# UnitedHealth AI Governance and the nH Predict Case > A sourced UnitedHealth AI governance case study on the unresolved nH Predict dispute, prior authorization data, and the limits of the public record. - Source: https://insureaiwire.com/unitedhealth-ai-governance/ - Publication: InsureAI Wire - Author: Simon Li - Updated: 2026-08-02 --- The nH Predict dispute puts four governance questions into one public record. What did the system do? Where did recommendation end and coverage authority begin? Which outcomes were monitored? Could the plan reproduce work performed by an affiliate or contractor? The litigation remains unresolved, and the federal inspector general report discussed below does not mention algorithms. Those limits make the case useful: it requires a reader to separate allegation, company position, government data, and inference before drawing a control lesson. UnitedHealth's scale makes those evidence seams visible. The company said in April 2026 that it expected to invest nearly $1.5 billion in AI-related initiatives during the year.[^2] In March it described Avery as the latest of more than 1,000 AI solutions across UnitedHealth Group.[^7] Neither statement proves that a particular system is safe, effective, or responsible for a disputed decision. They show why a large carrier needs to connect each deployed system to inventory, authority, testing, monitoring, escalation, and third-party records. For a walk through the regulatory documents, use the [insurance AI governance map](/ai-governance-in-insurance/) and the guide to the [NAIC AI Evaluation Tool's four exhibits](/naic-ai-evaluation-tool/). This article stays with the case and the limits of its public evidence. --- ## What the company says it is deploying UnitedHealth's first-quarter 2026 prepared remarks describe a portfolio of administrative and consumer-facing tools.[^2] The examples below are company statements; independent audit findings are unavailable: - **Optum Real**, an Optum Insight product for administrative work including claim adjudication and coverage validation.[^2] - **PreCheck**, an Optum Rx prescription approval tool. The company says it cut approval time from more than eight hours to under thirty seconds, with missing-information denials down 68% and appeals down 88% on the company's account [^2]. - **Avery**, a generative AI assistant for care navigation, available to roughly 6.5 million employer-sponsored members and 160,000 Medicare Advantage members, with a stated target of 20.5 million members in total by the end of 2026 [^7]. - **Avery's scheduling function**, which the company says can call network primary-care providers on a member's behalf.[^7] The count is the point. A carrier with three AI systems can keep them in a spreadsheet. A carrier reporting more than a thousand needs governance infrastructure that operates at a different order of magnitude. Regulatory or independent verification of these performance figures is unavailable. They are the company's own statements, and that is worth holding in mind for the rest of this piece. > **Table [row-headers]:** What UnitedHealth's AI program covers, on the company's own account > > | System | Function | Scale as reported | > |--------|----------|-------| > | Optum Real | Claim adjudication and coverage validation support | Company describes administrative use | > | PreCheck | Prescription approval (Optum Rx) | 8 hours → 30 seconds | > | Avery | Care navigation chatbot | 6.5M commercial plus 160K Medicare Advantage now, 20.5M targeted by year-end | > | Avery scheduling | Calls network primary-care providers on a member's behalf | Company-described capability | --- ## Where the governance pressure converges UnitedHealth's AI program combines scale with concentration in operations that a carrier may place in its own high-risk tier. The NAIC Evaluation Tool then asks detailed questions about systems in that tier. **[Prior authorization](/glossary/prior-authorization/)** is the clearest overlap. The [health operations guide](/ai-in-health-insurance-claims-prior-auth-risk-adjustment/) explains the distinct records for authorization, claims, and risk adjustment. The NAIC Evaluation Tool's [Exhibit C](/glossary/exhibit-c/) collects a detailed record on each system the company has itself put in the high-risk tier, and the tool says in as many words that those risk criteria are set by the insurance company [^3]. Nothing in it designates prior authorization, or any other operation, as high risk. A carrier whose model decides whether a service is covered, when, and on what conditions has to reach that judgment on its own and then live with it. On the Q1 2026 call the company reported that nearly 95% of UnitedHealthcare's requests now arrive electronically, that about half of those are processed live, and that 90% are approved within a single business day on average [^2]. Speed is the part that gets reported. What Exhibit C asks of a carrier clearing that volume is narrower and harder: the model's name and version, when it went in, whether it was built inside or bought, how its outputs were tested and validated and how they are watched now, when it was last