In a two-month window in 2022, physicians working for Cigna denied more than 300,000 claims using an internal tool called PxDx, spending an average of 1.2 seconds reviewing each one before signing off on the denial — according to internal company data reviewed by ProPublica and The Capitol Forum. One former Cigna doctor described the workflow bluntly: doctors would "click and submit" batches of fifty claims at a time, with no chart review. Cigna has said PxDx was designed to speed up approval of routine, low-risk claims, not to drive denials at scale — but a federal class action in California, allowed to proceed by a judge in March 2025, alleges the opposite: that the system automated wholesale denial with only the appearance of physician review.
This is the least-discussed AI story in healthcare. Everyone is talking about AI reading radiology scans or drafting clinical notes. Almost no one is talking about the fact that the single highest-volume application of AI in the entire healthcare system is deciding whether a bill gets paid — and that this decision now gets made, on both sides of the transaction, by competing algorithms that have never met each other but effectively negotiate in real time, at machine speed, over every American's medical bill.
Two algorithms, one claim, zero humans in the room
The mechanism is straightforward once you see it as an arms race rather than a single tool:
On the provider side, revenue-cycle-management (RCM) vendors use AI to do three things before a claim ever reaches a payer: predict which claims are likely to be denied and pre-empt the denial with better documentation, automatically select billing codes that match payer-specific coverage policy, and draft appeal letters when a denial does land. Cohere Health, a prior-authorization platform used by several national payers, reports that roughly 85% of prior-authorization requests now receive an immediate AI determination, with the remaining 15% routed to a nurse or physician reviewer; the company has processed more than 15 million PA submissions through its API layer. Waystar, one of the largest RCM platforms in the country, markets AI models that flag high-risk authorizations before submission and generate denial-appeal drafts, citing internal client data claiming billions of dollars in denials prevented industry-wide.
On the payer side, insurers use AI for utilization review and, in the most litigated cases, for denial automation itself. The Cigna PxDx system is one example. A second, more consequential case involves UnitedHealth Group's post-acute-care subsidiary, NaviHealth, and its nH Predict algorithm. A class action filed in federal court in November 2023 alleges that nH Predict was used to estimate how many days of skilled nursing or rehab care a Medicare Advantage patient would need — and that the tool had roughly a 90% error rate, measured by how often its projected discharge dates were overturned on appeal. The suit alleges the company kept relying on the algorithm's shorter stay predictions because fewer than 1% of denied members ever file an appeal. UnitedHealth has said the tool is not used to make coverage determinations and that final decisions are based on CMS criteria and plan terms — a defense now being tested in court.
Regulators are drawing a line, unevenly
The Centers for Medicare & Medicaid Services responded to the nH Predict litigation and similar concerns with clarifying guidance tied to its 2024 Medicare Advantage final rule (CMS-4201-F). The rule and subsequent FAQ make explicit that an algorithm or predictive tool cannot, by itself, be the basis for terminating post-acute coverage; MA plans must still base coverage decisions on the individual patient's circumstances, including physician recommendations and clinical notes, not solely on a model's output. It also bars plans from using undisclosed internal coverage criteria embedded in an algorithm or letting AI silently drift a plan's coverage standard over time. Critically, CMS's posture is permissive, not prohibitive: it does not ban AI from coverage determinations, it bans AI from being the sole and final word.
That is a narrower guardrail than it sounds. It applies to Medicare Advantage, not commercial insurance where the Cigna litigation sits, and it regulates the payer side only — there is no comparable federal rule constraining how aggressively provider-side RCM vendors optimize codes or auto-generate appeals. The result is a regulatory environment built one lawsuit and one FAQ memo at a time, chasing a technology deployment that scaled faster than the rulemaking cycle.
Where this goes next
Both cases have kept moving through 2026. In the nH Predict litigation, a federal magistrate ordered UnitedHealth in March 2026 to turn over broad discovery on the algorithm's design and use, and a ruling on class certification is expected by mid-2026. The Cigna PxDx case, by contrast, has seen no ruling beyond the March 2025 decision allowing it to proceed. The more consequential 2026 shift may be legislative rather than judicial: more than 240 AI-in-healthcare bills were introduced across 43 states in 2026, and Indiana, Alabama, Maryland, Iowa, Washington, Georgia, and Utah have all enacted laws barring insurers from using AI as the sole basis for a claim denial, downcoding, or prior-authorization refusal — a direct legislative response to cases like PxDx and nH Predict that the 2024 federal CMS rule left unaddressed.
The economics here are unusually clean, which is exactly why adoption has outpaced scrutiny: a percentage point of denial rate, or a day of length-of-stay, is worth real money at the scale of a national payer or a multi-hospital health system, and AI is simply better than a human reviewer at finding those percentage points thousands of times a day. That is not inherently a scandal — it is the same logic that makes AI valuable in fraud detection or underwriting. The scandal risk shows up specifically at the point where a prediction about a population gets applied to a decision about one patient, without anyone checking that the two match. Expect the next two years to bring more state-level AI-in-claims legislation modeled on California's and New York's recent bills, more discovery-driven lawsuits revealing internal override and appeal-success rates vendors have never had to disclose, and a slow but real shift toward tools that keep a licensed clinician's judgment as the final, auditable step rather than a rubber stamp. The health systems and payers that treat that transparency as a design requirement, not a legal defense, will be the ones still standing when the regulation catches up.