UnitedHealth Group, the largest health insurer in the United States, is spending $3 billion on artificial intelligence in 2026. According to Stat News, the company employs 22,000 software engineers and has built more than 1,000 AI solutions across its organization. On March 26, 2026, UnitedHealthcare launched Avery, a generative AI companion that schedules appointments, coordinates care, and helps members navigate their insurance. It is a genuinely useful product. It also shares a company with nH Predict, an AI algorithm facing a class action lawsuit that alleges it denied Medicare Advantage claims at a rate that appellants overturn more than 90 percent of the time.

Two AIs, One Company

UnitedHealth is running two fundamentally different experiments in healthcare AI simultaneously, and the contrast between them is the most useful case study available for any healthcare IT leader planning AI deployments.

The first is member-facing. Avery, available through the UnitedHealthcare app, books primary care appointments on a member's behalf, learns from each interaction, and personalizes navigation for insurance benefits. According to UnitedHealthcare's press release, Avery is currently available to 6.5 million commercial members and will expand to 20.5 million commercial, Medicare, and Medicaid members by the end of 2026. It is designed to reduce the friction that leads patients to delay care, skip preventive appointments, and miss medication refills.

The second is payer-facing. nH Predict, developed by UnitedHealth subsidiary naviHealth, generates recommendations on post-acute care coverage for Medicare Advantage members. According to the class action lawsuit proceeding in federal court, nH Predict regularly overrides physician recommendations and generates coverage denials for care that physicians have determined is medically necessary. The lawsuit cites an appellate reversal rate exceeding 90 percent, meaning nine out of ten appealed denials were ultimately reversed. A federal judge has allowed the case to continue, and according to Insurance News Net, ordered UnitedHealth to produce AI decision documentation by April 29, 2026.

The Governance Gap in Healthcare AI

The tension between Avery and nH Predict is not a UnitedHealth-specific problem. It is a preview of the governance challenge every healthcare organization will face as AI moves deeper into clinical and administrative workflows.

AI in healthcare can serve two goals: patient outcomes or operational efficiency. In a well-designed system, these align. Keeping patients healthier costs less than treating preventable crises. In a poorly designed system, operational efficiency wins by narrowing authorizations, shortening coverage windows, and generating denials for care the algorithm considers statistically atypical.

The problem is that patients cannot tell which goal an algorithm serves. When Avery books a primary care appointment, the member benefits directly. When nH Predict denies a skilled nursing facility stay, the member may have no idea that an algorithm made that decision or how to challenge it. According to Stat News, patients frequently do not know when AI is influencing decisions about their care, or whose interests the system optimizes for.

The regulatory response is forming. The DOJ is conducting criminal investigations into UnitedHealth's Medicare billing practices. The class action will produce discovery that, for the first time, requires a major insurer to document in detail how an AI made coverage decisions at scale. The FDA is simultaneously re-examining what constitutes a breakthrough AI device, according to Stat News analysis from April 2026, shifting its focus toward AI systems that solve problems physicians cannot rather than those that simply replicate existing capabilities.

What This Means for Healthcare IT Leaders

The UnitedHealth situation clarifies a design decision that every healthcare organization building or buying AI must make: who does the system serve, and how do you prove it?

AI that faces members and clinicians needs to optimize for outcomes and trust. Avery is a reasonable model. It provides navigation value without making consequential coverage decisions. When Avery makes a poor appointment suggestion, the failure is low-stakes and reversible. The governance bar is lower because the harm ceiling is lower.

AI that makes coverage, authorization, or clinical pathway decisions operates in high-stakes, irreversible territory. A skilled nursing denial affects a real patient's recovery. The governance requirements must match: documented decision logic, physician override requirements, real-time audit trails, disparate impact monitoring across patient demographics, and direct integration with the appeal process. Building these controls in retrospect, which is what UnitedHealth is now being compelled to do through litigation, is far more expensive and reputationally damaging than designing them in from the beginning.

What To Do About It

1. Classify every healthcare AI deployment by decision stakes before launch. Member navigation tools, scheduling automation, and benefit explanation AI operate at low stakes. Prior authorization AI, coverage determination AI, and care pathway recommendation AI operate at high stakes. Each class needs a different governance framework, and that classification should happen before deployment, not after an adverse event.

2. Build immutable audit trails for every algorithmic decision. Any AI system that influences a coverage, authorization, or care decision must log every decision with the inputs, model version, confidence score, and timestamp. Courts are now requiring this documentation on discovery timelines measured in weeks. Organizations that have it respond in days. Organizations that do not spend months reconstructing records under legal pressure.

3. Require physician override for any AI that touches clinical judgment. An AI system that can override or bypass physician recommendations without a documented, auditable escalation path is a liability, not an asset. Physician override is not a flaw in the design. It is the compliance control that separates defensible AI from indefensible AI.

4. Track your AI's appellate reversal rate as a performance metric. If your organization uses AI for coverage determinations and is not monitoring how often those determinations are reversed on appeal, start now. A reversal rate above 30 percent signals model miscalibration. A rate above 50 percent is a governance emergency that needs immediate escalation.

HRIM's Take

The UnitedHealth story is not a cautionary tale about healthcare AI. It is a clarifying one. AI will be embedded in healthcare administration at every level over the next five years, and most of it will benefit patients. The question for every healthcare organization is whether they design their governance frameworks before deployment or in response to litigation. UnitedHealth built member-facing AI that members appreciate, and payer-facing AI that courts are now scrutinizing. The difference is not the technology. It is whether the system was designed to serve the patient or to optimize around them. The organizations that answer that question honestly before deployment will be the ones that scale AI without a federal judge ordering their documentation by a specific date.