Federal healthcare agencies are accelerating efforts to expand artificial intelligence's role in healthcare, with clinical AI emerging as a top strategic priority for the Centers for Medicare & Medicaid Services.
At a healthcare AI event hosted by the Consumer Technology Association (CTA) in Washington, D.C., on Wednesday, Stephanie Carlton, deputy administrator at CMS, outlined the agency's evolving AI strategy as it aims to establish clearer pathways for AI market access and regulation while also exploring reimbursement frameworks for AI-enabled technologies.
Carlton also serves as chief clinical AI officer at CMS, a newly created position that signals the Trump administration's push to advance AI in clinical care.
A centerpiece of CMS' AI strategy is the ACCESS model—Advancing Chronic Care with Effective, Scalable Solutions. Carlton hinted that the agency would soon expand that model. Announced in December and launched in July, ACCESS is a new 10-year program for value-based chronic condition management that leans heavily on technology and AI to scale to large populations of patients.
"We have said that we want to announce more tracks and more opportunities to engage in that, so the team has been working very carefully on those additional tracks. We're still committed to an outcomes framework for patients. We like total cost of care. We like an outcomes framework, so we'll continue that theme," Carlton said at the CTA event but did not provide specific details on upcoming updates to the model.
Carlton said CMS' AI strategy consists of four primary pillars: building public trust, expanding healthcare data sharing and interoperability, establishing clearer pathways for AI market access and regulation and developing reimbursement frameworks for AI-enabled technologies.
Agency leaders are considering critical questions as the industry moves from general wellness and clinical decision support into higher-risk AI functions. Carlton said CMS is exploring how to regulate and deploy increasingly sophisticated AI tools, pointing to state-level pilot programs as one avenue for evaluating higher-risk use cases.
CMS is also considering how to develop reimbursement frameworks for these technologies, she noted. "When do we pay for it? When does that make sense? What's the difference between paying for things you can get in general apps that most of the frontier models make available quite cheaply versus what is actually performing a medical function and is reasonable and necessary for medical care, which is our standard," she said.
With the ACCESS model, more than 150 healthcare organizations were accepted to participate in the launch of the tech-enabled chronic care model. The model will provide predictable, recurring, outcomes-based payments for technology used to treat diabetes, hypertension, chronic kidney disease, obesity, depression and anxiety.
The ACCESS model is an early test of whether AI tools can improve measurable health outcomes, providing an evidence base and laying the groundwork for a broader role in managing total cost of care, Carlton asserted.
"It's the first step towards a greater vision of what is the role of AI in managing total costs of care that either holds quality constant or dramatically improves it, which I think it can," she said.
When asked what she thinks the ACCESS model will achieve in two years, by 2028, Carlton said, "I think we're hoping in December 2028 we will have seen that technology can have a massively deflationary impact on healthcare costs, and I think we've also done a lot of work with engaging tech companies in being partners and providers in Medicare, which you can do today through ACCESS, and we're testing to see how that works. That's a paradigm shift: paying tech companies versus just paying clinicians and healthcare facilities."
She stressed that as HHS pushes forward on AI, CMS is coordinating with other agencies. CMS' work dovetails with efforts at the Food and Drug Administration (FDA) to explore regulatory approaches to healthcare AI with a focus on safety and quality.
In August, the FDA issued a discussion paper on regulatory considerations for generative artificial intelligence-enabled medical devices. The FDA is seeking feedback on how to assess risks, evaluate the safety and effectiveness of GenAI-enabled medical devices before market authorization and monitor device performance after deployment. The FDA is seeking public comment on the discussion paper, which is open until Oct. 19.
The discussion paper outlines a possible framework for evaluating GenAI-based medical devices based on two factors: how independently they act, from providing information to taking actions, and the potential harm from incorrect outputs. The agency also said it was considering a "competency-based" evaluation framework, combining extensive benchmarking with real-world clinical validation.
