UnitedHealth's $3 Billion AI Push Raises the Governance Bar for Insurers

A nurse drives to a patient’s home. Before she arrives, an AI system has read the medical chart aloud in summary form so she knows what to expect. The same system, at the same company, listens to millions of customer calls to find patterns in complaints. In a trial phase, it is now calling doctors’ offices to schedule appointments on behalf of patients.

These are pilot projects no longer. They are live operations at UnitedHealth Group, the largest U.S. health insurer, which plans to invest $3 billion in artificial intelligence across 2026 and 2027. The company says it is already seeing a 2-to-1 return ¹.

For a compliance officer at a smaller carrier, the natural response is to watch from a distance. UnitedHealth’s scale, its litigation exposure, and its regulatory relationships are not yours. But the governance questions its AI program raises are universal, and the scrutiny it is under is a preview of what happens when AI moves from back-office efficiency to front-line decision-making at scale. The NAIC Model Bulletin and the AI Systems Evaluation Tool were not written with UnitedHealth in mind, but UnitedHealth is the case that will test whether they work. For a walk through what those tools ask, see our guide to the NAIC AI Evaluation Tool’s four exhibits.

This article covers what UnitedHealth is building, where its AI program intersects with the NAIC governance framework, and what any carrier should document before its own AI investment reaches a fraction of that scale. If you are still building your AI inventory, start with our guide to the NAIC Model Bulletin’s four pillars.

---

What $3 billion buys

UnitedHealth’s AI investment breaks into two parts. Roughly one-third goes to software products and platforms, accelerating Optum Insight’s shift toward an AI-first services model. The remaining two-thirds flows into end-to-end processes across the group ².

The signature applications include:

• Optum Real, a real-time coverage verification system that has processed about a billion transactions since launch, allowing providers to check whether a service is covered before delivering it ¹.
• PreCheck, a prescription approval tool that has cut approval time from more than eight hours to under thirty seconds, with missing-information denials down 68% and appeals cut 88% ².
• Avery, a generative AI assistant for care navigation, currently available to roughly 6.5 million employer-sponsored members and 160,000 Medicare Advantage members, with over 20 million slated for access by year-end ².
• AI agents calling doctors’ offices to schedule appointments, currently in trial ¹.
• Crimson AI, used for operating room scheduling, and Members Like You, for care pathway assistance ².

The company has identified more than 1,000 AI use cases ². That number alone is the point. A carrier with three AI systems can keep them in a spreadsheet. A carrier with a thousand needs a governance infrastructure that operates at a different order of magnitude.

What UnitedHealth’s AI program covers

System	Function	Scale
Optum Real	Real-time coverage verification	~1 billion transactions
PreCheck	Prescription approval	8 hours → 30 seconds
Avery	Care navigation chatbot	6.5M members now, 20M+ by year-end
AI agents	Appointment scheduling	Trial phase
Crimson AI	Operating room scheduling	Live
Members Like You	Care pathway assistance	Live

---

Where the governance pressure converges

UnitedHealth’s AI program is not just large. It is concentrated in exactly the areas the NAIC Evaluation Tool labels high-risk.

Prior authorization is the clearest overlap. The NAIC Evaluation Tool’s Exhibit C concentrates on systems that materially influence decisions about coverage, pricing, or claims settlement ³. Prior authorization is all three: it decides whether a service is covered, when, and under what conditions. UnitedHealth processes nearly 95% of its requests electronically, with about half processed live and 90% approved within one business day ². The speed is impressive. The governance question is whether the documentation behind those approvals would survive an Exhibit C review.

Post-acute care denials are where the pressure has already arrived. A class action lawsuit claims UnitedHealth and its subsidiary naviHealth used an AI algorithm to limit care, including denials of skilled nursing facility admissions ⁴. A federal inspector general report found that naviHealth denied 14% of skilled nursing facility admission requests, a higher rate than other Medicare Advantage organizations, and that 97% of those denials were overturned on appeal ⁵. The company disputes the claims and says the algorithm does not dictate coverage decisions ¹. But the pattern, high initial denial rates with near-total reversal on appeal, is the kind of adverse outcome tracking that the NAIC Model Bulletin’s consumer disclosure pillar requires carriers to monitor ³.

Agentic AI, systems that act with limited human oversight, is the emerging frontier. The trial of AI agents calling doctors’ offices to schedule appointments sits at the edge of what the NAIC tool defines as automated: it decides on its own within set parameters ³. If that trial scales, it becomes an Exhibit C system by definition, with the same documentation requirements as a claims denial model.

---

The governance gap at scale

UnitedHealth’s situation is not unique in kind. The scale is different, and that is the point. The governance gaps that appear at a thousand use cases are the same gaps that appear at ten, just harder to close.

The inventory problem. Exhibit A of the NAIC Evaluation Tool asks for a quantitative inventory of every AI system, sorted by operational area ³. UnitedHealth has identified 1,000 use cases. The practical question is not whether they are all listed. It is whether the list is current, whether it captures the systems operated by third-party vendors, and whether it distinguishes between supportive, augmented, and automated applications. A use case that started as a recommendation engine and evolved into an automated approval system is a common drift pattern. The inventory must catch it.

The vendor responsibility problem. The NAIC Model Bulletin is explicit: insurers remain fully responsible for third-party AI systems ⁶. UnitedHealth’s naviHealth unit is a subsidiary, not an independent vendor, but the governance principle is the same. The parent company is accountable for the algorithm’s outputs, its training data, its bias testing, and its adverse outcome tracking. The bulletin does not recognize a distinction between “our AI” and “their AI.” It recognizes whether the AI influences a consumer decision. If you have not yet mapped which of your vendor systems touch coverage, see our guide to the NAIC Model Bulletin’s four pillars.

