Courts are judging AI coverage denials by what insurers promised

Courts are judging AI coverage denials by what insurers promised. In the 3 lawsuits that have gone furthest, judges have not ruled on whether an insurer may use an algorithm. They have asked a narrower question. The question is whether the plan kept its promise that clinicians would make coverage decisions, or whether a tool made them instead.

That framing has let the core claims survive motions to dismiss, in Medicare Advantage and in employer coverage alike. It also points to what plans will have to prove. A plan that says a physician decides must be able to show that the physician’s review was real.

Key takeaways

• The core claims are surviving. In the UnitedHealthcare and Humana cases, breach of contract and good faith claims survived motions to dismiss. In the Cigna case, an ERISA fiduciary claim survived.
• The promise is the standard. The courts have asked whether insurers kept commitments about who makes decisions, not whether AI is allowed.
• Preemption trims the cases but does not end them. In the Medicare Advantage cases, claims about how coverage decisions are made were preempted. Claims about the insurer’s own contract were not.
• Exhaustion has not blocked them. The Humana court held that claims challenging the use of AI in the process did not require plaintiffs to exhaust Medicare appeals first.
• Discovery is opening the black box. In March 2026, the UnitedHealthcare court ordered production of the insurer’s claims policies and its AI review board materials, with some categories reaching back to 2017.
• Plans should be able to prove review. A clinician’s signature on a denial is not evidence that the clinician decided. Plans need records that show the review happened.

The cases so far

The 3 cases were filed in 2023, after reporting by ProPublica and other outlets on how large insurers used automated tools in coverage decisions. Each has passed the motion to dismiss stage with key claims intact.

Compiled from the Georgetown Health Care Litigation Tracker and court records. Rulings are described from coverage by ArentFox Schiff (April 2026) and Courthouse News (2025). The claims are allegations, and none of the cases has reached a decision on the merits.

Lokken and UnitedHealthcare

The Lokken case was filed in November 2023 by the families of 2 Medicare Advantage enrollees. It challenges UnitedHealthcare’s use of nH Predict, a tool from its naviHealth subsidiary that estimates how long a patient will need post-acute care. In February 2025, the court granted the motion to dismiss in part. Claims for breach of contract and breach of the implied covenant of good faith survived, and other claims were dismissed as preempted by Medicare Advantage law, according to ArentFox Schiff.

The surviving claims turn on a promise. Plaintiffs allege that UnitedHealthcare’s plan documents said “clinical staff and physicians” would make claim decisions. The court refused to split discovery into phases, describing the issues as “enmeshed factual and legal issues.”

In March 2026, a magistrate judge largely granted the plaintiffs’ motion to compel. UnitedHealthcare was ordered to produce its policies and training on post-acute claims. The order also covered documents analyzing nH Predict and materials on its internal AI review board. Some categories reach back to January 2017. The court also required the identities of medical directors and care coordinators who took part in denials for about 300 putative class members. A later summary by Freeman Mathis & Gary noted that requests for the tool’s source code and underlying data were denied.

The ruling cited a Senate investigation that found UnitedHealthcare’s post-acute denial rate “more than doubled” after it began using naviHealth and nH Predict in 2019. Our analysis of post-acute utilization management covered that report.

Barrows and Humana

The Barrows case, filed in December 2023, makes similar allegations about Humana’s use of nH Predict. In an August 2025 order, the court let the contract and good faith claims proceed, along with a claim of fraud by omission. It dismissed several claims under state insurance statutes, which the plaintiffs had largely conceded were preempted.

The court described the central question as whether Humana’s AI use “violated its contract with insureds.” It did not ask whether AI use is prohibited under the Medicare Act. Plaintiffs allege that Humana’s contracts promised “individualized assessment by medical professionals.” A motion for class certification is scheduled for November 2026, according to the Georgetown tracker.

Kisting-Leung and Cigna

The Cigna case concerns employer coverage, not Medicare. Plaintiffs allege that Cigna used an algorithm called PXDX to deny claims in bulk without individual review by medical professionals. ProPublica reported that a single Cigna physician denied as many as 60,000 claims in a month, at about 1.2 seconds each, as Courthouse News noted.

In March 2025, Judge Dale Drozd let an ERISA claim for breach of fiduciary duty proceed for some plaintiffs. He held that Cigna’s reading of its plan’s medical necessity requirement, “as allowing an algorithm to make the decision so long as a medical director pushes the button,” was an abuse of discretion. A claim under California’s unfair competition law survived for some plaintiffs. Claims for wrongful denial of benefits were dismissed with leave to amend, because plaintiffs had not tied them to specific plan terms.

What the courts are converging on

The promise is the standard

None of the rulings says an insurer may not use AI. Each asks whether the insurer did what it told members it would do. In the Medicare Advantage cases, the promises sit in the evidence of coverage and other plan documents. In the Cigna case, the standard is the plan’s own medical necessity terms, read under ERISA’s fiduciary duty.

The words that matter are often ordinary ones. Phrases about clinical staff or individualized assessment appear in member materials that were written to reassure, not to define a process. In these cases they have become the standard the process is measured against.