Carlton compared the discussion paper's concept of benchmarking to a doctor earning a medical degree by demonstrating core competencies, and she likened clinical confirmation to a medical residency where AI tools are evaluated in real-world settings to ensure they are safe, effective and ready for broader use.
The FDA's current review framework is not well suited to generative AI, noted Rick Abramson, M.D., associate director for digital health and director of the Digital Health Center of Excellence in the FDA's Center for Devices and Radiological Health, during a panel with healthcare officials later in the day.
"I'm absolutely confident that the evidentiary standard for FDA authorization of generative AI tools will change because it's [a] square peg and round hole. Where I want the industry to contribute most and what I want to know is what do we have right, what do we have wrong, where do we need to improve, and what do we need to do together to advance the way that we think about this," Abramson said.
As the FDA works through a regulatory approach, CMS' role is to determine which health outcomes should be monitored and measured, Carlton noted. "That's an area we would love input on because we are working through it, and you'll hear more soon," she said. "We are working on a bit of a framework on how to evaluate that, but I think there are some principles around health outcomes, and then how we construct a model dictates which health outcomes we'll pay attention to. Is it primary care? Is it specialty care? Is it total cost of care? That has implications for which health outcomes we'd be monitoring."
CMS is also thinking through what reporting measures make sense for AI payment models to collect only the data needed to ensure the tech improves patient outcomes, rather than creating a burdensome new reporting infrastructure.
"At some level, we don't want to build a whole quality industrial complex around reporting. We want to have exactly what is necessary to make sure the technology is working and that we're not causing harm; that it's actually getting to a better-than-today standard for patients," she noted.
While regulators sketch out the roadmap for healthcare AI regulation and policy, physicians and health tech executives continue to debate the opportunities and risks of AI in real-world medical practice. At one point during the day, there was a tense discussion about the potential use of fully autonomous AI versus human-in-the-loop clinical care.
Even in the case of AI-enabled prescription refills, which is being tested in Utah, there are situations where a human physician's cognitive judgment is necessary to ensure safe patient care, noted Jesse Ehrenfeld, M.D., past president of the American Medical Association.
"I've had patients walk in the office, and you know something's not right. The clinical context has changed. There's something that says that a refill is not safe today. You know, there's a tremor. There's something else going on, and that's when you execute your cognitive function to decide how do I manage what's happening in front of me. There are pieces of that that are just impossible to automate because you don't have the context fed into an autonomous system, and that's where we want to marry the best of both. We want to scale capacity, but we also make sure that we don't lose the opportunity to escalate as appropriate," said Ehrenfeld, who is now global chief medical officer at Aidoc.
Marc Paradis, principal and founder of SIYOM Consulting, argued that discussions should focus on AI's future potential rather than its present limitations. He suggested that AI will soon go beyond today's text-based large language models and will be able to detect subtle clinical signals, such as changes in voice, tremor or other physiological markers.
"One of the things you also have to remember about AI down the line is it can eventually move into areas where humans can't see, can't smell, can't touch, can't think, can't feel, and that's the world I think we should be talking about now because that world will come, and it will come much faster than we think," Paradis said, noting the potential for innovations like hyperspectral imaging during surgery, enabling the AI to identify patterns and abnormalities that clinicians cannot see.
John Whyte, M.D., CEO of the American Medical Association, immediately pushed back, arguing that AI and human intelligence are fundamentally different and cautioning against overstating AI's capabilities without strong evidence.
"What we should be talking about is what is the evidence base that we need to make decisions as it relates to safety, as it relates to patient outcomes," Whyte said. AI should be held to the same evidence standards as other healthcare interventions, he argued, also noting that healthcare is longitudinal and often extends beyond isolated tasks.
Laura Adams, senior advisor at the National Academy of Medicine, also challenged the common assumption that clinicians should always remain "in the loop" when AI is used in patient care. Requiring physician review of every AI decision could create unnecessary delays and inefficiencies, particularly in cases where AI has already demonstrated strong performance, such as image analysis, Adams said.
The goal, she argued, should be to determine when clinician involvement meaningfully improves outcomes and when it may actually reduce the benefits of AI.