Healthcare Payer Loss-Ratio AI: How AI Is Transforming MLR Management in 2026

UnitedHealth's full-year 2025 adjusted medical care ratio jumped to 88.9 percent from 85.5 percent in 2024, a 340 basis-point deterioration. CEO Stephen Hemsley announced $1.5 billion in AI investment with nearly $1 billion in 2026 operating cost reductions, many AI-enabled. Cigna delivered $6 billion in net income up 73 percent on an 82.2 percent MLR. Here is what payer AI for MLR management actually looks like in 2026, including the cautionary tales (nH Predict, PXDX) that defined what not to do.

Is PA AI dead after nH Predict and PXDX?

No. PA AI done with clinical review, transparent appeals, and provider feedback (the Cohere Health pattern) is the new standard. The dead pattern is the high-volume / no-review / no-appeal model.

UnitedHealth Group reported on January 26, 2026 that its full-year 2025 adjusted medical care ratio was 88.9 percent compared to 85.5 percent in 2024, a 340 basis-point deterioration (UnitedHealth Group press release via BusinessWire). UnitedHealthcare's segment MLR hit 85.1 percent (up 330 bps year over year) and Q4 2025 reached 86.3 percent. CEO Stephen Hemsley publicly tied the recovery plan to AI: $1.5 billion in AI investment across the company, with nearly $1 billion in 2026 operating cost reductions anticipated, many AI-enabled (Fierce Healthcare).

That is the headline of the entire 2026 payer AI story. Every major U.S. health insurer is at or above the 85 percent MLR ceiling for large-group plans. Q1 2025 MLR comparisons from Oliver Wyman put Aetna CVS Health at 87.3 percent, Elevance at 86.4 percent, UnitedHealthcare at 84.8 percent, and Cigna at 82.2 percent (Oliver Wyman, May 2025). The winners and losers list is now visible: Cigna's 2025 net income was nearly $6 billion (up 73 percent year over year) on the lowest MLR; CVS Health's 2025 net income dropped 61.8 percent to less than $1.8 billion; Elevance's net income fell 5 percent to $5.7 billion (Becker's Payer Issues).

This blog is the practical operating brief for payer AI in MLR management in 2026. The categories that work, the ones that produced class actions, and what a defensible 2026 deployment looks like.

Where the AI is actually moving the MLR needle

Four AI categories are showing up in payer earnings calls and in vendor revenue 2025-2026.

Payment integrity. Codoxo closed an oversubscribed $35M Series C on December 17, 2025, led by CVS Health Ventures. Codoxo's platform now supports payment-integrity operations for more than 80 million covered lives, with 100 percent customer retention and 125 percent net revenue retention in 2025 (BusinessWire). CVS Ventures leading the round is the clearest payer-side endorsement of pre-claim cost prevention as MLR strategy.

Prior authorization done right. Cohere Health raised a $90M Series C in May 2025 on reported metrics: 94 percent provider satisfaction, up to 85 percent real-time authorization approvals, 1-3 percent inpatient medical-expense savings in utilization management, 8x ROI for payments (PR Newswire). The PA AI category that works ties directly to MLR (1-3 percent inpatient savings) without the class-action exposure.

Care management for high-cost members. GLP-1s and gene therapies are the cost story of 2026. AI for high-cost-claimant identification, care coordination, and specialty pharmacy steerage is the highest-leverage MLR move per dollar invested. Most payers are running this work internally rather than buying.

Stars and HEDIS automation. Medicare Advantage Star Ratings hit a low point in 2025 with marginal improvement in 2026, and V28 risk adjustment implementation completed for the 2026 plan year (Oliver Wyman, October 2025). The double-whammy of Stars decline and V28 risk adjustment compression is crushing 2026 MA revenue, making AI for documentation, HEDIS gap closure, and Stars improvement indispensable.

The cautionary tales that defined what not to do

Two named PA-AI failures shape every 2026 deployment conversation.

UnitedHealth nH Predict. UnitedHealth's NaviHealth subsidiary used an AI system called nH Predict, built on a database of 6 million patients, to evaluate post-acute care coverage. The system reportedly had a 90 percent error rate, and the company's denial rate jumped from 10.9 percent to 22.7 percent after implementation (Healthcare Finance News). Judge Tunheim allowed the class-action claims to proceed on February 13, 2025.

Cigna PXDX. Cigna's PXDX algorithm let doctors deny over 300,000 claims while spending just 1.2 seconds reviewing each request (Alignmt AI).