That approach makes the plan’s words decisive. A plan that promises individualized review by medical professionals has to deliver it. A plan that uses a tool to make decisions while describing a clinician’s judgment invites exactly these claims.

Accuracy has not been tested

None of the rulings so far says whether the tools were accurate. The Humana court accepted the complaint’s allegations as true at the pleading stage and did not evaluate nH Predict. That question may come later through discovery and expert evidence. A tool could be accurate on average and still produce decisions that break a plan’s promises, if no clinician actually reviewed the cases. The contract theory does not depend on the tool being wrong. It depends on who made the decision.

Preemption narrows the cases

Medicare Advantage law preempts many state law claims, and both Medicare Advantage courts dismissed claims under state insurance statutes. The Humana court held that preemption must be analyzed claim by claim. Claims about how coverage decisions are made, which federal rules govern, were preempted. Claims that turn on interpreting the insurer’s own contract were not, and the court followed the Lokken analysis in reaching that result.

The practical effect is that plaintiffs cannot use these cases to challenge Medicare coverage rules. They can use them to hold an insurer to its own words.

Exhaustion has not been a barrier

Medicare Advantage enrollees usually must use the program’s appeal process before suing over a denial. The Humana court held that most of the plaintiffs’ claims did not arise under the Medicare Act, because they challenged the use of AI in the process rather than individual benefit decisions. Those claims did not require exhaustion. For the claim tied most closely to benefit decisions, the court waived exhaustion. It cited irreparable harm and futility.

That reasoning matters for any plan. If a claim is about the design of the decision process, the appeals process may not be the only path a member has.

Why the cases reach beyond the defendants

The theory in these cases is not limited to Medicare Advantage or to post-acute care. Any plan whose documents promise clinician review can face the same question. Medicaid managed care plans and plans sold on the exchanges publish member handbooks and coverage policies that describe how decisions are made. Employer plans have plan documents that ERISA courts read for exactly this purpose, as the Cigna ruling showed.

Discovery is a second reason the cases matter. The Lokken order reached a plan’s AI governance, including the membership and work of its internal AI review board. Plans that have created such boards should assume their minutes and decisions could be examined someday. That is an argument for running them well, with written reasons for approving each tool and evidence that someone monitors it after deployment.

Regulators are moving in the same direction. CMS told Medicare Advantage plans in February 2024 that a prediction alone cannot be the basis to terminate post-acute services, and that algorithms alone cannot deny an admission. Several states have written human-review requirements into insurance law. The litigation adds a third source of pressure, and it is the only one that can produce damages for members.

Vendors are part of the picture too. In both Medicare Advantage cases, the tool came from a vendor owned by or working for the insurer. A plan that delegates review or licenses a tool still owns the promise it made to members. The contract with the vendor should say who reviews and what records the vendor keeps, so the plan can show its process if asked.

What meaningful review would have to show

The cases will turn on facts that discovery is now producing. Whatever the outcome, they have defined the question every plan using AI in coverage decisions should be ready to answer. When a clinician signs a denial, something in the record has to show that the clinician decided.

A clinician who never disagrees with the tool is not reviewing it.

A signature is not enough. The Cigna ruling’s phrase about a medical director who “pushes the button” describes a review that exists on paper only. Evidence of real review would include several kinds of records.

1. Who reviewed. The reviewer’s identity and qualifications, matched to the type of service. Our credentialing coverage explains how plans verify the people who make these decisions.
2. What the reviewer saw. A log of the records the reviewer opened, with times. A decision made without opening the record is not a review of it.
3. Why the reviewer decided. Reasons that cite the patient’s record, not a template that would fit any case. CMS has said a prediction alone “cannot be used as the basis to terminate post-acute care services.”
4. How often the reviewer disagreed with the tool. A reviewer who overrides the tool sometimes is exercising judgment. A reviewer who never does may be confirming its output.
5. What happened on appeal. The share of denials overturned when a member or provider appeals, tracked by reviewer and by tool. A high reversal rate is a sign the first decision was not careful.

The fourth measure deserves emphasis. Tools that recommend denials will be right much of the time, and a good reviewer will often agree. Agreement in every case is different. It suggests the tool is deciding and the reviewer is approving its work. Our case for auditing payer AI like a reviewer described how plans can measure agreement between tools and people, the way they measure agreement between 2 human reviewers.

Plans can build this evidence before anyone asks for it. A sample of denials each month can go to a second reviewer who does not see the first reviewer’s decision. The agreement rate between them, and between each reviewer and the tool, becomes a standing measure. Reviewers whose agreement with the tool is unusually high get a closer look, and so do tools that reviewers override often. None of this requires new technology. It requires deciding that the plan will measure review, not only count it.

There are limits to what records can show. Time spent on a case is a weak measure, because some decisions are simple. Access logs can be gamed. Disagreement rates depend on the tool’s quality as well as the reviewer’s diligence. Records like these do not prove careful judgment on their own. Together they make a careless pattern visible, and they let a plan show a court or a regulator what its reviewers actually did.

