CMS's WISeR Medicare AI prior authorization model pays vendors a percentage of care they deny — and a Corrective Action Plan vendor denied more requests than it approved. That's a governance structure failure, not an AI failure.
Table of Contents #
The Electronic Frontier Foundation spent months suing CMS for records on its WISeR (Wasteful and Inappropriate Services Reduction) model. The roughly 1,000 pages they got back, released in September 2026, tell a story that anyone who has shipped AI systems inside regulated institutions will recognize immediately: governance controls designed on paper, missing in practice, and a payment structure that rewards exactly the wrong behavior.
WISeR is a voluntary CMS Innovation Center model that launched January 1, 2026, in six states — New Jersey, Ohio, Oklahoma, Texas, Arizona, and Washington. The model outsources Medicare prior authorization for thirteen medical services to six private technology vendors who use AI to evaluate requests. The services covered include steroid injections, nerve stimulators, arthroscopic knee surgery, sleep apnea treatments, and incontinence devices. The model runs through December 31, 2031, and CMS has already signaled potential expansion to air ambulance transport, cancer treatment, and MRI scans.
The pitch was straightforward: AI-assisted prior authorization would reduce administrative burden, catch inappropriate care, and save Medicare money. The governance story embedded in those FOIA documents is something else entirely.
The Governance Gap #
The foundational problem with WISeR is not that it uses AI. It’s how vendors get paid.
Under WISeR’s payment structure, vendors receive compensation tied to a percentage of expenditures they identify as wasteful and successfully prevent — meaning denials that are not subsequently overturned on appeal. Vendors are not paid for the requests they approve. They are paid when they deny, and only when that denial holds. CMS added quality score adjustments to guard against inappropriate denials, but the WISeR Data Reporting Guide shows that low quality scores reduce vendor payments by only 5 to 10 percent. That is not a meaningful deterrent. A vendor that systematically denies borderline requests, absorbs the modest quality score penalty, and profits on the net reduction in payouts has a rational business case for over-denial that the current payment structure does not correct.
Virtix Health, one of the six WISeR vendors, makes this visible in the data. In the program’s first three months, Virtix denied more prior authorization requests than it approved — a total of more than 5,944 denials across the two vendors for which EFF obtained early data. CMS required Virtix to submit a Corrective Action Plan. But the Corrective Action Plan requirement is a governance response after the harm has occurred, not a control that prevents it. The question the EFF records do not answer — because CMS has not released the data — is what the denial accuracy rate actually was. How many of those denials were clinically appropriate? What share were overturned on appeal? Without that data, the Corrective Action Plan is a governance gesture.
The human review requirement does not close this gap the way CMS describes it. CMS’s position is that a qualified clinician must review every denial recommendation before it issues. But research consistently shows that AI recommendation systems exercise strong anchoring effects on human reviewers — the human tends to confirm, not override, particularly under time pressure and high volume. The EFF records offer limited evidence that WISeR’s clinical reviewers have either the authority or the information to systematically override vendor AI recommendations. A signature on an AI-generated denial is not meaningful human oversight. It is liability transfer.
What the Examiner Will Find #
Congressional scrutiny is already escalating. The Senate voted on — and rejected, narrowly — an effort to halt the WISeR pilot. That vote signals the governance gaps are visible at the oversight level, and that this program will face sustained political and regulatory examination through its 2031 run date.
What oversight bodies will be looking for, and what is currently missing: First, denial accuracy rates broken out by vendor, service type, and state — and whether those rates diverge significantly from clinical expectation or historical approval baselines. The EFF records show denial totals but not accuracy. Second, appeal overturn rates by vendor. Under WISeR’s payment structure, the appeal overturn rate is effectively the only financial check on over-denial — and it’s a weak one at 5–10% payment adjustment. Third, bias and disparate impact analysis. Any prior authorization AI system processing Medicare claims is touching a population with protected characteristics across age, disability status, and potentially race and geography. There are no public records of bias testing for any WISeR vendor.
Regulators examining health plans that participate in similar AI-assisted prior authorization programs will look for operational evidence — not written policies, but audit logs, decision records, override rates, and vendor monitoring cadences. Plans that point to a vendor contract as their governance control will get an MRA. The vendor contract establishes the relationship; it does not demonstrate that the clinical review process is functioning, that the AI system is performing within tolerances, or that patients are not being systematically harmed by incorrect denials.
State insurance regulators are entering this space explicitly. California SB 306 and Oregon HB 3134 both impose prior authorization AI requirements on health plans. The EFF records suggest CMS’s federal model was not built to meet the bar those state laws establish. That creates a compliance gap for plans operating across state lines — a different regulatory standard applies to the same AI decision depending on which state the patient is in.
Model Risk Reality #
From a model risk perspective, the WISeR program exhibits the hallmarks of a system that was deployed without completing the validation requirements it would face in any model risk framework aligned to SR 11-7 or its successors. The readiness failure at Innovaccer — which serves Ohio — is the clearest evidence. About a month before the January 2026 launch, Innovaccer informed CMS that it would go live without full portal functionality or end-to-end testing. The vendor cited changing requirements and unclear governance. Its response was to auto-affirm all incoming requests until the system was fully operational. CMS launched on schedule anyway. Auto-affirmation avoids harm in the near term, but it means the vendor was running an untested system in a live production environment on Medicare beneficiaries. The testing that should have happened in pre-production happened in production, on real patients, with real care decisions at stake.
