Clinical AI's Payment Problem: Why Fee-for-Service Could Backfire

The core question in healthcare AI has shifted. It is no longer whether the technology can perform clinical tasks reliably. It is whether the payment system built around human labor can absorb a technology that scales without human constraints.

A report from the Peterson Health Technology Institute, an independent non-profit that evaluates digital health systems, makes a direct case: applying today's fee-for-service reimbursement rules to autonomous clinical AI risks driving healthcare spending up, not down. That finding deserves attention from every health system executive, payer, and investor tracking the AI adoption curve.

"We need to get payment right to expand access to high-quality care while lowering costs," said Caroline Pearson, PHTI's executive director. The stakes are structural, not incremental. Payment policy, she argues, will determine whether clinical AI becomes a genuine cost-saving tool or an expensive add-on layered atop an already bloated system.

The Structural Mismatch

Fee-for-service works because clinician time is finite. A physician can only see so many patients in a day, and reimbursement scales with that natural ceiling. Autonomous AI has no such ceiling.

"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, while AI "can generate far more clinical activity than any human workforce, without a proportional increase in clinician time." Tying payment to volume, in other words, could inflate spending at a scale fee-for-service was never designed to handle.

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 existing billing codes. Autonomous AI does not.

Pearson's hypertension example is instructive. Algorithms that adjust medication based on remote blood pressure readings, with minimal human oversight, can manage a protocolized condition effectively. But there is no clinician encounter to bill, no code that captures the work, and often no mechanism to pay the technology developer directly. The payment infrastructure simply has no slot for this kind of care.

This is not a minor gap in the coding manual. It is a signal that reimbursement architecture, built decades ago around physician time, cannot simply be stretched to fit a technology that operates on a different cost curve entirely.

Outcomes-based payment is PHTI's preferred alternative, but the report is careful not to oversell it. "Outcomes-based payment works best for chronic diseases with clear, measurable metrics," Pearson said, pointing to conditions where near-term results are easy to track. It "begins to fall apart in areas like primary care or prevention," where outcomes take years to materialize and attribution gets murky. No single model, she argues, will cover every use case.

There is also a practical adoption problem layered on top of the payment design question. Health systems face substantial upfront licensing and implementation costs before any AI tool delivers measurable value, and the eventual returns are often uncertain and distributed across multiple stakeholders, few of whom bore the initial expense.

Pearson's proposed fix is dynamic pricing. Rates should start high enough to stimulate early adoption, then adjust downward as utilization scales, marginal costs fall, and real-world evidence accumulates. That requires "regular, longitudinal review," a discipline current fee schedules rarely apply with any rigor. It is a more demanding administrative task than the periodic rate updates payers currently manage, but Pearson frames it as necessary given how fast AI capabilities and costs are likely to shift.

Payment design cannot be separated from the broader adoption ecosystem either. Safety validation, licensure, liability exposure, and clinician workflow redesign all factor into whether a reimbursement model actually drives uptake or sits unused. Get the incentive structure right but ignore liability concerns, and adoption stalls regardless.

Pearson frames the moment as a rare opportunity to rebuild the underlying economics of care delivery, not just bolt AI onto existing billing codes. "If autonomous AI tools simply get layered into fee-for-service, we risk exploding healthcare spending," she said. Her preferred alternative: CMS and commercial payers should design models that explicitly demand technology both improve health outcomes and lower spending, rather than rewarding activity alone.

The comparison Pearson draws to electronic health records is pointed, and it should worry anyone who lived through that transition. "One of the last great technology revolutions in healthcare was electronic health records, and it was a missed opportunity," she said. Instead of the simplification and coordination EHRs promised, the industry got "complexity and data silos that we're still working to break down." Her warning is explicit: once a technology becomes entrenched in clinical workflows and billing systems, correcting the incentives around it becomes exponentially harder. The window to design payment models correctly is now, before autonomous AI tools are wired into hospital operations at scale.

The Bottom Line

This report is a policy warning dressed as a payment analysis, and it deserves to be read that way. The risk PHTI identifies is not that clinical AI fails to deliver value. It is that poorly designed reimbursement mechanics allow AI to generate volume and revenue without a corresponding improvement in outcomes or cost efficiency, replicating the EHR experience at a larger scale.

For investors and health system leaders, the signal is clear: watch how CMS and commercial payers structure reimbursement for autonomous AI tools over the next 12 to 18 months. Fee-for-service extensions to AI-generated activity should be treated as a red flag for margin risk across the system, not a growth catalyst. Outcomes-linked, dynamically priced models, while harder to implement, represent the more sustainable path, and companies building AI tools with measurable, near-term outcome data will be better positioned to negotiate favorable payment terms than those relying on volume alone.