Autonomous Clinical AI Could Break Fee-for-Service, Peterson Institute Warns

Healthcare's next big cost problem may not be the technology itself. It may be how we pay for it.

That is the core thesis of a new report from the Peterson Health Technology Institute, an independent non-profit that evaluates digital health systems and advises healthcare decision-makers. The finding is stark: existing reimbursement models, built around human clinician time and effort, were never designed for AI that can act with minimal or no supervision. Apply those models unchanged, and the report warns spending could rise without any corresponding gain in outcomes.

For investors and health system executives tracking the clinical AI buildout, this is not a peripheral compliance issue. Payment policy will decide whether AI becomes a genuine cost-and-access lever or simply another billable layer stacked on top of an already expensive system.

"We need to get payment right to expand access to high-quality care while lowering costs," said Caroline Pearson, PHTI's executive director.

The fee-for-service math problem

Traditional healthcare spending carries a built-in governor: clinicians can only work so many hours a day. Fee-for-service payment, tied directly to that time and effort, has an implicit ceiling. Autonomous AI removes it.

"The biggest risk of paying for clinical AI through existing fee-for-service mechanisms is that total healthcare spending will increase rapidly," Pearson said. Today's spending is "somewhat constrained by supply limits on clinician effort," she noted, but AI systems face no such constraint. They can generate clinical activity "far more" than any human workforce, without a proportional rise in clinician hours. Tying payment to volume, she said, "risks inflating spending at a scale fee-for-service was never built for."

The report draws a sharp line between two categories of AI. Assistive AI, where a clinician reviews and approves every action, fits reasonably well into current billing structures. A physician's time is still the unit being paid for; the AI is simply a tool alongside it.

Autonomous AI is a different animal entirely. Pearson's example is hypertension management: a well-protocolized condition where algorithms can adjust medication based on remote blood pressure readings with little to no human intervention. The clinical logic works. The billing logic does not. There is no clinician time to log, no encounter code to file, and often no mechanism to pay the technology developer directly at all.

That gap is not a minor technicality. It is the reason PHTI argues for an entirely new payment architecture, one that rewards measured results rather than logged activity, with pricing that adjusts as evidence accumulates and clinical roles shift. "With technology, we can both collect outcome measures in real time and set payment levels tied to those outcomes," Pearson said.

Outcomes-based payment is not a universal fix, however. It works cleanly for chronic diseases with well-defined metrics, blood pressure, glucose control, and similar measurable endpoints. It "begins to fall apart," in Pearson's words, in primary care or prevention, where outcomes may not materialize for years. That unevenness is precisely why PHTI insists no single reimbursement model will cover every AI use case.

There is also a harder, more practical wrinkle: adoption economics. Health systems face substantial upfront licensing and implementation costs before any AI tool generates a dollar of return, and those returns are often diffuse, shared across payers, providers and patients who did not necessarily bear the upfront cost. Outcomes-based models, while directionally correct, "may still lack the design features needed to drive adoption," Pearson said.

Her proposed fix is a form of managed dynamic pricing: reimbursement rates that start high enough to reward early adopters and fund innovation, then step down over time as utilization scales, marginal costs fall, and real-world evidence builds. That requires "regular, longitudinal review," she said, a level of ongoing oversight payers have not historically applied to fee schedules. It is a harder administrative lift than a static rate card, but Pearson argues the dynamic nature of the technology leaves payers little choice.

Payment redesign alone will not settle the matter, either. Pearson frames it as one piece of a larger puzzle that includes safety and efficacy evaluation, licensure and liability frameworks, and provider workflow changes needed to build clinical confidence. Get the payment model right without solving liability, and adoption stalls anyway. Get liability sorted without fixing payment, and hospitals have no financial reason to deploy the tools at scale.

The stakes, in Pearson's view, run through both directions of risk. "If autonomous AI tools simply get layered into fee-for-service, we risk exploding healthcare spending," she said. The alternative: force CMS and commercial payers to build models that demand technology "both improves health and lowers spending," which she argues could simultaneously expand access, improve affordability and lift outcomes.

Pearson points to a cautionary precedent. Electronic health records were the last major technology wave to sweep through healthcare, and she calls the payment and policy response to that wave "a missed opportunity." Instead of simplification and coordination, the industry got fragmented data silos that persist today, years after adoption. "Once a technology is entrenched, tearing it out is a lot harder," she said. That history is the urgency underpinning PHTI's call to act now, before autonomous AI billing patterns calcify into the next set of hard-to-reverse defaults.

The Bottom Line

The report does not oppose clinical AI adoption. It argues that payment design, not clinical capability, will be the deciding factor in whether AI lowers system-wide costs or inflates them. For healthcare investors, the signal is clear: watch CMS and commercial payer rulemaking on AI reimbursement as closely as any clinical trial data. Companies building autonomous AI tools without a credible outcomes-based payment pathway may find strong clinical validation but a much narrower route to revenue than fee-for-service assumptions would suggest.