Medicare's AI pilot pays contractors a share of every denial, records show
The administration's Medicare experiment pays private companies when they refuse care — and its own actuaries wrote that down months before a Senate hearing where the man running the program said they don't.
Medicare's WISeR pilot pays its contractors a share of the money they save by denying care, and the administration's own actuaries said so in writing. The program — Wasteful and Inappropriate Service Reduction — launched in January in six states: New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. It requires prior authorization for roughly a dozen services, from epidural steroid injections for pain to cervical fusions, and it runs until 2031. Documents released by the Electronic Frontier Foundation this month show how the money works. CMS sets the payment rate at 25 percent of the regional benchmark cost of each request a contractor declines — "a novel payment approach where the model participants are compensated based on a share of averted expenditures" — with the effective rate landing lower after discounts that outside analyses put closer to 20 percent.
The CMS Office of the Actuary was blunter. Participants, it wrote in a memo, "will have an incentive to deny as many claims as possible."
That memo got tested on September 16, when Sen. Patty Murray asked Chris Klomp — CMS deputy administrator, who has been directing Medicare since last year and is now nominated for deputy secretary of Health and Human Services — whether the contractors make more money if they deny care. "Just 'yes' or 'no,'" she said. "My understanding is no," Klomp replied. Murray's follow-up went into the record: "Okay, but they do deny care. They're paid to deny care."
The rollout was as rough as the incentives suggest. One vendor, Innovaccer, asked CMS to delay the launch, then configured its system to auto-approve everything until it finished building the real one. Another, Virtix, denied more requests than it approved in the first quarter and was put on a corrective action plan after missing the program's 72-hour decision window; the company says that ended in August and its average turnaround is now about 1.18 days. CMS records contain rows where a request was logged as Medicare Part A when it was a Part B claim — the inpatient-versus-outpatient distinction that determines who pays.
The appeal data is the part that should worry CMS most. Only about 11 percent of Medicare Advantage denials get appealed, and roughly 80 percent of those appeals succeed. If anything like that ratio carries over, most of these denials are wrong and almost nobody is checking. Klomp told Murray the agency has "significant financial penalties" for inappropriate denials. The released documents show the penalty is a 10 percent cut to that 25 percent payment for a contractor scoring below 60 out of 100 — a failing grade by school standards — and EFF's corrected count is 5,944 denials across two vendors in three months.
This is the other half of a claims fight we covered earlier today — insurers blame hospitals' AI coding for $942 million in extra claims. Both sides of American healthcare now run AI against the other side's paperwork, and the party with the least visibility is the patient. The GAO ruled in May that the administration set WISeR up without following proper procedure, a House committee vote to force more documents out failed last week, and CMS plans to expand the program to oncology, advanced imaging, pacemakers and genetic tests. Murray says she is going to do everything she can to stop it.
Anthropic says Claude is now drafting the daily situation reports for the Ebola outbreak in the eastern Democratic Republic of the Congo, and that the job went from a full day to under an hour. The company's feature published this week describes a collaboration with CEPI, the WHO Regional Office for Africa and the DRC's national biomedical research institute on the Bundibugyo strain, which has no approved vaccine. A WHO AFRO data team built a Claude workflow that pulls case and lab counts out of each health zone's slide deck, checks them against the previous day's report and flags trend changes before summarizing districts' submissions. Anthropic also says the institute's lab — which sequenced essentially all the outbreak's genomes — is using its Claude Science workbench to assemble viral genomes, and that CEPI used Claude to build a dashboard tracking vaccine development.
The outbreak itself is well documented, the AI claims currently are not. WHO's external situation report for data through September 20 puts the toll at 7,733 confirmed cases and 3,732 deaths, a fatality rate of 48.3 percent, across 63 of the country's 167 health zones with Ituri province the epicenter. CEPI has redirected $100 million of its own funding toward Bundibugyo vaccine work, and its chief medical officer has called this outbreak potentially the worst the world has seen. But Claude's role rests on Anthropic's account alone — CEPI's own site and WHO's situation reports make no mention of it, and every other write-up traces back to the company's page. Read the partnership claims as a vendor's testimony until a health agency confirms them.
If an algorithm denies your care, who should have to prove it got it right — the contractor or the government? Tell us in the comments.
Sources: Ars Technica · Electronic Frontier Foundation · EFF records: WISeR participation agreements · Sen. Patty Murray, HELP hearing · STAT News · KUOW / NPR · Anthropic · WHO AFRO situation report 19 · WIRED (via DNYUZ)
