The Denial Reads Back

On March 12, 2026, a federal judge in Minnesota ordered UnitedHealth Group to produce internal documents about the nH Predict algorithm. The order asked for the design records that would show whether the tool was built to override a treating physician's judgment on post-acute care for Medicare Advantage patients. The order came in Estate of Lokken v. UnitedHealth Group, a class action that has been grinding forward since 2023 and that survived a preemption motion in February 2025 on breach-of-contract grounds. Judge John Tunheim's February 13, 2025 ruling established that a Medicare Advantage plan cannot invoke Medicare preemption to escape its own contractual promises. The March 12 order operationalized that ruling. The model has to come out of the box.

Nineteen days later, on March 31, 2026, the first tranche of a separate CMS transparency requirement came due. Every Medicare Advantage, Medicaid managed care, CHIP, and Marketplace payer had to begin publicly reporting prior authorization turnaround times, denial rates, appeal rates, and overturn rates. The dashboards are live. The denial engines that ran quietly in the back office for a decade now sit on public record.

By January 1, 2027, roughly four months from today, the same payers must expose four HL7 FHIR APIs, including a Prior Authorization API that returns an approval or denial status in a machine-readable envelope with the clinical rationale attached. Da Vinci CRD, DTR, and PAS. SMART on FHIR. OAuth. The rule is CMS-0057-F, finalized January 17, 2024. It is not a suggestion.

I am writing this because the payor tech AI consulting market is still selling a product from 2023.

The old advantage was silence

The economics of the American health insurance denial were straightforward for a long time. Deny in volume, be right on maybe eighty percent, and reap the twenty percent that never appeal. Cigna's PXDX algorithm let doctors clear denials in batches of hundreds at 1.2 seconds of review per claim, a workflow ProPublica reported in March 2023 and a class action has been picking at ever since. UnitedHealth's nH Predict, per plaintiff filings, produced denials that were overturned on appeal at rates approaching ninety percent. That was not a bug in the profit model. That was the model.

The number that carried the book was 11.5 percent. Kaiser Family Foundation analysis put the 2024 Medicare Advantage prior authorization appeal rate around that figure, meaning almost nine out of ten disputed decisions went unappealed. If overturn on appeal runs at 80.7 percent, as a January 2026 Health Affairs study from Stanford's Health Policy team documented, the denial engine's expected value is not the denial itself. It is the friction that stops the patient or the practice from filing at all.

Every incumbent payor tech stack was optimized for that gap. The prior authorization workflow was designed to be exhausting on purpose. The letter arrives on paper. The fax is the interface of record. The peer-to-peer window is 48 hours during business days in a specific time zone. The clinical rationale is a paragraph of copied CPT guidance with no reference to the specific patient. Everything about the artifact discourages a response.

That is what the AI wave first sold into. Faster denial. Cheaper denial. Denial at higher first-pass accuracy against the plan's medical necessity criteria. Cohere Health, Waystar's Myndshft (acquired 2022), Optum's internal stack, dozens of platform vendors pitching utilization management. The pitch was the same denial at lower cost per determination. Some of the technology is quite good. The bet under the pitch is that 11.5 percent holds.

It does not hold in 2026.

Four forces collapsed on the same quarter

The reason the incumbent stack breaks is that four independent forces converged inside a single three-month window between March and June 2026, and each one attacks a different pillar of the old model.

The March 31 CMS transparency rule made denial and overturn statistics a matter of public record for the plans that serve Medicare Advantage, Medicaid managed care, CHIP, and Marketplace enrollees. That is a change in the observability of the denial. The denial used to be a private transaction between plan and provider. It is now a data point in a NAIC-adjacent dashboard that a plaintiff's firm, a state attorney general, a benefits consultant, or a competitor can pull on demand. When the overturn column reads 78 percent for cardiac imaging in Michigan, that denial engine is public data.

The state-law wave hit at the same time. Minnesota's HF2500 passed both chambers of the legislature in May 2026 and prohibits health insurance carriers from using algorithms or artificial intelligence as the sole basis for a prior authorization decision. It requires a licensed physician to review any denial. According to the Transparency Coalition, at least fourteen additional states passed comparable AI-in-health-insurance measures in the first half of 2026, with Holland & Knight tracking companion bills advancing in California, Texas, Illinois, and New York. This is no longer one regulator. It is a fifty-state patchwork with civil penalties, and each state's statute has its own definition of what "sole basis" means.

