CMS's AI prior authorization pilot: what it means for providers and payers

Prior authorization has been one of the most contentious friction points in U.S. healthcare for years. The administrative burden it imposes on providers, the delays it introduces into patient care, and the growing evidence that commercial insurer AI-driven denial algorithms are generating inappropriate denials have combined into a political situation in which CMS felt pressure to demonstrate that AI in prior authorization could be used differently — to reduce friction rather than optimize denials.

The agency’s response is a six-state prior authorization pilot that is now live, designed around a two-track triage model that represents a substantively different philosophy than the approach commercial insurers have been taking. Understanding the design of the pilot, its implications for providers in participating states, and where this policy is likely to go is essential for health system and medical group leadership.

How the two-track triage model works

The CMS pilot operates on a triage principle: not all prior authorization requests require the same level of review, and routing them appropriately can reduce delays for straightforward cases while preserving appropriate scrutiny for genuinely complex ones.

Track one is an automated approval pathway. Prior authorization requests that meet defined clinical criteria — high-confidence algorithmic determinations based on diagnostic coding, clinical documentation, and historical approval patterns — are approved automatically, typically within minutes of submission. The algorithm in this track is calibrated for specificity: a request only lands on track one if the model is highly confident it would have been approved by a human reviewer anyway. Borderline cases do not default to automatic approval; they default to track two.

Track two is a human-assisted review pathway. Requests that do not meet track-one criteria are routed to a CMS contractor reviewer, but with AI-generated structured summaries of the clinical documentation and a flagging system that highlights the specific criteria the request most closely resembles. The AI in track two is not making a decision — it is organizing information to help a human reviewer make a faster, better-informed decision.

This design is deliberately conservative. CMS made a deliberate choice not to use AI for automated denials in this pilot, in contrast to commercial insurer algorithms that have drawn congressional scrutiny and litigation over automated denial rates. The pilot’s AI handles only automatic approvals and human-review triage — denials always go through a human reviewer.

Why this differs from commercial insurer algorithms

The commercial insurer AI prior authorization controversy has centered on algorithms that generate high volumes of automated denials based on statistical patterns, with human review that may be perfunctory or missing for individual cases. The public evidence for this includes CMS’s own inspector general report on Medicare Advantage denial rates, litigation against major insurers, and whistleblower accounts from clinical reviewers describing algorithms that left little discretion.

CMS’s pilot design reflects a deliberate counter-positioning. By limiting AI to automatic approvals and human-triage support — not automatic denials — the agency is modeling an approach where AI accelerates care that would have been approved anyway, rather than creating new barriers to care that require appeals to overcome. This is both a substantive clinical care policy and a political signal about how Medicare should use AI differently than Medicare Advantage plans have.

Whether the design is too conservative to demonstrate meaningful administrative savings is a fair question. The pilot’s ROI will depend on how much of the prior authorization volume is captureable on track one, and early data will be watched closely by both health system advocates who want the program expanded and commercial insurers looking to legitimate their own approaches by pointing to the government doing something similar.

What providers in the pilot states should monitor

The six states participating in the pilot are operating within Medicare fee-for-service, not Medicare Advantage. Providers in these states should be tracking several things.

First, track-one approval rates by procedure category. If certain high-volume, guideline-concordant procedures are being consistently approved on track one, that is useful information for workflow planning — those requests require minimal human time to generate and approve, and the documentation burden can be streamlined accordingly.

Second, track-two review cycle times compared to pre-pilot baselines. The case for the pilot’s track-two design rests on the claim that AI-organized documentation summaries allow human reviewers to make decisions faster. If cycle times are not improving on track two, that is an early signal that the AI summary quality is insufficient or reviewer training is inadequate.

Third, denial rates and appeal outcomes on track-two requests. The pilot’s safeguard is that denials always involve human review, but the quality of that human review can still be assessed through appeal success rates. If appeals are overturning a high proportion of track-two denials, the review process has a calibration problem.

The policy trajectory toward national scaling

CMS has been explicit that the six-state pilot is a proof-of-concept for national expansion, and the trajectory of federal prior authorization policy points in a clear direction. The Improving Seniors’ Timely Access to Care Act, which Congress passed in 2022 and which CMS has been implementing, established electronic prior authorization requirements for Medicare Advantage. The AI pilot extends that logic into fee-for-service Medicare.

If the pilot produces strong data on administrative cost reduction and maintained or improved clinical appropriateness rates, the case for national scaling will be substantial. The political economy also favors expansion — prior authorization reform has unusual bipartisan support, and demonstrating that AI can reduce friction without increasing inappropriate denials gives CMS a policy win that crosses party lines.

For providers, the long-term implication is that AI-mediated prior authorization is coming to Medicare at scale. The organizations best positioned for that transition are those that have already invested in structured clinical documentation and EHR-integrated prior authorization workflows — because the track-one pathway’s value to providers scales directly with the quality and consistency of the documentation feeding into it.