Medicare is weeks away from letting AI screen prior authorisation requests

On 1 January, traditional Medicare will begin doing something it has largely avoided for decades: requiring advance approval for a set of covered services, with artificial intelligence helping decide which requests get through.

The programme is called WISeR — Wasteful and Inappropriate Service Reduction — and it runs as a six-year Centers for Medicare & Medicaid Services model in six states: Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington [s1]. Six technology companies have been selected as model participants to use enhanced technology including artificial intelligence to support medical necessity determinations: Cohere Health, Genzeon, Humata Health, Innovaccer, Virtix Health and Zyter [s1].

The services in scope are narrow and deliberately chosen: skin and tissue substitutes, electrical nerve stimulators, and knee arthroscopy for osteoarthritis, among other high-cost Part B items [s1]. Inpatient care, emergency care and services where a delay would itself create risk are excluded [s1].

The design feature everyone argues about

CMS's position is that coverage and payment rules do not change — WISeR changes only who has to ask first, and how [s1]. Two safeguards are built in. A non-affirmation decision cannot be issued without review by a licensed clinician [s1]. And providers with a demonstrated record of compliance can be exempted from the requirement altogether, subject to monitoring [s1].

The objection is not really about the algorithm. It is about the money. Companies running the model receive a percentage of the savings associated with their work [s2] — compensation tied to the savings produced by their denials [s1].

In November, six House Democrats introduced the Seniors Deserve SMARTER Care Act to repeal the model. Representative Ami Bera argued that the incentive structure creates a "dangerous incentive to put profits ahead of patients' health"; Representative Mark Pocan called the use of AI to determine medical necessity "extremely reckless" [s2]. The bill's other sponsors are Representatives Rick Larsen, Suzan DelBene, Kim Schrier and Greg Landsman [s2].

Whether it moves is a separate question from whether the model starts. As of now, it starts.

What prior authorisation already does, measured

The argument over WISeR is being conducted mostly in the abstract. There is a large body of survey data on what prior authorisation does in the plans that already use it, and it is worth putting on the table before the pilot generates any of its own.

The American Medical Association surveys 1,000 practising physicians annually — 400 primary care, 600 specialists [s3]. In the 2025 survey, 95% reported that prior authorisation at least sometimes delays access to necessary care, and 79% reported that it can at least sometimes lead patients to abandon a recommended course of treatment [s3]. Ninety-two per cent said the process has a somewhat or significantly negative effect on clinical outcomes [s3].

On harm: 26% of physicians reported that prior authorisation had led to a serious adverse event for a patient in their care; 20% reported it had led to a hospitalisation; 22% to a life-threatening event or an intervention needed to prevent permanent impairment; and 8% to a patient's disability, permanent bodily damage, congenital anomaly or death [s3].

On burden: practices complete an average of 40 prior authorisations per physician per week, at a cost of about 13 hours of physician and staff time [s3]. Ninety-four per cent said the process increases physician burnout [s3].

Two caveats belong here. These are physician perceptions, self-reported, from a sample of physicians who chose to respond — not adjudicated adverse-event records. And they describe commercial and Medicare Advantage prior authorisation, not WISeR, which does not exist yet.

But they establish the baseline against which the model will be judged, and they explain why the profession is not neutral. The same survey found that 60% of physicians are concerned that augmented intelligence increases, or will increase, prior authorisation denial rates [s3].

What would count as evidence

WISeR is a demonstration model, which means it is supposed to generate findings. The useful questions are not whether AI can read a claim — it can — but whether the model's determinations track medical necessity as a clinician would assess it, how often non-affirmations are overturned, how long the process adds to a course of treatment, and whether the savings survive once downstream costs from delayed or abandoned care are counted.

None of those are answerable in January. The model runs six years [s1], and the first meaningful data will be the denial and appeal rates rather than anything about outcomes.

There is also a structural point about participation. WISeR is optional for providers in the affected states [s1], and gold-carding exempts the compliant [s1]. Both features shape who ends up inside the model, which will complicate any comparison between WISeR states and the rest of Medicare.

What to watch

Three things. Whether the repeal bill attracts any Republican co-sponsors, which would change its odds considerably. Whether CMS publishes non-affirmation and overturn rates by participant, which is the single most informative disclosure available. And whether the clinician-review requirement [s1] holds as a real check or becomes a signature step — a distinction that will only be visible in how quickly reviews are completed.