The CMS-0057-F operational requirements took effect January 1, 2026, before the FHIR API deadline. Payers now owe a decision inside 72 hours for urgent requests and seven calendar days for standard requests, with the January 2027 API deadline forcing every part of the workflow onto rails a machine can consume. This is the pillar most payor tech vendors are actively working on, and it is also the pillar that most quietly guarantees the old model cannot survive. Once the denial is a structured FHIR resource with a documented rationale in a Da Vinci CRD payload, the appeal becomes a structured FHIR resource too. Both sides of the trade become code.

The final force is the one the incumbents have been slowest to price. On the provider and patient side, the auto-appeal is here. Muni Health, Claimable, Fight Health Insurance, and a widening set of provider-embedded tools inside Waystar, R1, and Availity now generate a full appeal letter from the denial rationale and the patient's clinical record in under a minute. The marginal cost of an appeal has been collapsing for eighteen months. The 11.5 percent appeal rate is not going to hold when the appeal itself costs less than a stamp.

Multiply the four forces together and the incumbent denial book does not survive contact with 2027. The Forbes piece Dara Abasiita ran on June 9, 2026 called this out under a headline that named it exactly: the algorithm counted on no one appealing.

The math of an eighty percent overturn

Consider the balance sheet of a mid-sized regional payer with a Medicare Advantage book of 400,000 lives. Suppose that book generates roughly 200,000 prior authorization decisions per year, and suppose the plan denies fifteen percent of them at first pass. That is 30,000 denials. Under the old 11.5 percent appeal rate, the plan sees roughly 3,450 appeals a year, of which around 2,780 come back overturned. The plan pays for those but keeps the remaining 26,550 denials as retained margin against its medical loss ratio target. The math works because the appeal never arrives.

Now hold the denial engine constant and change one variable. Suppose auto-appeal tools push the appeal rate from 11.5 percent to 60 percent, a number that provider-side vendors are already reporting for specialty pharmacy and orthopedic books. The plan now sees 18,000 appeals a year, of which around 14,500 come back overturned. The retained-margin bucket collapses from 26,550 denials to about 15,500. That is a roughly forty percent reduction in the load the denial engine was carrying on the medical loss ratio.

Now add the CMS transparency dashboard. When the state regulator sees a public 78 percent overturn rate on cardiac imaging denials for this plan, an examination follows. The examination looks not at the individual denials but at the design of the tool that produced them. That is where the March 12, 2026 order in the nH Predict case matters as precedent, because it establishes that the design records are discoverable and that plan-level preemption arguments do not shelter contract claims. The exam ends in a corrective action plan, sometimes a market conduct fine, and in California and New York it can end in a bar on the specific tool.

Now add the state law. Minnesota's HF2500, taken at face value, converts a batch-denial workflow into a per-decision physician review that the plan must staff and document. The unit economics of the batch approach depend on the ratio of denials issued to physician hours consumed. A statute that resets that ratio toward one to one is not a compliance overlay. It is a redesign of the operating cost of the utilization management function.

Any one of these four forces is a headwind. Together they mean the denial engine that generated the old margin is not the same asset it was in 2023.

What most payor tech AI consulting is still selling

The market for payor tech AI consulting split around 2024 into two visible camps and one quieter camp that matters more than either.

The first visible camp sells faster denial. Better clinical criteria matching, tighter clinical documentation retrieval, cleaner Da Vinci CRD payloads that make it easier for the plan to say no with the rationale attached. Every vendor here is real. The product being optimized, however, has a shrinking market. Selling faster denial into a market where the denial is being priced back at the plan is like selling faster hooves to a livery stable in 1912.

The second visible camp sells compliance tooling. FHIR API implementation, Da Vinci profile validation, prior authorization API gateways. This work has to happen. January 1, 2027 is a real deadline with real financial penalties. The problem with treating this as the strategic answer is that it is a floor rather than a ceiling. Every payer will have the API by 2027, or they will not be in the Medicare Advantage business. The API does not differentiate anyone. It merely makes the denial legible on the way out the door.

The third camp is quieter, and it is the one that matters. It sells the design of an adjudication process that can survive being read back. Read back by a state examiner reading the public overturn dashboard. Read back by opposing counsel deposing the tool. Read back by a provider's AI appeal generator that will take the denial rationale, cross-reference it against the same clinical evidence base the plan used, and file the strongest available counter-argument inside 24 hours. The strategic asset for a payer in 2027 is a defensible decision, produced by an adjudication process built to be read back at every turn.

A defensible decision has an audit trail with named clinical sources. It reconciles the plan's medical necessity criteria to a specific patient's chart in a way a physician actually reviewed and signed. It documents the alternative treatments considered. It is grounded in a plan-specific coverage policy that traces to a public evidence base. When the same denial is generated on the same patient the following day, it produces the same rationale, because the process is reproducible and not a stochastic pull from a proprietary black box.