tested, and how it is reviewed against the unfair trade practices and unfair claims settlement laws [^3]. Volume is what makes those records hard to assemble after the fact. **Post-acute care denials** are where the scrutiny has already arrived. A putative class action in Minnesota claims UnitedHealth and its subsidiary naviHealth used the nH Predict algorithm to limit post-acute care, including denials of skilled nursing facility admissions. On February 13, 2025 Judge John Tunheim dismissed five of the seven counts with prejudice, holding them preempted by the Medicare Act, and let the breach of contract and implied covenant claims proceed on the ground that deciding them requires only reading the insurance agreements [^4]. In March 2026 a magistrate ordered [broad discovery into how the tool was built and overseen](/news/unitedhealth-nh-predict-discovery/) [^8]. No court has ruled on whether the allegations are true. The company disputes them: it says the algorithm was not used to make coverage decisions, that it does not dictate how long patients can stay in care facilities, and that it is used in line with regulators' guidance [^1]. Separately, a federal inspector general report examining nineteen [Medicare Advantage](/glossary/medicare-advantage/) organizations found that requests processed by naviHealth were denied 14% of the time, against 11% for requests the plans handled internally and 9% for other contractors [^5]. Of the naviHealth denials that were appealed, 97% were overturned [^5]. The useful comparison sits in the appendix: naviHealth was at 96.6% and the other contractors at 97.4%, against 88.6% for denials the plans reviewed in-house [^5]. On appeal, naviHealth's denials therefore held up marginally better than the rest of the contractor field. The report's headline 95% covers all nineteen organizations and includes UnitedHealth Group, which received 42% of the appeals and overturned 99.7% of them [^5]. Only 18% of SNF denials were appealed at all, which is what turns a reversal rate into a question about the other 82% [^5]. The words algorithm, artificial intelligence, and predictive model appear nowhere in the report's thirty-two pages; where the OIG pointed at a cause, it pointed at the training and oversight the plans gave the contractors doing the reviews [^5]. Set the litigation aside and what remains is a monitoring signal of exactly the kind the NAIC Model Bulletin expects a carrier to catch on its own [^6].
Two comparisons from the HHS inspector general report on skilled nursing facility prior authorization, June 2026. Denial rates by who processed the request: naviHealth 14 percent, the Medicare Advantage plans handling requests in-house 11 percent, other contractors 9 percent. Of the denials that were appealed, the share overturned in the enrollee's favor: other contractors 97.4 percent, naviHealth 96.6 percent, plans reviewing in-house 88.6 percent. naviHealth denies at a higher rate than its peers, but on appeal its denials do not stand out. Only 18 percent of denials were appealed at all, by enrollees or their providers. SNF DENIAL RATE, BY WHO PROCESSED THE REQUEST 14% NAVIHEALTH 11% PLANS, IN-HOUSE 9% OTHER CONTRACTORS OF DENIALS APPEALED, SHARE OVERTURNED OTHER CONTRACTORS 97.4% NAVIHEALTH 96.6% PLANS, IN-HOUSE 88.6% Enrollees and their providers appealed only 18% of SNF denials.
FIG. 1: NAVIHEALTH STANDS OUT ON DENIALS, NOT ON APPEALSSOURCE: HHS OIG, REPORT OEI-09-24-00331, JUNE 2026, KEY TAKEAWAYS AND APPENDIX C
**Agentic AI**, systems that act with limited human oversight, is the emerging frontier. Agents that place scheduling calls sit at the automated end of the tool's autonomy bands, which the tool defines as operating without human intervention [^3]. Exhibit C applies only after the insurer has classified the system as high risk. The autonomy band is one of the thirteen fields recorded at that point; the insurer's own criteria control entry into the exhibit [^3]. The governance question is what happens the first time an agent built to that pattern is pointed at a benefit question instead, and whether the classification moves with it. --- ## The governance gap at scale UnitedHealth makes familiar governance gaps visible at unusual scale. The same gaps can appear at ten use cases; a thousand simply makes them harder to close. **The inventory problem.** [Exhibit A](/glossary/exhibit-a/) of the NAIC Evaluation Tool wants every AI system counted and bucketed by operational area [^3]. UnitedHealth puts its own count above a thousand. The [inventory playbook](/ai-inventory-by-line-of-business/) owns the register design; the case question is whether that register can keep pace as a recommendation feature evolves into an automated action. **The vendor responsibility problem.** The NAIC Model Bulletin expects an insurer's program to cover AI systems whether they were built in-house or acquired from a third party [^6]. naviHealth is a UnitedHealth subsidiary. The case still tests whether the plan can reproduce work performed elsewhere in its corporate structure. The [vendor risk assessment](/ai-vendor-risk-assessment/) carries the diligence, contracting, and ongoing-oversight detail. **The adverse outcome tracking problem.