Medicare Put an Algorithm in Charge of Saying No. Its Vendors Are Paid a Share of What It Refuses.

In Texas, in the first months of this year, roughly 62% of prior authorisation requests submitted under Medicare’s new model were approved on the first pass. After a physician reviewed the ones that had been refused, the approval rate rose to about 84%.

Read those two numbers together. Something told twenty-two out of every hundred patients that the care their doctor had ordered would not be paid for — and then a human being looked at those same files and said yes. Not a policy change. Not new evidence. A second opinion, on a decision the first reviewer should not have got wrong.

The first reviewer was software.

Every official description of this programme contains the same reassurance: it does not change Medicare coverage or payment policy. That is true, and it is the most revealing sentence in the entire story. The list of what Medicare covers is untouched. What changed is whether any particular person receives it.

Coverage did not change. Access did. Those have never been the same thing.

On 1 January 2026 — in practice, a fortnight later, after the portals failed — the Centers for Medicare and Medicaid Services launched the Wasteful and Inappropriate Service Reduction model, in Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. It runs for six years, to the end of 2031.

The mechanism is straightforward. For a defined list of services — skin and tissue substitutes, electrical nerve stimulators, knee arthroscopy, among others — providers must now seek authorisation before treating. Private contractors, using artificial intelligence and machine learning, process the request and issue an affirmation or a non-affirmation. A licensed clinician must sign off before any refusal is issued, and CMS is unambiguous that algorithms hold no final authority. The administrator, Mehmet Oz, has said the model will help root out waste in original Medicare.

The significance is not the technology. It is the venue. Traditional Medicare has never routinely required prior authorisation — that is the machinery of private insurance and Medicare Advantage, and the reason many people choose traditional Medicare in the first place. A gate that existed in the commercial market has been installed in the public one, and automated on arrival.

Congress noticed. In May the Government Accountability Office determined the model was subject to the Congressional Review Act, opening a window to force a vote. Resolutions followed in both chambers. The House Appropriations Committee moved on 9 June to block its funding; thirty-one House Democrats wrote on 22 June demanding implementation data. On 16 July the Senate voted 46 to 50, along party lines, against advancing the repeal.

The model survives. It has five and a half years left to run, and a stated path to expansion.

The keeper is paid by the refusal. The contractors operating the model are compensated, in part, on a share of what CMS terms averted expenditures — the money not spent because a request was not approved. Their revenue is a function of their denials.

No accusation of bad faith is required here, and none is being made. The point is structural and it is simple: an incentive shapes a system whether or not anyone intends it to. We have built a gate, pointed it at elderly patients, and arranged for its keeper to be paid by the refusal. Whatever that machine learns, it will learn it in that direction.

It is wrong at a rate we can measure, and only sometimes catch. The gap between 62% and 84% is not a rounding error; it is a fifth of the affected population being told no before being told yes. For comparison, prior authorisation approval rates in Medicare Advantage often exceed 90%. And a Senate committee report in 2024 found that AI tools deployed in prior authorisation settings had been associated with denial rates sixteen times higher than decisions made without them.

That gap is only closed for the patients whose refusal gets challenged. In Medicare Advantage in 2023, just 11.7% of denials were ever appealed — and around 82% of those appeals succeeded. Read that pairing slowly. The overwhelming majority of denials are never contested, and the overwhelming majority of contested denials turn out to have been wrong. The system’s error rate is knowable only in the sliver of cases where someone had the stamina to fight.

Nobody can see the reasoning. In March, a digital rights organisation filed a Freedom of Information suit against CMS seeking to learn which technologies and AI tools the contractors are actually using. That a lawsuit was necessary is the finding. The logic now standing between an American pensioner and a procedure their physician ordered is a trade secret.

This publication has argued from its first issue that approval is not access, and access is not adoption — three gates, each losing patients the last one had cleared. What has appeared in Medicare this year is a fourth, and it is different in kind from the three before it.

The first three gates are decisions about a product. A regulator approves a medicine. A payer decides to fund it. A system adopts it. Each is made once, for a population, by a named institution, on a published record, against criteria that can be read and contested. You may loathe the decision, but you can find it, argue with it, and cite it.

The fourth gate is a decision about you. It is made in seconds, on your file, by a system whose reasoning is proprietary, operated by a company whose revenue rises when the answer is no. And because coverage policy is officially unchanged, there is no decision at the level of principle to appeal at all. The service remains covered. It was simply not covered for you, this time, in your case.

There is no policy to appeal, because no policy was made.

That is what makes this the most efficient form of rationing yet devised. It produces no announcement, no committee minute, no published criterion, no negative recommendation anyone can point at. It generates only outcomes — millions of individually plausible ones — and any pattern in them is visible solely to whoever holds the data. A health system can now become dramatically less generous without a single line of its stated generosity changing.

The honest counter-argument deserves to be made, because it is real. Low-value care exists, and it is expensive: CMS put the services in scope at somewhere between $1.9 billion and $5.8 billion of low-value spending in 2022 alone. Prior authorisation is already routine across private insurance. Automation genuinely can speed approvals as well as refusals, and the model includes a gold-carding mechanism to exempt clinicians with clean approval histories — an intelligent feature. CMS delayed two of the more sensitive services in April, which suggests some responsiveness. The objection is not to using software. It is to using software that is paid by the refusal, unexaminable in its logic, and installed in the one place in American health care that had been left alone.

Three things follow.

First, for anyone working in access, the last mile is now a live risk and nobody is modelling it. A product can be approved, reimbursed, guideline-recommended and formulary-listed, and still be refused at the point of use by an automated utilisation-management decision. Every dossier we build is addressed to a committee. Not one is addressed to the algorithm — and the algorithm is increasingly the thing standing between the patient and the prescription.

Second, Europe should stop treating this as an American curiosity. The NHS is deploying an AI triage tool inside a £10 billion technology programme, reaching more than 200,000 patients within twelve months and every app user by April 2028, with an early trial cutting phone queues by 29%. The intent is entirely benign and the tool may well be excellent. But the architecture is the same architecture: an algorithm placed between a patient and the system, deciding what happens next. The question worth asking early — before it is embedded rather than after — is who is paid what, and whether anyone outside the building can inspect the logic.

Third, and least comfortably: we wrote last month that artificial intelligence was being aimed at drug discovery, the one gate the industry already wins, and that no model was pointed at the gate where patients actually wait. That was wrong. One is.

It is simply not working for them.

Approval was never access. Reimbursement was never access. And now, quietly, in six American states, coverage is not access either — because between the benefit a person is entitled to and the treatment they actually receive, someone has installed a machine that is paid a share of everything it refuses to let through.

Share

Beyond Approval publishes weekly strategic intelligence on market access, HTA, pricing, and regulatory shifts that reshape how medicines reach patients.