Journal / Essay · Health

The Math of the Claim Denial

Insurers use algorithms to deny claims in bulk, while roughly 1% of denials are ever appealed.

Being insured used to mean a doctor decided what care you needed. Increasingly, a piece of software decides in under two seconds — and it's counting on you never filing the appeal that could overturn it.

And Then You're Dead · August 2026 · 8 min read

Having health insurance is supposed to mean a doctor decides what care you need and the insurer pays for it. That's not always how it works anymore. A wave of lawsuits, a Senate investigation, and a wave of federal rulemaking now describe a different sequence: a patient's diagnosis, age, and vitals go into a piece of software, the software predicts how much care that patient "should" need, and the claim gets denied the moment the prediction says stop — sometimes before a human ever opens the file.

This isn't a story about lacking coverage. It's a story about paying the premium, having the policy, filing the claim exactly the way you're supposed to — and losing anyway, to a system built to make losing look procedural.

The Algorithm That Said No

UnitedHealth's post-acute care subsidiary, naviHealth, built a tool called nH Predict. It draws on a database of roughly six million past patients and, given someone's diagnosis, age, and living situation, forecasts exactly how many days of nursing home or rehab care they'll need — then flags the date UnitedHealthcare should stop paying.

Gene Lokken, a 91-year-old Wisconsin man, broke his leg and developed a skin infection after knee replacement surgery. His doctors recommended continued nursing home care. UnitedHealthcare cut off payment anyway, according to the lawsuit his estate filed in Minnesota federal court in November 2023. His family paid $12,000 to $14,000 a month out of pocket until he died about a year later. A federal judge let most of the case proceed past a motion to dismiss in February 2025.

The complaint alleges UnitedHealth kept using an algorithm it knew was wrong roughly nine times out of ten, because it also knew only about two in a thousand patients would ever challenge it.

The claim isn't that the algorithm was occasionally mistaken. It's that its error rate, on the rare occasions someone appealed, ran close to 90% — and that the company's own reviewers were pressured to follow its output rather than override it, according to the complaint.

Denied in 1.2 Seconds

Cigna runs a different system, called PxDx, and it works on a different part of the claim: not how long care should last, but whether a specific test or procedure matches a specific diagnosis code closely enough to pay for automatically. When it doesn't, PxDx flags a "mismatch," and a Cigna medical director signs off on the denial — often in batches of hundreds at a time, without opening the patient's file.

A 2023 investigation by ProPublica and The Capitol Forum found that one Cigna medical director denied around 60,000 claims in a single month. Company records the reporters reviewed showed Cigna doctors spent an average of 1.2 seconds per claim. Over a two-month stretch in 2022, the system was used to reject more than 300,000 payment requests.

Multiple class-action lawsuits followed — first in California, then Connecticut — arguing that batch-signing denials without reading a single chart violates state laws requiring individualized physician review of medical claims. Courts have allowed both cases to move forward.

What the Senate Found

In October 2024, the Senate Permanent Subcommittee on Investigations released a report on UnitedHealthcare, Humana, and CVS/Aetna — three insurers that together cover nearly 60% of all Medicare Advantage enrollees. The subcommittee reviewed more than 280,000 pages of internal documents spanning 2019 to 2022.

CVS's prior-authorization request volume rose 57.5% over the same period — far outpacing its roughly 40% enrollment growth — after the company deployed an AI tool called Post-Acute Analytics in April 2021, built specifically to reduce skilled-nursing-facility spending. The subcommittee's report describes automation as a primary driver behind all three companies' rising denial rates for exactly the care category — nursing homes, inpatient rehab, long-term care — where patients are least able to just pay out of pocket and wait.

What Regulators Did About It

The Centers for Medicare and Medicaid Services finalized a rule in April 2023, applicable to coverage starting January 1, 2024, that governs how Medicare Advantage plans set prior-authorization and coverage criteria. In February 2024, CMS issued a follow-up FAQ specifically addressing AI: insurers may use algorithms to help make coverage determinations, but an algorithm cannot be the sole basis for a denial. Every determination has to reflect the individual patient's condition, their treating physician's recommendation, and their clinical notes — not just where they land in a larger dataset of similar patients.

That's a real federal standard, not a suggestion. It's also not self-enforcing. Nothing in the rule automatically catches an insurer that runs the algorithm and treats its output as the individualized determination in practice. That gap is exactly what the Lokken and Cigna lawsuits are arguing happened — and what makes the litigation, not just the rule, the thing actually testing whether insurers comply.

A State Went Further Than the Federal Rule

CMS's rule only reaches Medicare Advantage. California passed a law aimed at the rest of the market. Senate Bill 1120 — titled the Physicians Make Decisions Act, sponsored by the California Medical Association on behalf of the state's 50,000 physicians — took effect January 1, 2025, and applies to health plans and disability insurers doing business in the state generally, not just Medicare's private-plan carve-out. It bars companies from denying, delaying, or modifying care based on medical necessity using AI alone, and requires that any such decision be made by a licensed physician or another qualified healthcare provider with expertise in the specific clinical question — not just a person signing off on whatever the software already decided.

The bill's own author put the reasoning in blunt terms: "An algorithm cannot fully understand a patient's unique medical history or needs, and its misuse can lead to devastating consequences." What makes California's version different from the federal FAQ isn't the standard itself — CMS's language about individualized review reads similarly — it's who enforces it. The state's Department of Managed Health Care and Department of Insurance can levy administrative penalties directly, without a patient having to file a federal lawsuit first. Whether that authority gets used against a major insurer is still an open question a year into the law; the mechanism to use it, at least, now exists in a way it didn't before 2025.

What Happens If You Go Over the Insurer's Head

The KFF numbers below cover appeals filed with the insurer itself — the same company that denied the claim reviewing its own decision. There's a further step, required nationwide under the ACA for medical-necessity disputes: an independent external review, conducted by an outside organization with no financial stake in the outcome. A Health Affairs analysis of denials that reached independent medical review across four states from 2019 through 2023 found that almost half were overturned.

Independent review isn't free of friction either — it comes after the internal appeal already failed, requires a patient to know the option exists, and still runs on the insurer's own coding of what happened. But the overturn rate at that stage tells a consistent story with everything else in this piece: denials that get a genuinely outside look, whether by an independent medical reviewer or a court examining an algorithm's error rate under oath, fail at a far higher rate than the insurer's own paperwork would suggest. The reviewers aren't the bottleneck. Reaching them is.

The Appeals Nobody Files

Zoom out from Medicare Advantage and post-acute care to the broader insurance market, and the same shape of math shows up again. KFF's analysis of 2024 ACA marketplace plans found insurers denied 19% of in-network claims and 37% of out-of-network claims — a combined 20% of everything submitted.

Two different numbers, two different insurance markets, one identical structure: the overturn rate for people who actually appeal is meaningfully higher than the denial rate would suggest is fair — 34% in the general marketplace data, close to 90% in the specific nH Predict cases cited in the Lokken complaint. But an overturn rate only helps the person who files. Fewer than one in a hundred people denied a marketplace claim do. The rest of the math never gets a chance to run.

And Then You're Dead

None of this requires a conspiracy. It requires a denial that's cheap to issue, an appeal that's expensive to file, and a patient too sick, too old, or too exhausted to do the paperwork. An algorithm that's wrong nine times out of ten is still profitable if only two in a thousand people ever find out. That's not a bug in how automated claims review got built. Court filings, a Senate investigation, and a federal agency's own rulemaking all now say, in their own language, that it might be the design.

That's it. That's the whole thing.

Read about having no coverage at all · or the debt left over either way · or see every essay