When an AI denial tool reaches federal court: in Lokken v. UnitedHealth, the breach-of-contract claims over 'nH Predict' survive Medicare preemption

The problem

Medicare Advantage plans decide how long an insurer will pay for post-acute care: the skilled nursing, therapy and inpatient rehabilitation a patient needs to recover from a serious injury or illness [source]. A putative class action pending in federal court in Minnesota alleges that UnitedHealthcare made those decisions with an AI model instead of a clinician: the suit “claims that UnitedHealth’s use of an AI tool called nH Predict led to arbitrary coverage denials without consideration of the patients’ health status or the involvement of a physician” ([source]).

This is the case TIN’s verification thesis is built to examine: an AI step placed inside a consequential professional workflow, and an independent institution (here a federal court) forced to decide what that workflow actually produced. The court’s own words, not a vendor’s, are the record.

What was built, and what is alleged

nH Predict is a predictive model. The complaint, filed on November 14, 2023 in the U.S. District Court for the District of Minnesota as No. 0:23-cv-03514, opens by alleging the case “arises from Defendants’ illegal deployment of artificial intelligence (AI) in place of real medical professionals to wrongfully deny elderly patients care owed to them under Medicare Advantage Plans by overriding their treating physicians’ determinations as to medically necessary care based on an AI model that Defendants know has a 90% error rate” [source]. The court’s own order recites the allegation the same way: “Plaintiffs claim that UHC used an artificial intelligence program, nH Predict AI Model (‘nH Predict’), in lieu of physicians to make coverage determinations” ([source]). As reported, plaintiffs allege the model “compares them to a database of 6 million patients” using inputs such as the patient’s age, diagnosis, living situation and physical function, and that “the model then apparently makes determinations about how much care a patient needs and provides a target discharge date” [source].

Plaintiffs allege in the complaint itself that they were denied care “based on an AI model that Defendants know has a 90% error rate” [source], a figure the legal press reported as an allegation plaintiffs “maintain that defendants are aware of” ([source]), and a second outlet reports the complaint’s claim that “more than 90% of claim denials and over 80% of preauthorization denials are overturned on appeal” [source]. That conflict is shown, not resolved: UnitedHealth denies the premise. Its spokesperson stated “The naviHealth Predict tool is not used to make coverage determinations. The tool is used as a guide to help us inform providers, families and other caregivers about what sort of assistance and care the patient may need both in the facility and after returning home” ([source]), and the same reporting that carried the 90% figure noted UnitedHealth “denies using nH Predict” [source].

The lead plaintiff’s facts anchor the stakes. “Lead plaintiff Lokken, age 91, fell while at home, fracturing his leg and ankle,” received physician-recommended physical therapy “for less than three weeks” before coverage was denied, and the family “paid roughly $12,000 to $14,000 per month out of pocket until he passed away a year later” [source].

The outcome

On February 13, 2025, the court ruled on UnitedHealth’s motion to dismiss. The order states its own holding: “The Court will grant in part and deny in part UHC’s motion to dismiss, allowing the breach of contract and breach of implied covenant of good faith and fair dealing claims to proceed” [source], signed “DATED: February 13, 2025 at Minneapolis, Minnesota. JOHN R. TUNHEIM” ([source]). The legal press reported it the same way, that Judge John Tunheim “granted in part but denied in part the Minnetonka-based health insurance giant’s motion to dismiss” [source], and a second outlet quoted the ruling directly ([source]).

Most of the plaintiffs’ seven Minnesota-law causes of action did not survive: UnitedHealth “successfully argued that the Medicare Act’s broad preemption provision resulted in most of plaintiff’s claims being preempted” [source]. The two that survived did so on a narrow, contract-based reasoning. The order states the plaintiffs’ “claims for breach of contract and breach of the implied covenant of good faith and fair dealing, however, survive preemption because those claims do not aim to regulate the same subject matter as the federal standards” ([source]). The legal press paraphrased the same reasoning, that deciding the surviving claims would require the court “only … to investigate whether UHC complied with its own written documents,” namely “its statement that claim decisions would be made by ‘clinical services staff’ and ‘physicians’ when it allegedly used artificial intelligence” [source].

The practical effect, per the trade reporting, is that the case proceeds and “could reopen the discovery process and allow attorneys representing the aggrieved Medicare Advantage enrollees and their families to seek documents from UnitedHealth, especially pertaining to its use of AI tools” [source]. This is a procedural ruling, not a finding of liability: nothing here decides whether UnitedHealth in fact used nH Predict to deny care, only that the two contract-based claims may be litigated on the merits.

Why this matters for AI-led operations

The court did not weigh in on whether AI is good or bad. It drew a narrower and more useful line for anyone deploying automation in a regulated workflow: the live question is “whether UHC complied with its statement that claim decisions would be made by ‘clinical services staff’ and ‘physicians’ when it allegedly used artificial intelligence” [source]. The exposure is not the model, it is the gap between what an organization tells customers its process is and what an automated process actually does. That gap is exactly what independent verification against the record is meant to close.

A note on source strength. The adjudicated holding and the nH Predict allegations are now quoted directly from the primary court documents, the February 13, 2025 order (Doc. 91) and the complaint (Doc. 1), whose text was extracted verbatim from the archived PDFs and independently corroborated word for word by Minnesota Lawyer and Skilled Nursing News. The weakest load-bearing element is the 90% error rate: even though it is now quoted from the complaint itself, it is a plaintiffs’ allegation, not an adjudicated or independently measured figure, and UnitedHealth denies the tool makes coverage decisions at all. Treat it as a contested claim, not a verified outcome.

Related case files

• Mobley v. Workday, the other landmark case of an automated decision system judged in federal court, where AI hiring tools face a certified nationwide age-discrimination collective.
• Italy’s Garante fines Foodinho, a regulator pricing the harm of an algorithm making consequential decisions about people, here rider management rather than medical coverage.
• Mata v. Avianca, the canonical example of an AI step inserted into a professional workflow with no verification behind it, and a court measuring the result.