The adverse outcome tracking problem. The HHS inspector general report found that 97% of naviHealth’s denials were overturned on appeal ⁵. That figure is a signal. It suggests either that the initial denial criteria were too strict, or that the appeal process was too lenient, or that the algorithm was not well-calibrated to the population it served. What it does not suggest is a system that was being monitored for disparate impact. The NAIC Model Bulletin’s risk management pillar requires ongoing testing for unfair discrimination ⁶. A 97% overturn rate is the kind of pattern that testing should surface, and the OIG’s own wording is that it “raises concerns about initial denials.”

---

What your own program should look like

You are not UnitedHealth. Your AI budget is not $3 billion. But the NAIC Model Bulletin and the Evaluation Tool apply to you both, and the questions they ask do not scale with revenue.

Start with the systems that touch coverage decisions. Prior authorization, utilization management, claims adjudication, and settlement recommendations are the systems that land in Exhibit C. If you have AI in any of those areas, the documentation requirements are the same regardless of your size. The tool asks for decision logic, human override mechanisms, monitoring logs, and bias testing ³. Those are not enterprise-grade luxuries. They are the minimum.

Watch the drift from supportive to automated. UnitedHealth’s Optum Real began as a real-time verification tool. Its PreCheck tool began as a prescription approval accelerator. Both are now closer to automated decision systems than to human support tools. The NAIC tool records autonomy levels in Exhibit A, but for risk assessment in Exhibit C, what matters is whether the system influences the outcome ³. If your AI recommends and a human approves, the system is still in scope. The question is whether you can show the human approval means something.

Document the appeals. The naviHealth case is a lesson in what happens when adverse outcomes are not tracked. A 97% overturn rate on appeal is an operational success only if your metric is speed. By any measure of accuracy or fairness, it is a warning. The OIG’s own wording is that it “raises concerns about initial denials.” The NAIC Model Bulletin’s consumer disclosure pillar requires procedures for adverse outcome review ³. If you cannot produce the data on how often your AI-driven decisions are reversed, you cannot show you are monitoring.

---

What to do this week

You are not going to build a $3 billion AI program in a week. You can, however, find out whether your governance documentation is proportionate to your AI exposure.

Map your coverage-touching AI. List every system that influences a coverage, pricing, or claims decision. For each one, note whether it is supportive, augmented, or automated, when it was last validated, and whether it has been tested for disparate impact. If the list is shorter than you expected, you may have systems that were never logged as AI.

Pull your appeal data. For any AI-assisted denial or prior authorization system, produce the last quarter’s appeal rate and overturn rate. If you cannot, that is a gap. If the overturn rate is high, that is a signal. If you are not sure what counts as high, see our guide to why the NAIC Evaluation Tool’s Exhibit C targets claims AI.

Check your vendor files. For any third-party AI system, verify that your contract includes audit rights, model documentation requirements, and ongoing monitoring provisions. The NAIC Model Bulletin treats vendor AI as your AI ⁶. Your diligence files should reflect that.

---

UnitedHealth’s $3 billion bet is a business story. Its governance implications are a regulatory story. The two stories converge at the same question: can you show, system by system, that your AI is making decisions you would stand behind in an exam room? The carriers that answer yes before the question is asked will be the ones that define the standard. The carriers that wait for the exam notice will be building under pressure, at a scale they did not choose.

InsureAI Wire tracks NAIC and state-level AI governance developments weekly. Subscribe for updates on the evaluation tool pilot, state adoption maps, and compliance checklists.

Footnotes

1. John Tozzi, Bloomberg, “UnitedHealth’s $3 billion AI push has bots calling doctors,” Spokesman-Review, June 19, 2026: https://www.spokesman.com/stories/2026/jun/19/unitedhealths-3-billion-ai-push-has-bots-calling-d/ ↩ ↩² ↩³ ↩⁴

2. Elizabeth Casolo, “UnitedHealth is spending $1.5B on AI this year. Here’s where the money is going,” Becker’s Payer Issues, April 21, 2026: https://www.beckerspayer.com/virtual-care/unitedhealth-is-spending-1-5b-on-ai-this-year-heres-where-the-money-is-going/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. NAIC, “AI Systems Evaluation Tool 4.0,” 2026: https://content.naic.org/sites/default/files/inline-files/AI%20Systems%20Evaluation%20Tool%204.0%20%28Clean%29.pdf ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

4. Jeff Lagasse, “Class action lawsuit against UnitedHealth’s AI claim denials advances,” Healthcare Finance News, February 14, 2025: https://www.healthcarefinancenews.com/news/class-action-lawsuit-against-unitedhealths-ai-claim-denials-advances ↩

5. HHS Office of Inspector General, “Medicare Advantage Organizations Overturned Nearly All Appealed Prior Authorization Denials for Skilled Nursing Facility Admission, Raising Concerns About Initial Denials,” Report OEI-09-24-00331, June 8, 2026: https://oig.hhs.gov/reports/all/2026/medicare-advantage-organizations-overturned-nearly-all-appealed-prior-authorization-denials-for-skilled-nursing-facility-admission-raising-concerns-about-initial-denials/ ↩ ↩²

6. NAIC Model Bulletin, “Use of Artificial Intelligence Systems by Insurers,” adopted December 4, 2023: https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf ↩ ↩² ↩³