What plans should write down

The cases also argue for clearer plan documents. Members and providers are entitled to know what role a tool plays in a decision. State law is moving the same way. California’s SB 1120 requires that “final medical necessity determinations may be made only by a licensed physician” or qualified licensed provider, as our preemption analysis described.

A plan’s documents could say plainly what its tools do. A tool may approve a request that meets published criteria. A tool may recommend a decision or summarize the record for a reviewer. A denial is made by a clinician who reviews the record, and the denial notice states the clinician’s reasons. Those sentences describe a defensible process, and a plan that writes them down commits itself to running one.

The denial notice is where the promise becomes visible to a member. Since January 2026, plans covered by CMS’s 2024 rule must give a specific reason for every prior authorization denial. A reason written by the reviewer, citing the member’s record, shows that a person engaged with the case. A reason copied from the tool’s output, the same for every member, shows the opposite.

The same clarity belongs in contracts with vendors. A plan that delegates utilization review remains responsible for the decisions made in its name. NCQA’s 2026 updates added requirements for how credentialing delegation agreements address the integrity of information, as our NCQA analysis described. A delegation agreement for review that uses AI should also say who makes denials and what evidence of review the vendor keeps and shares.

Where Redlake stands

Redlake builds tools that health plans use in utilization review. Our first rule is that a machine can say yes, but only a person can say no. The litigation shows why the rule has to be real and not nominal. A clinician who signs what a tool produces has not made the decision the rule requires.

Our other rules are built for the evidence these cases demand. Every recommendation shows its reasons, so a reviewer sees what the tool saw and can disagree with it. Every decision can be reconstructed. The record shows who reviewed the case and why they decided it. Those records protect members first. They also protect plans that do the work, because they turn a claim about process into a question of fact the plan can answer.

The design choice that matters most is where the tool’s work ends. A tool can organize the record and apply published criteria. When the criteria are met, it can approve. When they are not, it hands the case to a clinician with the reasons laid out. The clinician’s decision is recorded as the clinician’s own.

We think every plan using AI in coverage decisions should track how often its reviewers disagree with its tools. A plan that cannot answer that question cannot show that its clinicians are deciding.

What to watch in the rest of 2026

• Class certification in Barrows. The motion scheduled for November 2026 will show whether the Humana case proceeds on behalf of a broad group of enrollees.
• Discovery in Lokken. Watch for disputes over the materials the court ordered, which could describe how nH Predict was used in practice.
• The amended Cigna claims. Watch whether plaintiffs tie their wrongful denial claims to specific plan terms, as the court required.
• New cases. Watch for suits against other insurers and vendors, particularly where plan documents promise clinician review.

About this analysis

Method. The case descriptions come from published rulings and law firm summaries of the 3 leading cases on algorithmic coverage decisions, with case documents and status from the Georgetown Health Care Litigation Tracker.

Data notes. The descriptions of each case reflect the allegations and the courts’ rulings at the motion to dismiss and discovery stages. None of the courts has decided whether the insurers breached their contracts or duties. None of it is legal advice. The evidence framework is our proposal.

Sources. ArentFox Schiff (April 2026). Freeman Mathis & Gary (July 2026). Georgetown Health Care Litigation Tracker, including the August 2025 order in Barrows. Courthouse News (2025). CMS, answers on the CY2024 final rule (February 2024).

Frequently asked questions

What is the Lokken lawsuit against UnitedHealthcare?

It is a proposed class action in Minnesota federal court, filed in November 2023, alleging that UnitedHealthcare used the nH Predict tool to deny post-acute care to Medicare Advantage enrollees. Breach of contract and good faith claims survived a motion to dismiss, and the court ordered broad discovery in March 2026.

Have courts ruled that insurers cannot use AI to deny claims?

No. The rulings so far ask whether insurers kept promises about who makes coverage decisions. They have not held that AI use is prohibited, and none has decided the merits.

Why did some claims survive Medicare Advantage preemption?

Courts held that contract claims require only interpretation of the insurer’s own contract and are not preempted. Claims about how coverage decisions are made, which federal rules govern, were preempted.

Could these claims be brought against other kinds of plans?

The theory is not limited to Medicare Advantage. Any plan whose documents promise review by clinicians could face a similar claim, including Medicaid managed care and exchange plans. The Cigna case shows how the theory works for employer plans under ERISA.

What should plans using AI in coverage decisions do now?

Plans should check that their member documents describe accurately what their tools do. They should also keep records that show clinician review, such as access logs and case-specific reasons. Tracking how often reviewers disagree with the tool helps show that the review is real.

What is nH Predict?

It is a tool from naviHealth, a company owned by UnitedHealth Group, that estimates how long a patient will need post-acute care such as a skilled nursing stay. The lawsuits against UnitedHealthcare and Humana allege that the insurers relied on its estimates to end coverage for post-acute care.

What did the court say about Cigna’s PXDX algorithm?

In March 2025, the court held that reading Cigna’s plan to allow an algorithm to decide “so long as a medical director pushes the button” was an abuse of discretion. An ERISA fiduciary claim proceeded for some plaintiffs.