertain	5
Amitabh Chandra	Uncertain	6
Lanhee J. Chen	Uncertain	7
Michael Chernew	Agree	8	But would come with meaningful risks that need to be addressed
Janet Currie	Disagree	7	While many clinicians make questionable prescribing decisions, or fail to consider things like drug interactions, AI should not be used to prescribe medications without oversight.
Lesley Curtis	Uncertain	7
David Cutler	Uncertain	6	It depends in part on how frequently the algorithm is updated and whether pharmaceutical coimpanies are allowed to pay for placement on the app.
Julie Donohue	Uncertain	9	AI licensed prescribers may simply substitute for human prescribers by offering services at a lower price. Inequities in access to essential medicines may increase if low-income, uninsured individuals are not able to afford these services. The effects from this change depend on payment structure, organizational and oversight. FDA making a change in the absence of payment reform will not solve the affordability issue.
Joseph Doyle	Disagree	6
David Dranove	Agree	8
Stacie Dusetzina	Strongly disagree	7	“Safely improve” is the main concern / disagreement that I have. Also, relative to what – limits on prescribing practices for non MD clinicians?
Jose Esarce	Strongly disagree	8
Elliott Fisher	Strongly disagree	9
Richard Frank	Uncertain	2
Craig Garthwaite	Disagree	5	I don’t see a lot of high quality evidence that this is a stumbling block to prescriptions.
Darrell Gaskin	Disagree	8	Using A.I. to prescribe medication is tantamount to self prescribing.
Martin Gaynor	Uncertain	5	I doubt that the FDA examines and tests AI in ways that are relevant and important for this issue in deciding on clearances.
Sherry Glied	Uncertain	4
David Grabowski	Disagree	7
Jonathan Gruber	Uncertain	5
Vivian Ho	Uncertain	10	Access will depend on a wider set of regulations. Incumbents (I’m thinking physician associations) can construct barriers to entry.
Jason Hockenberry	Uncertain	9
Haiden Huskamp	Disagree	7
Benedic Ippolito	Disagree	3	Among other concerns, I expect insurers to take a skeptical view towards AI-generated prescriptions and adjust utilization management accordingly before blindly paying.
Anupam Jena	Uncertain	7
Nancy Keating	Disagree	3	Not sure how this would work, but I would want to see studies to demonstrate safety.
Aaron Kesselheim	Disagree	8	Maybe eventually with appropriate guiderails, but such a step seems too premature at this point
Jonathan Kolstad	Agree	9
R Tamara Konetzka	Uncertain	7	Again depends on degree of oversight. Not clear that the workforce bottleneck is in prescribing.
Rick Kronick	Disagree	8
Valerie Lewis	Strongly disagree	7	It is hard to see given the science on AI (and the lack of science in many domains) how the FDA could possibly be up to speed enough on this to adequately screen and license algorithms to prescribe medications. (It’s mind blowing to me that for example this would mean from a federal perspective there are many states where an NP, psychologist, or pharmacist cannot prescribe but an AI algorithm necessarily developed very recently would be trusted with this). I cannot imagine enough safeguards on this given the state of AI at the moment that this would be viable in any setting. The agentic piece of AI simply is not advanced enough or remotely vetted enough to handle this.
Nicole Maestas	Uncertain	7
Tom McGuire	Agree	8
Ellen Meara	Uncertain	5
Ateev Mehrotra	Uncertain	6	The key word in the question that tripped me up was “safely.” We don’t know how effective they will be and what error rate is acceptable. I think it is important to test this idea, but rigorous testing is key.
David Meltzer	Agree	5
Joseph Newhouse	Disagree	5	I doubt it would on net much affect access but it may improve prescribing, including reducing overprescribing (polypharmacy).
Sean Nicholson	Disagree	7	This could improve provider efficiency but is not likely to affect access substantially. Getting a physician to refill a prescription is not a major access barrier.
Steve Parente	No response
Stephen Patrick	Strongly disagree	8	I’ve not seen clear safety analyses. Perhaps there are data out there, however, there are elements that seem like they are moving faster than safety analyses … this is likely one of them.
Harold Pollack	No response
Daniel Polsky	Agree	7
Ninez Ponce	Uncertain	5
Thomas Rice	Uncertain	5
Meredith Rosenthal	Uncertain	10	Again, I think AI will mostly be improve quality and efficiency but not access.
Joseph Ross	Strongly disagree	8	Allowing AI algorithms to newly prescribe medications presents substantial risk for misuse and waste.
Brendan Saloner	Disagree	6	In theory this could work, but the current FDA is far too lax in its oversight and the Trump Administration is not taking any strong lines with the AI industry.
Kosali Simon	No response
Jon Skinner	Disagree	5
Ben Sommers	Uncertain	5
Neeraj Sood	Agree	6
David Stevenson	Disagree	3
Kevin Volpp	Agree	5
Rachel Werner	Don’t know	0

Meet our panel of experts

Cornell Health Policy Center

Cornell Health Policy Center

Artificial Intelligence in Health Care

Key Findings

Survey Questions

Individual Survey Responses

Question One

Question Two

Question Three

Recent Surveys

Private Equity in Health Care

Vaccine Injury Compensation Program

Most Favored Nation Drug Pricing

Affordable Care Act Subsidies

Medicare Advantage