** Every contractor with enough appeals for the OIG to publish a rate came in at 96 percent or higher, and at the plans reviewing in-house the figure still ran near ninety [^5]. That is a signal without being a diagnosis. It is consistent with initial criteria that were too strict, with an appeals process more generous than the intake, or with review capacity that never caught up to volume. The public data leave the cause and underlying monitoring unknown. The [model monitoring playbook](/ai-model-monitoring-insurance/) owns threshold and response design; this case shows why denial, appeal, reversal, and access measures need to be read together. The OIG's own wording is that the pattern "raises concerns about initial denials." --- ## What the case exposes, and what it does not Four control questions emerge from the public record. **Table [row-headers]:** Evidence seams and control questions raised by the UnitedHealth public record | Evidence seam | Question raised by the record | |---|---| | Scope | Can a company keep a current view of more than a thousand uses as features and purposes change? | | Authority | Where does verification or recommendation end and the coverage determination begin? | | Outcomes | Which denial, appeal, reversal, and access measures are reviewed together, with the right denominators? | | Organizational boundary | Can the plan reproduce work performed by an affiliate or contractor as part of its own decision process? | The public record leaves any causal relationship between nH Predict and the OIG's cross-industry findings unestablished. The report is silent on algorithms [^5], while the litigation allegations remain pending [^9]. A universal risk tier for prior authorization is equally unsupported because the Evaluation Tool leaves those criteria to the insurer.[^3] Operating methods remain with the health-operations, inventory, vendor, and model-monitoring owners linked where each issue first appears. A case study should expose the control question without becoming a replacement playbook. --- The record contains no finding against UnitedHealth. The two surviving claims are in fact discovery, under a schedule that does not reach trial readiness until the end of 2027 [^9]. The OIG report describes a reversal pattern across nineteen organizations and never mentions an algorithm at all. The company's position is that the model was not used to make coverage decisions. The case is therefore more useful as a specimen than as a verdict. It shows which four questions a large AI program has to answer for every system it runs: what the system does, who can override it, what is monitored, what was tested. And it shows those questions arriving from a court, an inspector general, and a state insurance department at once, in whatever order they please. Those questions transfer to a carrier running ten systems and no dedicated budget. The public record supplies no verdict to transfer. [^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]: UnitedHealth Group, [First Quarter 2026 Results: Teleconference Prepared Remarks](https://www.unitedhealthgroup.com/content/dam/UHG/PDF/investors/2026/unh-q1-2026-remarks.pdf), April 21, 2026. The remarks state that the company expected to invest nearly $1.5 billion in AI-related initiatives in 2026 and provide the prior-authorization, PreCheck, and Optum Real figures used here. [^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 [^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 [^4]: Memorandum Opinion and Order on Defendants' Motion to Dismiss, *Estate of Gene B. Lokken et al. v. UnitedHealth Group Inc. et al.*, No. 0:23-cv-03514-JRT-SGE (D. Minn. Feb. 13, 2025), Doc. 91. Counts 3 through 7 dismissed with prejudice as preempted by 42 U.S.C. § 1395w-26(b)(3); Counts 1 and 2 survive because the court "will only need to evaluate compliance with the insurance agreements" (Conclusion, p. 23): https://storage.courtlistener.com/recap/gov.uscourts.mnd.211721/gov.uscourts.mnd.211721.91.0_2.pdf [^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/ [^7]: UnitedHealth Group, "UnitedHealthcare introduces AI companion, empowering people with simpler navigation and a personal experience," March 26, 2026: https://www.unitedhealthgroup.com/newsroom/2026/2026-03-26-uhc-introduces-ai-companion-empowering-people-with-simpler-navigation-personal-experience.html [^9]: Order Granting Joint Motion to Amend Pretrial Scheduling Order, *Estate of Gene B. Lokken et al. v. UnitedHealth Group Inc. et al.*, No. 0:23-cv-03514-JRT-SGE (D. Minn. May 21, 2026), Doc. 175: fact discovery due March 11, 2027, dispositive motions due July 29, 2027, ready for trial December 6, 2027. Docket: https://www.courtlistener.com/docket/68006832/estate-of-gene-b-lokken-the-v-unitedhealth-group-inc/ [^8]: Order on Motion to Compel Discovery, Estate of Gene B. Lokken et al. v. UnitedHealth Group Inc. et al., D. Minn., March 9, 2026: https://litigationtracker.law.georgetown.edu/wp-content/uploads/2023/11/Estate-of-Gene-B.-Lokken-et-al-v.-UnitedHealth-Group-Inc.-et-al_2026_3_9_ORDER-ON-MOTION-TO-COMPEL-DISCOVERY.pdf