That is an operating model. No vendor sells it as a feature.

Payor tech AI consulting has to change what it sells

For AI consulting that actually helps payors, the work in 2026 is not tool selection. It is closer to what forensic accountants do when they redesign a book of business that has to survive an audit.

Start with the observability layer. Every prior authorization decision, whether generated by an AI-assisted workflow or a fully human review, has to land in an evidence store the plan itself can query as if it were a state examiner. If the plan cannot answer the question "show me every denial for spinal fusion in Q2 2026, grouped by the specific medical necessity criterion applied, with the physician reviewer named and the average review time attached" inside ten minutes on its own data, it has no defensible read on its own exposure. Building that layer is table stakes.

Then rebuild the clinical rationale generator. Structured rationale in a Da Vinci CRD payload that names the coverage policy, the evidence citations, and the patient-specific findings it relied on. The rationale must be a document a physician would author. Not a template. Not a generic clinical criteria dump. When an auto-appeal reads that rationale, it should find a coherent argument to answer, not a boilerplate paragraph it can shred in one paragraph.

Then reconcile the plan's medical necessity policies to the current evidence base. A quiet finding across recent state examinations has been that plan-level coverage criteria in a meaningful share of denied cases lag the current clinical evidence by several years. When the plaintiff's expert produces a 2024 systematic review contradicting the plan's 2019 policy, the denial fails not because the AI got it wrong but because the policy underneath the AI was wrong. This is a governance workstream, not a modeling workstream. It has to be built.

Then instrument the human-in-the-loop. Minnesota's law and its coming peers do not just require a physician's involvement. They require a physician who actually reviewed the specific case and can be named on the record. The plan needs a review workflow that captures reviewer identity, time on case, and the specific clinical basis for the physician's concurrence with or override of the AI recommendation. When a court demands the design records, this is what the court is looking for.

Finally, price the appeal into the P&L up front. The days when appeal volume was a rounding error are ending. Plans that model their prior authorization P&L as if 60 percent of denials will now be appealed within 72 hours, and that reserve accordingly, will price their premiums correctly. Plans that model it as if 2024 assumptions hold will underprice their book and find out during rate review.

The vendor answer is not the answer

Every one of the operational shifts above can be assisted by a vendor. Cohere Health can do the clinical intelligence side. Availity can do the eligibility and API rails. Waystar and Optum can do the workflow. Snowflake and Databricks can do the evidence store. This is not a market where the tools are missing.

What is missing is the architectural work that connects the tools to the strategic reality that the denial is now a public act on a machine-readable rail with a machine-generated counter-response inside 24 hours. That work is a decision by the operator, one level above the CIO and the Chief Medical Officer, about what the plan actually wants its denial rate and its overturn rate to say about it on a public dashboard for the next decade. No vendor sells that.

I have watched payor executives in the last eight months treat CMS-0057-F as an IT project, treat the state-law wave as a compliance line item, and treat the auto-appeal side as a curiosity. That framing misses the interaction effect. The four forces do not compound linearly. They compound the way risk compounds in a book where every layer of the underwriting model was correlated in ways nobody priced.

The plans that come through 2027 with intact margins will be the plans that spent 2026 redesigning the adjudication process itself around the assumption that every decision will be inspected by a machine, by a regulator, and by a court. Tool selection and vendor bake-offs are the ground floor of that work, not the work itself.

The architecting move

Payor tech AI consulting is our natural home at Agor AI Advisory. We build the design work. Not the Da Vinci payload, not the FHIR gateway, not the utilization management platform. The upstream architecture that decides how the plan makes a defensible decision, at what latency, with which human review model, and against which reconciled evidence base. Then we work with the tool vendors already in the stack to configure them to that architecture rather than the reverse.

A vendor can sell you a faster denial. A vendor can sell you a compliant API. Neither of them can sell you a book of business that survives the reading-back that CMS, fourteen state legislatures, and every provider's auto-appeal are already doing to your existing one. That is architectural work, and it has to be done by someone whose job is the architecture rather than the sale of a component.

If your plan is heading into 2027 with a denial engine that assumes an 11.5 percent appeal rate, you are underwriting a book that already reset. The window to redesign is short. The redesign itself is possible. It is not a vendor conversation.

The Denial Reads Back

The old advantage was silence

Four forces collapsed on the same quarter

The math of an eighty percent overturn

What most payor tech AI consulting is still selling

Payor tech AI consulting has to change what it sells

The vendor answer is not the answer

The architecting move

Sources

Related Content

Adversarial Inference

The Underwriter Got Here First

The Wait Was The Business

Want this kind of automation working for your business?