Medicare's AI Prior Authorization Pilot Pays Vendors To Deny Care. The Senate Just Let It Continue.

Last Thursday, the Senate voted 46 to 50, along party lines, to block a bill that would have shut down WISeR, the CMS pilot testing AI-driven prior authorization inside traditional Medicare. Republicans held the line. The Trump administration had sent lawmakers a handout defending the program days before the vote. WISeR keeps running in six states.

Most of the coverage since has framed this as an AI story: is the technology accurate enough, is it fair to patients, will it hallucinate a denial. Those are real questions. They are not the question that should worry you most if you have run claims operations for a living, which I have, for 25 years.

The question is who gets paid, and for what.

The incentive is the design flaw, not the model.

WISeR stands for Wasteful and Inappropriate Service Reduction. CMS contracts with outside technology vendors to review prior authorization requests for a defined set of services in original Medicare, something the program almost never allowed before this pilot. The part that should get more attention than it has: vendor compensation is tied to what CMS calls averted expenditures, the dollar value of the care the AI recommends against approving.

Read that again slowly. The company reviewing whether your mother's procedure gets approved does better financially the more procedures it recommends denying.

I have sat on the other side of contracts like this. Not this one, but the structure is familiar. When you build a financial incentive around an outcome instead of around accuracy, you do not need a biased algorithm to get biased results. You need a well-functioning algorithm sitting inside a badly designed contract. The model will learn, or be tuned, or simply be selected and kept in production, because it produces the outcome the contract rewards. Nobody has to instruct it to deny more claims. The incentive does that work on its own.

This is not a new pattern in claims operations. It predates AI by decades. What is new is the speed and the scale at which a contingency-style incentive can now be executed, at machine speed, across a Medicare population that has almost no history of prior authorization at all.

The data backs up what the incentive predicts.

The American Medical Association's most recent physician survey found that six in ten doctors expect AI to increase prior authorization denials, not reduce administrative friction the way vendors pitch it. That is not a fringe number. That is a majority of the people who actually submit these requests telling you what they expect the incentive to produce.

Meanwhile, AI is not arriving into claims operations. It is already there. A National Association of Insurance Commissioners survey of 93 insurers across 16 states found 84 percent already use AI or machine learning somewhere in utilization management, disease management, or prior authorization. The debate over whether AI belongs in this process is several years behind the actual adoption curve. The debate that matters now is what governs it once it is in.

Some states already answered this. Washington might override them.

A number of states did not wait for Congress. Illinois requires that only a licensed clinical peer, not an algorithm, can issue a denial based on medical necessity. Alabama requires that any AI used in prior authorization weigh the enrollee's actual clinical history, not just pattern-match against prior claims. Texas gives its insurance commissioner standing audit rights over the automated systems insurers use. Utah requires disclosure to regulators, providers, and enrollees when AI is doing the reviewing. As of this spring, at least 25 states had issued guidance built on this model.

At the same time, the current administration's AI policy framework pushes for federal preemption of exactly this kind of state law, on the argument that a patchwork of state rules slows AI deployment down. Whether that preemption survives the legislative process is genuinely unresolved. But the practical effect, if it goes through as proposed, is that a state like Illinois could lose the authority to require a human being to sign a denial, right as the federal government is piloting a program engineered to produce more of them.

What actually needs to be true before AI touches a denial.

I do not think AI has no place in prior authorization. Claims volume in this country is enormous, a meaningful share of it is genuinely routine, and automation done right speeds up the majority of requests that should be approved without friction, freeing clinical staff to spend real time on the requests that are actually complicated. That is a legitimate use of the technology and I have built toward exactly that kind of automation myself.

What has to be true first, every time, no exceptions:

A human being with clinical authority signs the final denial, not a dashboard. CMS's own 2023 Medicare Advantage rule already requires this for medical necessity determinations. WISeR is testing AI in a part of Medicare that never had this guardrail built in from the start, which is precisely backwards.

The vendor's incentive has to be decoupled from the denial rate. Pay for accuracy, pay for turnaround time, pay for reduced appeal overturn rates. Do not pay a percentage of what got denied. That single contract term explains more of the outcome than the model architecture ever will.

The algorithm and its outcomes have to be open to audit by someone other than the vendor grading its own homework. Texas has this right. CMS, as of this month, does not, at least not in a way regulators or the public can currently see without litigation.

CMS just stood up a new Office of Health Technology Products to oversee exactly this kind of federal AI adoption. Whether that office has the authority and the independence to actually govern WISeR, instead of defend it, is the thing worth watching over the next two quarters, not whether the AI itself is technically sound.

The bigger point.

Every AI adoption conversation I have run in payer operations comes back to the same question: what is the incentive structure the model is sitting inside. Get that wrong, and the most accurate model in the world will still produce outcomes nobody intended and nobody will fully own. Get it right, and AI becomes what it should be in this industry, a way to clear the routine work faster so clinical judgment gets spent where it actually matters.

WISeR is not a referendum on whether AI works in healthcare. It is a live case study in what happens when nobody asks who the incentive was built to reward before the pilot goes live.

Gigi Gierbolini-Carrabbia is the founder of GGenesis Strategic Solutions LLC. She spent 25 years running enterprise payer operations, including claims stabilization through a full core payer platform conversion, before building GGenesis, CaliberSuite™, and Talvori™. Certified in AI Strategies for Business Transformation: Generative and Agentic Intelligence. Based in Broussard, Louisiana. Serves clients nationally.

Sources: STAT News, "GOP senators oppose bill to end AI prior authorizations in Medicare," July 16, 2026. KFF, "Regulation of AI in Prior Authorization and Claims Review: A Look at Federal and State Consumer Protections," May 6, 2026. Undark, "Will AI Fix Prior Authorization, or Make It Worse?," July 15, 2026.