A model risk officer reviewing this would flag the absence of what any serious validation program requires for a model with this risk tier: independent challenge of the AI recommendation logic, performance benchmarks against a clinical standard, pre-deployment bias testing across the population it serves, and a defined operating envelope with escalation protocols when the model operates outside expected parameters. The WISeR records, as released by EFF, contain none of this. That is not necessarily because CMS did not ask for it — EFF specifically requested records on vendor accuracy, bias, and hallucination testing, plus audits and monitoring, and CMS is still producing documents. The absence of public records on these items is itself a governance signal.
The 83-day response delay at Genzeon — against a 72-hour standard that is supposed to prevent exactly this kind of care interruption — is the operational manifestation of a model running without an adequate monitoring and escalation framework. In a properly governed clinical AI deployment, an SLA breach at day four triggers an alert. A breach at day seven triggers escalation. An 83-day breach means the monitoring system either did not exist or was not functioning. Provider feedback obtained by EFF describes the human cost: patients calling in pain because procedures were delayed, surgeries cancelled, patients crying at bedside. These are not edge cases. They are what a clinical AI deployment that fails its SLA agreement at scale looks like.
The SuperML Take #
The WISeR story is being covered as an AI rollout problem. It is not. It is a governance structure problem that AI amplified.
Every element of what went wrong here — perverse payment incentives, rubber-stamp human review, absent pre-deployment validation, no bias testing, an 83-day SLA breach with no escalation — would have created harm at volume in a manual prior authorization system. AI made it faster, more opaque, and harder to audit. The vendors can now process thousands of cases per day, which means that a systematic bias in the denial logic, or a systematic over-denial incentivized by the payment structure, affects far more patients far more quickly than a manual reviewer making the same error ever could. That is what “AI at scale in a regulated industry” actually means: the failure modes scale with the throughput.
What a rigorous governance framework for a program like WISeR would actually require is not complicated, but it requires the willingness to build it before launch, not after the FOIA records come out. Vendor payment must be decoupled from denial outcomes. Any payment structure where the vendor’s revenue goes up when it denies care has to be treated as a conflict of interest, not a quality incentive. The clinical review requirement must be operationalized to mean something: reviewers need the full clinical record, the AI’s recommendation with its confidence score and the features that drove it, a documented mechanism to override, and override rate reporting as a governance metric — not a signature on a queue. Bias and accuracy testing must happen in pre-production, against the actual population the system will serve, with the results disclosed before the system goes live. And SLA compliance has to be monitored in real time, with automatic escalation, not discovered through FOIA requests eight months later.
The Senate’s failed attempt to halt WISeR signals that the political consensus on AI prior authorization is not stable. The program runs through 2031, and if the denial accuracy data, when released, shows systematic over-denial correlated with geography or demographics, the political and legal exposure for both CMS and the participating vendors will be significant. Health plans watching this model should be treating WISeR as a preview of the examiner questions they will face on their own prior authorization AI programs — and using it to test whether their own governance infrastructure would survive the same FOIA request. If the answer is no, now is a better time to fix it than when the records are released.
Sources #
- Electronic Frontier Foundation. “New Records Reveal Problems with Medicare’s AI Prior Authorization Experiment.” EFF Deeplinks, September 8, 2026. https://www.eff.org/deeplinks/2026/09/new-records-reveal-problems-medicares-ai-prior-authorization-experiment
- Cohen, Lena. “7 Potential Problems with Medicare’s AI Prior Authorization Experiment.” MedCity News, November 2026. https://medcitynews.com/2026/11/7-potential-problems-with-medicares-ai-prior-authorization-experiment
- “One Medicare AI Prior Authorization Request Went Unanswered for 83 Days as Vendors Often Missed the 72-Hour Standard.” Medical Daily. https://www.medicaldaily.com/medicare-wiser-ai-prior-authorization-records-delays-478712
- “Medicare AI program hit WA with ‘staggering’ denial rate, records show.” Rep. DelBene Congressional Office. https://delbene.house.gov/news/documentsingle.aspx?DocumentID=4475
- “Senate rejects effort to halt Medicare AI prior authorization pilot.” Healthcare Finance News. https://www.healthcarefinancenews.com/news/senate-rejects-effort-halt-medicare-ai-prior-authorization-pilot
- FTI Consulting. “AI in Prior Authorization: Advanced Oversight Required.” https://www.fticonsulting.com/insights/articles/ai-use-prior-authorization-decisions-requires-advanced-oversight
- “CMS launched Medicare WISeR AI pilot despite vendor warning.” AI Weekly. https://aiweekly.co/alerts/cms-launched-medicare-wiser-ai-pilot-despite-vendor-warning
Enterprise AI Architecture
Want more enterprise AI architecture breakdowns? #
Subscribe to SuperML.