Medicare's six-state experiment in artificial-intelligence-assisted prior authorization reduced the use of several targeted services in its first six months, according to a JAMA research letter published October 26. Epidural steroid injections, the highest-volume service in the pilot, fell about 15 percent in the six states compared with the rest of the country. Other targeted services fell about 20 percent.
The study is the first independent evaluation of the WISeR model, short for Wasteful and Inappropriate Service Reduction, which launched on January 1, 2026, in Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington. Medicare pays contractors to review prior authorization requests for 41 services in traditional Medicare, and the contractors are allowed to use AI to screen requests. A human clinician must review a request before a denial, and the contractors' compensation includes a share of the savings the pilot generates.
The authors, Hannah Neprash of the University of Minnesota and Michael Barnett of Brown University, used electronic health record data from Epic Cosmos, a database of more than 300 million records. They compared the pilot states with all other states and the District of Columbia over two years before the launch and six months after. The design is observational, so the results describe an association rather than a proven cause, and the paper says the data cannot show whether the care that did not happen was unnecessary or needed.
The pilot is an experiment in saying no
WISeR is a test of a simple proposition: that a portion of Medicare spending goes to services of low or no value, and that prior authorization can screen them out. The pilot covers a list of 41 services that CMS considers prone to overuse, including skin substitutes, epidural steroid injections, cervical fusion, electrical nerve stimulator implants, and incontinence-control devices. Emergency care, inpatient-only procedures, and care that would be high-risk if delayed are excluded.
The model differs from ordinary prior authorization in two ways. It applies only to original Medicare, not to Medicare Advantage plans, which already use prior authorization extensively. And it pays the contractors partly on savings, a design meant to align their incentives with Medicare's budget rather than with the volume of approvals.
The contractors are Cohere Health in Texas, Zyter in Arizona, and Virtix Health in Washington. CMS has said the AI may screen requests but cannot issue a denial on its own. A clinician employed by the contractor must review the case first. How consistently that rule is followed is one of the questions the pilot's critics have raised, and the study does not address it.
Prior authorization itself is not new. Insurers and Medicare Advantage plans have used it for years, and doctors have complained about it for just as long. What is new is the scale and the automation. A model can review thousands of requests a day, flag the ones that deviate from a pattern, and route them to a human reviewer. That changes the economics of review: the marginal cost of a denial request falls, and the volume of reviews a contractor can process rises. The pilot's designers say that is the point, because a low-cost review can catch low-value care that a human reviewer would never have time to find.
The decline was largest at the start
The headline numbers are large enough to matter for a program that covers millions of people. Epidural steroid injections, which are commonly used for back and leg pain, fell about 15 percent. The other targeted services, taken together, fell about 20 percent. The comparison group showed no comparable change over the same period.
The timing is more interesting than the averages. In the first quarter of 2026, epidural injections in the pilot states were about 21 percent below the comparison states. By the second quarter, the gap had narrowed to about 9 percent. For the other targeted services, the reduction went from about 21 percent to about 18 percent. The effect was strongest when the program was new and then partly faded.
There are at least two readings. One is that doctors and patients adjusted: clinicians learned which requests would be approved, documentation improved, and some care shifted to settings or codes outside the pilot's list. The other is that the initial drop captured a backlog of requests that were never going to be approved, and that the steady state is the smaller number. The study cannot distinguish between them, and the authors say the fade is consistent with clinicians adapting quickly.
The control group is the strongest part of the design
The study's most persuasive feature is its control. CMS had planned to add other services to the pilot but never did. Those services showed no meaningful decline in the pilot states. If the drop in epidurals and the other targeted services were caused by something else, such as a regional shift in medical practice or a change in coding, the control services should have moved too. They did not.
The comparison across states has limits. The six pilot states differ from the rest of the country in demographics, medical practice, and insurance mix, and the study adjusts for those differences only through the comparison group. The authors also cannot rule out that some of the decline reflects care moving out of the measured settings rather than care not happening at all.