The lesson is not that PA AI is wrong. It is that PA AI without medical review, transparent appeals, and defensible documentation is now a class-action liability. In June 2025, more than 50 major health plans (including UnitedHealthcare, Aetna, Cigna, Humana, Elevance Health, Kaiser Permanente, Centene) pledged to the federal government to simplify PA processes and commit to real-time approvals for at least 80 percent of requests (Thompson Coburn). UnitedHealth itself reduced or removed reauthorization requirements for more than 180 drugs and is targeting elimination of PA for 30 percent of services still needing approval by year-end (TS2.tech). The 2026 PA market is unwinding the prior generation while leaving a vacuum for smarter AI.

What "AI ready" looks like for a payer in 2026

The payer AI deployment that survives a state insurance commissioner inquiry has four components.

Component 1: Documented model design and training data. Per the December 2025 SEC OCA stance on AI in regulated industries, every AI system that affects a coverage decision needs a documented model card, training-data lineage, and refresh cadence. The nH Predict deployment failed this test publicly.

Component 2: Clinical review pathway. Every adverse coverage decision driven in part by AI needs a documented clinical review by a qualified human reviewer. The 1.2-second-review pattern is the failure mode.

Component 3: Transparent appeals and provider feedback loop. Cohere Health's 94 percent provider satisfaction is the benchmark. The AI is a participant in the workflow, not a hidden gatekeeper.

Component 4: Audit trail and rebate reserve. ACA MLR rebates totaled nearly $958 million paid to more than 6 million consumers in 2024 (KFF). The MLR floor is real. The audit trail that proves the AI made its decisions inside the regulated framework is also what defends the rebate calculation.

The 90-day plan

Days 1 to 30: identify the top three MLR-impact use cases (typically a combination of payment integrity, PA-augmentation, and a high-cost-claimant care-management workflow), score the data layer against the four-component readiness rubric, document the clinical-review pathway for each.

Days 31 to 60: deploy one use case end-to-end with full audit trail and clinical review. Payment integrity is usually the lowest-friction starting point because the regulatory regime is well-defined and the savings attribution is direct.

Days 61 to 90: stand up the appeals and provider feedback loop, document the model design and training data, complete one state insurance department disclosure rehearsal.

By day 91 the payer has one production AI deployment, an audit trail that survives a state insurance commissioner inquiry, and a real MLR-impact number for the next plan-year filing.

Frequently asked questions

Is PA AI dead after nH Predict and PXDX?

No. PA AI done with clinical review, transparent appeals, and provider feedback (the Cohere Health pattern) is the new standard. The dead pattern is the high-volume / no-review / no-appeal model.

What's the right MLR-impact target for the first deployment?

Most payment-integrity deployments deliver 0.5 to 1.5 MLR points within four quarters of full production. PA-augmentation deployments deliver 1 to 3 points on managed inpatient categories. Star/HEDIS automation rarely moves MLR directly but improves the revenue side via Stars rebates.

What about Medicare Advantage V28?

V28 risk adjustment implementation was complete for the 2026 plan year. The compression is in the rearview. AI for HCC capture, documentation accuracy, and provider coding support is now table stakes and is required regardless of MLR work.

What does the SEC OCA stance mean for payer AI?

The SEC's December 2025 framing requires documented model design, data lineage, and human oversight for AI affecting financial reporting. For public payers, that includes any AI affecting MLR calculation or claims reserve estimation. Documentation is the burden.

What are the rebate-rule implications?

Federal MLR rebates triggered when individual or small-group MLR fell below 80 percent or large-group below 85 percent. AI investments that produce sub-floor MLR have to fund the rebate. AI investments that hold MLR right at or above the floor are the optimization target.

If you want the longer version of this analysis, including the four-component readiness rubric scored, the regulatory matrix, and the 90-day deployment plan with named vendors, our AI Readiness, Data Governance Consulting, and AI Workflow Automation Consulting practices ship the operating model. The full healthcare strategic context is in our 2026 Healthcare AI Spend Map. Anchor case studies: the Express Scripts AI Readiness and Kaiser Permanente Metric Governance engagements are the closest published Thinklytics work to the payer AI operating model described here.

How do we measure AI's actual impact on medical loss ratio?

Compare MLR cohorts with and without the AI intervention over a 12-month period, controlling for member mix and benefit design changes. Most payers see a 1.5 to 3 percentage-point MLR improvement attributable to AI across the four use cases combined. Single-use-case attribution is harder; multi-use-case rollups are cleaner.

Can a regional plan compete on AI with the national carriers?

Yes, in narrow domains. Regional plans win in network optimization and member-engagement AI because their geographic concentration produces higher data signal per square mile than national plans get. National carriers win in claims and fraud AI because raw claims volume drives model quality.

Where does Thinklytics start with a new payer engagement?

30-day audit on the unified claim-and-member view. Most payers think their core platform fragmentation is worse than it is, OR much better than it is. The audit produces an honest map of what's resolvable in 8 weeks, 12 weeks, and 6 months. Read more at healthcare analytics consulting.