medicare ai prior authorization in 2026: what the court record actually shows

Medicare now runs AI prior authorization on both sides of the program. Medicare Advantage insurers have faced lawsuits over it since 2023, and since January 2026 traditional Medicare has its own AI pilot in six states [source].

The public argument is about accuracy. How often is the model wrong, and how many denials get overturned on appeal? That is the obvious question, and the court record says it is not the one that decides liability.

The two federal courts that have ruled on Medicare Advantage AI denials never reached the model’s accuracy. They let the cases proceed on a narrower question: did the insurer do what its own documents said it would do?

What medicare ai prior authorization means

Medicare AI prior authorization is the use of an algorithm to help decide, before or during care, whether Medicare or a Medicare Advantage plan will pay for a service. The rules allow the algorithm to assist. They do not allow it to replace an individual assessment of the patient.

The proof: two courts, one question

In Lokken v. UnitedHealth, the plaintiffs allege UnitedHealthcare used the naviHealth “nH Predict” model to deny post-acute care to elderly Medicare Advantage members. On 13 February 2025 Judge John Tunheim dismissed most claims as preempted by the Medicare Act, but let breach of contract and breach of the implied covenant of good faith proceed. TIN’s Lokken case file quotes the reasoning: the court “would only be required to investigate whether UHC complied with its own written documents”.

Those documents say coverage decisions are made by clinical staff and physicians. That is the whole hook. The claim survives Medicare preemption because it asks about a promise, not about a regulation.

Barrows v. Humana, in the Western District of Kentucky, alleges the same tool was used to deny the same kind of care. On 14 August 2025 Judge Rebecca Grady Jennings let breach of contract, good faith, unjust enrichment and common-law fraud proceed, and dismissed four other counts, including insurance bad faith, with prejudice, per TIN’s Barrows case file.

The Lokken complaint alleges that “more than 90% of claim denials and over 80% of preauthorization denials are overturned on appeal” [source], and the Barrows complaint pleads, upon information and belief, that over 90% of denials are reversed ([source]). Both case files show that figure as a contested allegation, not a measured one. UnitedHealth states the tool “is not used to make coverage determinations”; Humana calls the allegations “sensationalized” and “false”.

Is AI allowed in Medicare prior authorization?

Yes, AI is allowed in Medicare prior authorization, as an input and not as the decision. On 6 February 2024 CMS told Medicare Advantage plans that an algorithm deciding coverage “based on a larger data set instead of the individual patient’s medical history, the physician’s recommendations, or clinical notes would not be compliant” [source].

The same guidance addresses post-acute care, the exact service in both lawsuits. Software “may predict how long patients may stay, but this prediction cannot be used as the predicate to terminate coverage” [source].

What the Senate found on post-acute denials

A Senate Permanent Subcommittee on Investigations report in October 2024 found that in 2022 UnitedHealthcare and CVS denied post-acute prior authorization requests at roughly three times their overall denial rates, and Humana at more than 16 times [source].

The report put UnitedHealthcare’s post-acute denial rate at 8.7% in 2019 and 22.7% in 2022, and named nH Predict as part of that strategy [source]. Those are Senate staff findings, not court findings, and the lawsuits have not adopted them.

How is AI used in the Medicare prior authorization pilot?

The WISeR model runs from 1 January 2026 to 31 December 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. It reviews three service groups, including skin and tissue substitutes and knee arthroscopy for knee osteoarthritis, using AI and machine learning “along with human clinical review” [source].

Its page makes one promise in plain words: “All recommendations for non-payment are determined by appropriately licensed clinicians” [source]. After Lokken, that sentence is no longer reassurance. It is the standard every WISeR denial can be measured against.

The bottom line

The industry is preparing to defend its models on accuracy, and the courts are not asking about accuracy. They are asking whether the sentence “a clinician decides” was true on the day a claim was denied. That is cheaper to prove and harder to fake than any error rate.

The transferable rule for anyone automating a yes or no decision: your liability lives in the sentence you wrote about who decides, so write only the one your logs can back.