The data source has limits of its own. Epic Cosmos covers records from health systems that use Epic, which is a large share of American hospitals but not all of them. Patients who received care outside that network, or who paid cash, may not appear. The study measures service use, not diagnoses, so it cannot say whether the patients who did not get an injection had a herniated disc, spinal stenosis, or nothing that warranted treatment.
The study cannot see the patient
The paper is explicit about what it does not measure. It does not measure pain, function, complications, or patient satisfaction. It does not measure total spending, so it cannot say whether the pilot saved money after accounting for substitute care. It does not measure whether patients who were denied an injection later had surgery, took opioids, or went without treatment.
That gap is the heart of the policy dispute. If the forgone injections were mostly low-value, the pilot is doing what it was designed to do. If a meaningful share were necessary, the pilot is imposing a cost on patients that the study cannot detect. Neprash put the problem plainly: the key question the researchers cannot yet answer is whether the care not received was wasteful or necessary.
The authors also note that the pilot's first six months may not represent its steady state. Prior authorization programs often show their largest effects at launch, when clinicians are unfamiliar with the rules, and smaller effects later as behavior adapts. Whether the fade continues, stabilizes, or reverses will determine how much of the initial decline was a one-time adjustment.
The fight has been going on without the data
WISeR has been contested since before it started. Medical groups have complained about delays and administrative burden, and some have said the automated screening produces denials that a human reviewer later reverses. A Senate resolution to overturn the pilot failed 46 to 50 in July, with supporters arguing that the program was imposed without enough evidence and opponents arguing that Medicare needs tools to control low-value care.
The study gives both sides something. Supporters can point to a measurable reduction in services that CMS itself identified as prone to overuse, with a control group that behaved as expected. Critics can point to the fade, the missing patient outcomes, and the fact that the contractors are paid partly on savings, which gives them a reason to deny.
The size of the affected population is part of the stakes. KFF estimated that the six pilot states had about 6.4 million traditional Medicare beneficiaries in 2024, and that about 207,500 of them received one of the pilot's services, roughly one-fifth of the 1.1 million beneficiaries nationally who did. The program is small relative to Medicare overall, but it covers a set of services that are common among older adults with back pain and other chronic conditions.
CMS has acknowledged problems with the rollout. A senior official said some components of the program were not working as intended, and reporting has described requests that took far longer than the 72-hour standard the program set, including one that waited 83 days. The agency has said it is monitoring the contractors and can adjust or terminate the pilot.
What the next six months will show
The next round of data will answer some of the open questions. If the decline in epidurals stabilizes around 9 percent, the pilot's steady-state effect is smaller than its launch effect, and the savings estimate will need to be revised down. If it falls again, the initial fade was a learning curve. If it disappears, the pilot may be mostly a temporary administrative burden.
The more important data may come from outside the claims records. Studies that track pain and function for patients who were denied care would tell the public what the service-use numbers cannot. So would an audit of how often the contractors' AI screening and human review disagree. Neither is part of the current evaluation, and both are within CMS's power to commission.
There is precedent for that kind of study. Evaluations of prior authorization in Medicaid and in Medicare Advantage have paired claims analysis with surveys of clinicians and patients, and some have found that denials are reversed on appeal at high rates. Those findings do not settle whether the initial review was appropriate, but they show that the administrative process itself can be a source of error. The WISeR study measures only the volume of care, not the quality of the decisions that produced the change.
For now, the study establishes two things. The pilot changed behavior quickly, and the change was largest at the start. Whether that counts as success depends on a question the data cannot answer: what happened to the patients who did not get the injection.
Primary sources
- Neprash HT, Barnett ML. Changes in Service Use Following the Introduction of AI-Enabled Prior Authorization in Medicare. JAMA. 2026. doi.org
- MedicalXpress, report on the JAMA study, October 27, 2026. medicalxpress.com
- Medical Daily, report on the pilot and the study's caveats, October 27, 2026. medicaldaily.com
- Medical Daily, report on the epidural injection findings, October 27, 2026. medicaldaily.com