5,944 denials: Medicare's AI pilot exposed

Two vendors. Ninety days. 5,944 denied requests. That is the number Medicare's own contractors reported to CMS in the first quarter of the Wasteful and Inappropriate Service Reduction model, the federal government's first attempt at AI-assisted prior authorization in Original Medicare, and it only became public because the Electronic Frontier Foundation sued for the records ("New Records Reveal Problems with Medicare's AI Prior Authorization Experiment," EFF, September 8, 2026, corrected September 15, 2026). If you have ever wondered what "AI reviewing prior authorization requests faster" looks like once you can actually see the paperwork, this is it.

What the FOIA records actually show

WISeR launched in January 2026 across six states, New Jersey, Ohio, Oklahoma, Texas, Arizona, and Washington, requiring prior authorization for 13 defined procedure categories in fee-for-service Medicare, including arthroscopic debridement for the osteoarthritic knee, epidural steroid injections for pain management, and cervical fusion (CMS, WISeR Model overview, 2026). Private companies under contract to CMS use AI to help process the requests, though CMS states a qualified human clinician reviews every denial. EFF filed a Freedom of Information Act lawsuit in March 2026 after CMS did not respond to a records request, and on September 8, 2026, EFF released roughly 1,000 pages of what it obtained: vendor contracts, internal CMS status reports, and provider complaint logs.

The headline figure is specific enough to check: two of WISeR's vendors denied 5,944 prior authorization requests between January and the end of March 2026, drawn from a March 30, 2026 CMS status report. One vendor, Virtix, denied more requests than it approved in that window and was required to file a corrective action plan. CMS also publicly commits to a 72-hour turnaround on decisions. The released status reports show that promise was not consistently kept: one request sat unanswered for 83 days.

So what for you: treat this as a paper trail, not a hypothetical AI-governance debate. It shows what happens when an AI-assisted approval system is built without enough runway, and it is the same category of system, faster automated review of clinical requests, that vendors are now pitching to independent practices under a friendlier name.

The financial incentive built into the model

The records confirm something providers have suspected since WISeR launched: the payment structure rewards denial. WISeR vendors are paid for requests they deny, but not for denials that are later overturned on appeal. CMS says it protects against inappropriate denials by tying vendor payment rates to quality scores that reflect timeliness and accuracy. But the WISeR Data Reporting Guide, also released through EFF's lawsuit, shows a low quality score only reduces a vendor's payment by 5 to 10%. That is a modest penalty against a revenue model built on saying no.

The rollout itself was rushed. One vendor, Innovaccer, told CMS roughly a month before launch that it would go live with software that lacked full functionality and had not been fully tested, and that it planned to auto-affirm requests until it could finish building the system, writing that "given CMS's decision not to delay the model start date, auto-affirming is the only path available." Provider feedback collected in the same records describes the human cost on the other side of the delays: patients waiting weeks for a decision on a kyphoplasty or an epidural injection they were already scheduled for, procedures canceled mid-wait, and, in the words of one provider quoted in the released feedback survey, patients calling offices "crying in pain because their procedures are being delayed."

So what for you: an AI system with a financial incentive to deny, a weak penalty for getting it wrong, and a rushed launch is not an AI failure story. It is a procurement and incentive-design failure that happened to use AI, and that distinction matters for anyone evaluating a vendor's AI claims on their own admin side.

Why this reaches physical therapy anyway

Physical therapist services are not among WISeR's 13 covered categories, and this site covered that distinction in July, when the concern was whether PT billing itself would come under WISeR review. It will not, directly. But two of the 13 categories that WISeR does cover, arthroscopic debridement for the osteoarthritic knee and epidural steroid injections, are common referral points on either side of a physical therapy plan of care. WISeR's own medical necessity review checks for documented prior conservative treatment, typically physical therapy, before approving those procedures (Noridian Medicare, WISeR Model provider guidance, 2026).

That means a referring orthopedist or pain specialist whose WISeR request gets stuck for weeks, or denied outright by a vendor with every incentive to deny it, is a referral that does not reach your waiting room on schedule, if it reaches it at all. The provider feedback in EFF's released records describes exactly this pattern: surgeries and procedures canceled while a WISeR decision sat unresolved. A practice that relies on a steady flow of post-procedure or pre-procedure referrals from orthopedic and pain management practices in New Jersey, Ohio, Oklahoma, Texas, Arizona, or Washington has a direct, if indirect, stake in how reliably those referring practices can get patients through WISeR's front door.

So what for you: if your referral pipeline runs through any of the six pilot states, treat clean, complete conservative-treatment documentation as something that now serves two purposes, your own record and a referring specialist's WISeR file, and worth flagging to referral sources who may not realize the connection yet.

The distinction worth holding onto

It would be easy to read this story as evidence that AI does not belong near clinical or administrative decisions. That is the wrong conclusion, and it is worth being precise about why. WISeR is AI deployed by a payer, against a provider and patient, with a financial reward for the AI's decision going one particular way, under a rushed timeline that the vendor itself flagged as too fast. Every one of those design choices, not the presence of AI itself, produced the denial rate, the delays, and the patient harm documented in EFF's records.

An independent practice adopting AI for its own scheduling, documentation, or billing workflow is making a different set of choices: the AI works for the practice and the patient, not against them, there is no incentive to produce a worse outcome faster, and a badly performing tool can be switched off the same week it is identified, not left running for months while a lawsuit works through the courts. The WISeR story is a useful stress test for that distinction. Before adopting any AI tool that touches a clinical or financial decision, on either the payer or the practice side, ask who is paid when the tool says no, how quickly a wrong decision gets corrected, and what the human review step in the workflow actually consists of. WISeR's own records show "a qualified human clinician reviews every denial" was true on paper and still produced an 83-day wait and thousands of denials in a single quarter.

So what for you: the recommendation is not caution about AI generally, it is specificity about incentives. Ask what a tool is optimized to produce before asking what it can do.

CMS has said it may expand WISeR to more urgent categories in future years, including air ambulance transport and cancer treatment, according to a June 2025 planning document also released through EFF's lawsuit. Whatever happens next with the pilot, the first quarter's numbers are now a matter of public record, not vendor marketing, which is a rare enough thing in AI healthcare coverage to be worth sitting with.

Practices weighing their own AI adoption, on the administrative side where the incentives point toward the patient rather than away from them, can use our AI Opportunity and Growth Assessment to map where AI genuinely reduces admin load without the tradeoffs WISeR illustrates. You can also book a free 20-minute call to talk through what that would look like for your practice.

Related: WISeR's 13 categories and what they mean for PT referrals  ·  Prior authorization's true cost for PTs  ·  8.2% by 2027: the physical therapist shortfall

Core figures (5,944 denials across two vendors in the first three months, the 83-day unanswered request, the 5-10% quality-score payment penalty, and the Virtix corrective action plan) are from the Electronic Frontier Foundation, "New Records Reveal Problems with Medicare's AI Prior Authorization Experiment," published September 8, 2026 and corrected September 15, 2026, drawing on approximately 1,000 pages of CMS records obtained through EFF v. CMS, a Freedom of Information Act lawsuit filed March 2026. WISeR's 13 covered procedure categories, the six pilot states, and the program's run through December 31, 2031 are from CMS's own WISeR Model overview. The conservative-treatment documentation requirement for WISeR-covered knee and pain-management procedures is from Noridian Medicare's WISeR Model provider guidance, previously verified on this site. The provider quote is drawn from feedback survey responses released in the same FOIA production. This article is for informational purposes only and does not constitute clinical, legal, or financial advice.