EFF Records Show CMS Warned WISeR Vendors Over Slow, Inaccurate AI Medicare Denials

On October 27, 2026, the Electronic Frontier Foundation published a second batch of government records about WISeR, the Medicare pilot that lets private vendors use AI-assisted tools to screen prior authorization requests. The 800-plus pages show that in the program's first three months, CMS sent requests for information or corrective action plans to at least four of the six vendors, citing late determinations, denials that looked inaccurate, poor customer service and, in one case, a provider portal that did not work. EFF says CMS did not publicly disclose the extent of those problems.

This is the part of the story that is documented. What the records do not show is how many patients were harmed, how often the AI's recommendation was overturned by a human, or what the redacted safeguards actually are. This post separates those.

TL;DR: what WISeR is and what the records add #

How WISeR works #

According to CMS, WISeR covers a pre-selected set of items and services, including skin and tissue substitutes, implantation of electrical nerve stimulators and knee arthroscopy for knee osteoarthritis. It excludes inpatient-only services, emergency services and services that would pose a substantial risk to patients if delayed. Vendors, called model participants, are paid a percentage of the expenditures associated with averted wasteful or inappropriate care, adjusted by performance on process measures such as provider experience.

CMS also says recommendations for non-payment must be made by appropriately licensed clinicians applying standardized, evidence-based procedures. That is the "human in the loop" claim at the center of the debate.

EFF sued for these records in March 2026. Its first release, in September, described delays, denials of care and operational chaos. The October release goes deeper into the vendor relationship.

What the new documents say #

Per EFF's summary of the records:

• Virtix. An RFI dated March 17, 2026 cited slow determination turnaround, denials that appeared inaccurate or inconsistent with criteria, and inadequate customer service.
• Zyter. An RFI dated February 27, 2026 noted reports that determination letters were not issued on time and that the company may lack adequate processes for provider and supplier concerns. EFF notes Zyter uses Palantir software infrastructure.
• Innovaccer. It received two corrective action plan requests and a "Performance Notice of Non-Compliance" on February 20, 2026 for failing to deliver a fully functional provider portal. Innovaccer said it had warned CMS months earlier it might not be ready and had submitted a proposed plan on January 16, 2026.
• A fourth vendor said it faced "technical and workflow stabilization challenges" because it joined about a month after the others, compressing its implementation timeline.

The records also include applications from 66 companies and task orders for the program's evaluation, implementation and monitoring. EFF says CMS launched WISeR six months after announcing it, over providers' calls to delay, and had itself acknowledged an "extremely condensed timeline."

Vendors described fixes. EFF's argument is that after-the-fact fixes cannot undo harms already suffered by patients who waited for care and providers who spent time fighting denials.

A short timeline #

Putting the dates together shows how quickly problems surfaced:

• January 1, 2026: WISeR goes live in six states.
• January 16, 2026: Innovaccer submits a proposed corrective plan, saying it had already warned CMS it might not be ready.
• February 20, 2026: CMS issues Innovaccer a notice of non-compliance over its provider portal.
• February 27, 2026: CMS sends Zyter a request for information about late determination letters.
• March 2026: EFF files its suit for records; CMS sends Virtix an RFI on March 17.
• September 2026: EFF's first release describes delays and denials.
• October 27, 2026: the second release, with 800-plus pages.

Every one of the dated vendor warnings falls inside roughly the first eleven weeks. That is the core of the "condensed timeline" criticism: the program was live in six states before several vendors could reliably send a letter or run a portal.

The incentive question #

The most contested design choice is payment. A vendor paid a share of the money it saves has a financial reason to deny. EFF says the government itself recognized that the payment structure gives vendors an incentive to deny as many claims as possible, and that one vendor's application, while touting human-in-the-loop workflows, continuous model monitoring and audit trails, denied more requests than it approved in the first few months.

CMS has not published vendor-level denial statistics that we could find, so how often each vendor denies requests remains unknown.

Is "a human reviews it" a real safeguard? #

CMS's FAQ says denials cannot be made solely by technology. EFF's response is that the label tells you little. It asks how many cases a reviewer handles, whether they can override the AI, and whether they see the patient's file. It cites reporting that UnitedHealth pressured clinical reviewers to accept AI recommendations and that a Cigna system let reviewers reject claims without opening patient files, averaging 1.2 seconds per case, and it points to research on over-reliance on automated recommendations.

This is the same failure pattern we see in other AI deployments: a human is nominally present, but the system is designed so that agreeing is the path of least resistance. We looked at the model-side version of that dynamic in sycophancy in chatbots and the consequences when a health-related assistant simply goes along in the GPT-4o seizure medication lawsuit.

What data goes in #

Vendor applications reference inputs such as provider behavior, patient demographics, utilization management datasets, provider and payer performance metrics and historical claims data. EFF notes that CMS guidance for Medicare Advantage plans warns against coverage decisions based on large datasets rather than the individual patient's history and records, and says more detail is needed to judge whether WISeR's inputs are appropriate. That detail is largely absent from what has been released.

What EFF is asking for #

EFF recommends that Congress impose greater legal limits on automated decision-making in Medicare, that government AI use come with transparency, oversight and legal limits, and that the public get more information on vendor safeguards, the substance of clinician review and the data used. It points to Colorado's AI Act as a model.

For a wider view of where US rules on automated decisions stand, see our guide to AI regulation across the EU AI Act and US policy. For another example of lawmakers pressing AI makers for answers, see the Hawley and Blumenthal letter on Google AI in classrooms.

What this means for people building decision systems #

If you build or buy AI that influences approvals, denials or eligibility, the WISeR record offers practical lessons.

1. Readiness is a safety property. The corrective action requests are mostly about operations: late letters, a missing portal, a vendor that joined late. In a high-stakes workflow, an outage or delay is itself a denial of service for the person waiting.
2. Define what human review means. State reviewer caseloads, time per case, whether the reviewer sees the full record, and how often they overrule the system. Publish those numbers.
3. Align the incentive. Pay based on accuracy and appeal outcomes, not on dollars averted.
4. Log and audit. Keep decision logs that can be sampled and tested for inconsistency, since "denials inconsistent with criteria" was the flagged problem.
5. Test beyond averages. Benchmarks likeOpenAI's MentalHealthBench show how hard it is to grade clinical behavior even with clinician rubrics; a pilot with real patients needs stronger measurement still.

If your systems use autonomous agents that touch sensitive records, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

What to watch #

Whether CMS publishes vendor-level denial and appeal-overturn rates; whether the redacted safeguard descriptions are released; whether Congress acts on automated-decision limits; and whether any of the four flagged vendors' corrective plans are shown to work. We will update this post if new records appear. This is separate from the Australian Medicare portal incident, which involves a different country's system.

Related reading #

Details reflect EFF's October 27, 2026 release and CMS's program page as of October 28, 2026; secondary figures are flagged as unverified.