Why fee-for-service could push AI's health care costs up, not down

Artificial intelligence's promise in medicine has always come with an asterisk: the technology can only improve care as fast as the payment system lets it. Caroline Pearson, executive director of the Peterson Health Technology Institute (PHTI), a nonprofit that evaluates the clinical and economic value of emerging health technologies, has spent much of the past year studying that gap. PHTI's report on AI and reimbursement concludes that clinical AI has real potential to improve outcomes, but that current fee-for-service payment models offer physicians few incentives to adopt it responsibly — and in some cases risk driving costs higher rather than lower.

The stakes for physicians are immediate. Two-thirds of doctors already expect AI fluency to boost their earnings, but that expectation depends on unresolved questions about who captures the value AI creates. Meanwhile, liability and reimbursement barriers remain among the biggest obstacles to real-world AI deployment, even as the technology itself keeps advancing.

The PHTI report digs into where AI payment policy is headed next: the difference between assistive and autonomous AI, new Medicare models that could let technology vendors bill payers directly, who bears liability when an algorithm gets it wrong, and why accountable care organizations — seemingly natural early adopters — have been slower to embrace AI than expected.

Medical Economics spoke with Pearson about the report to learn more.

(Editor's note: The following transcript has been edited for brevity and clarity.)

Medical Economics: For physicians, what should be the biggest takeaway from your report on AI and reimbursement?

Caroline Pearson: We have found that the potential for clinical AI to improve care is tremendous, but under the current payment models, we often lack the incentives to adopt that technology, and the payment models that we do have available to us risk really raising health care costs. So we're calling for a real reconsideration of how we want to pay for health tech to achieve the benefits that we all are looking for.

Medical Economics: The report mentions that AI should be lowering costs, not inflating them. Are there any signs that an AI payment model is heading toward inflation instead of lowering costs?

Pearson: So today we still have most of our health care that is paid for in fee-for-service models, and those payment rates are generally set on the basis of clinician time and effort that is expected for any given service or intervention. The tricky thing about technology is that, of course, the marginal cost of deploying the technology is relatively low, and so you actually have low time and effort, but an ability to bill a service many, many times, very cheaply. And so you just sort of do the math and say, okay, technology on a fee-for-service chassis could really increase costs, and we do see that happening in cases like remote patient monitoring and some other limited uses. And so you really want to say, how do we think about the outcomes that we're looking for in technology-based care, and how can we encourage a more value-based payment system?

Medical Economics: If AI lets doctors see more patients for the same pay per visit, is that bad for physicians, or good for them and bad for the system?

Pearson: Well, generally, if we could see more patients, if we actually improved access to care, that would be great, because we could stretch our precious health care workforce further. So I think that's not really the source of the concern. The source of the concern is really where AI may be increasing the amount of revenue per visit, but not actually improving the number of patients that are being seen or improving the clinical outcomes in those same visits.

Medical Economics: The report talks about assistive AI and autonomous AI. Can you explain the difference, and does the former eventually lead to the latter?

Pearson: I think in some cases it will, and in some cases we may always want to prioritize assistive AI. So assistive AI is really human clinicians using AI to make their job faster, more efficient or more accurate. There are administrative tools to help with diagnosis or research on appropriate treatment patterns, and there can be management tools that extend care into the patient environment. So those would all be assistive — they're still being deployed, overseen, and billed by the supervising physician. Autonomous AI would really enable technology to deliver some facets of care independently, and I think the easiest way to think about that is certain common forms of medication prescribing or medication titration. Think about some urgent care use cases, or even something like titrating medications for hypertension, where an AI tool could be doing that independently while the patient remains under the primary care of their PCP.

Medical Economics: Do you think there will ever be a time when billing codes are modified so doctors get paid for reviewing an AI recommendation?

Pearson: Yeah, we already see some codes like that. Remote patient monitoring pays physicians for not only reviewing patient data, but often we're seeing AI analysis of that data. So patients may be checking their blood pressure or their blood sugar at home, readings are being uploaded, and then the AI is reviewing that data and flagging things for the clinician. There are certainly other cases where you could have AI proposing diagnoses or treatment plans, and the doctor being paid to oversee it. So some of those cases are going to be important, but we need to think about both what is the right value for those payments, and how do we make sure that the clinician oversight is very effective. Obviously, we know that the more we use AI, sometimes it's hard for people to stay cognitively focused, and so we want to make sure that we're keeping our humans sharp, focused and spending their time on the highest-value clinical interventions.

Medical Economics: If an autonomous AI made a bad call on, say, a patient's medication, who would be liable — the doctor, the health system or the AI company?

Pearson: This is a major open area for legal work. In most cases, right now, it's not clear. Generally speaking, AI solutions like the prescribing need to operate under the physician's malpractice insurance, and that would be the physician's liability. Certainly, physicians are concerned about that — they're not inside the black box of the algorithm. And so I think in many cases we are seeing a push to say, how can that liability move to the technology vendor, and what would that look like? But we don't have the legal infrastructure in most cases to do that today. So this is an area of a lot of additional work and regulation that's going to be needed.

Medical Economics: Is there any chance that AI companies could eventually bill payers directly and cut physicians out of the revenue?

Pearson: There is some potential for that. We've seen a couple of examples where things are headed that direction. One example is the new Medicare model called the Access Model, which creates direct payments for technology-based care — technology companies can sign up to be paid directly by Medicare for services provided for chronic care management if they deliver clinical outcomes as set forth by the program, and there doesn't need to be any sort of traditional Medicare physician in the mix on that. So that's really the beginning in the Medicare program. The other example we saw recently is with Doctronic, which has been doing a pilot program for medication prescribing in Utah, where the model would be that those prescriptions get reimbursed by a payer directly. So I think there's a lot of interest in this, and this is where we've been calling out the need to reimagine not only the rules around when we think that would be beneficial, safe, and effective, and make sure that we preserve the role of the physician, but also what's the right payment level for that. Because, as I sometimes say, technology doesn't need a living wage, and if we're going to shift care from physicians to technology, we want to do that in a way that is very cost effective, so that we free up resources to spend on other high-value care elsewhere.

Medical Economics: The report mentioned that ACOs didn't always embrace the technology, even when the incentives seemed to line up. Why was that?

Pearson: This is a bit of a mystery, because you would hope that the ACOs would be sort of the first place where you would see some of the tech adoption, but I think this is one of the challenges with where we are in the development. In a fee-for-service model, lots of health systems are finding ways to deploy technology to increase their revenue, but in an ACO or other risk-based payment model, you really have to be confident that the technology is improving the clinical outcomes and reducing total cost of care. There's not a pure revenue play, and generally I think the ACOs have not found that the evidence available to them about the performance of these technologies is compelling enough to drive further adoption. So that is certainly something that we want to both encourage more evidence generation on, and then really educate those ACOs about which technologies might be worth integrating into their care.

Medical Economics: Is there anything else doctors need to know about this report that we haven't talked about?

Pearson: Well, I think this is a rapidly changing environment. There's going to be lots to track and to worry about and to learn over the next few years. And so I would highlight that another important component, in addition to payment, is really going to be how we do change management — how do we support the existing clinical workforce in understanding how to use these tools and how to redesign their own care delivery to maximize the benefit of the technology. I think we want to encourage folks to embrace the technology, but that really does require redesigning workflows and rethinking how clinicians are spending their time. So we're going to need to devote time and effort to that, both at a system level and at an individual provider level, to make sure that we can do this right.