How Payers Built AI to Deny Claims and What Happens Next

For most of the last fifty years, a denied health insurance claim was, somewhere along the way, the product of human judgment. A nurse reviewer or medical director read the documentation, applied the plan’s coverage rules, and made a call. The decision could be wrong, slow, or frustrating, but it was made by someone who had at least looked at the file.

That is not how a growing share of denials work anymore.

What payers actually built

Over the past decade, the largest US health insurers operating in Medicare Advantage, Medicaid managed care, and commercial markets have invested heavily in algorithmic and AI-driven systems that sit in front of the claims-review process. These systems take different shapes, but the architecture is similar in spirit.

At the simplest end are rule-based engines that compare the procedure code on a claim against the diagnosis code on the same claim, against the insurer’s internal list of “acceptable” pairings. If the combination falls outside the list, the claim is flagged for denial in batches and a clinician signs off on hundreds or thousands of those flags at a time, often without opening individual patient files.

At the more sophisticated end are predictive models trained on millions of historical patient records. These models score new requests against patterns in the training data, predicting for example how many days of skilled nursing care a patient with a given diagnosis, age, and functional status “should” need. When the patient hits the predicted threshold, the model triggers a coverage cutoff. The clinical team is then expected to keep length-of-stay within a narrow band of what the algorithm forecast.

What sits underneath both approaches is the same operational logic: a human reviewer is expensive, slow, and inconsistent. An algorithm is none of those things. If the regulatory framework allows a model to do the first pass and in most US payer markets it does, then the model does the first pass.

Why this changes what “denial” means

The most consequential shift isn’t technical. It’s that the unit economics of denying a claim have changed.

When a human reviewer denies a claim, the cost to the payer is the reviewer’s time. That sets a natural ceiling on how many denials a system can produce. When an algorithm denies a claim, the cost is essentially zero, and the ceiling is gone. Denials can now be issued at a rate that matches the rate at which claims arrive.

That matters because of a second, well-documented asymmetry: most denied claims are never appealed. The most recent KFF analysis of CMS data on Medicare Advantage shows roughly 90% of denied prior authorization requests are not appealed by the patient or provider. Of the small share that are appealed, more than eight in ten are partially or fully overturned.

Read those two numbers together and the picture is clear. The system isn’t engineered to be right — it’s engineered to be cheap to be wrong. When wrongful denials cost the payer almost nothing because patients don’t fight them, and when the cost of issuing those denials has dropped to near zero through automation, the rational behavior for a profit-maximizing payer is to deny more, not fewer.

What the next few years look like

Three trends are worth watching.

First, the practice is spreading from Medicare Advantage into traditional Medicare. Until recently, fee-for-service Medicare was largely insulated from this kind of utilization management. That’s changing. CMS launched a six-year pilot in January 2026 that brings AI-assisted prior authorization to a defined set of Part B services in six states. The official intent is to reduce waste and fraud. The operational reality is that traditional Medicare beneficiaries, people who specifically chose original Medicare to avoid managed-care gatekeeping are now subject to a version of it.

Second, regulators are responding to outputs, not models. The federal rules that took effect at the start of 2026 require payers to issue prior authorization decisions faster (72 hours for expedited requests, 7 days for standard), explain denials more clearly, and publish aggregate approval data. These are real improvements. But none of them regulates the internal models that generate the denial in the first place. A payer that uses an opaque AI system to deny a claim in 1.2 seconds is now required to do so within 72 hours and to explain its reasoning, which it can do with a generated paragraph. The asymmetry survives.

Third, the burden is migrating onto providers. Every denial that an algorithm issues becomes a packet that a clinic has to assemble, an appeal that a billing team has to write, a deadline that a practice manager has to track. The administrative cost of running an independent clinic in the US has been climbing for years; payer automation accelerates that curve. For specialty practices in fields like orthotics and prosthetics, behavioral health, and post-acute care where prior authorization volumes are highest, this is starting to determine which practices survive and which don’t.

The long-term effect

The long-term effect of payer AI is not, in our view, a sudden collapse of patient access. It’s something slower and harder to see: a gradual reweighting of who absorbs the cost of insurance.

When denials are cheap to issue and expensive to fight, the cost of an insurance company being wrong gets pushed downstream. Patients absorb it in out-of-pocket payments for care they thought was covered, or in foregone care they couldn’t fight for. Providers absorb it in administrative labor, billing staff, appeal writers, deadline trackers that doesn’t show up in any clinical metric but shows up clearly on the P&L. The insurance company keeps the difference.

The honest framing is that AI didn’t create this dynamic. The asymmetry between cheap denials and expensive appeals has existed for as long as managed care has existed. What AI did was remove the last remaining constraint on how aggressively a payer could exploit it. That’s the technology story. The policy and provider response is what comes next.

Where this leaves clinics

There are exactly two ways a provider can respond to a denial system that’s getting faster, cheaper, and more aggressive on the payer side.

The first is to absorb more of the cost, hire more billing staff, write more appeals, lose more revenue to write-offs. This is the path most clinics are on by default. It is not sustainable for independent practices.

The second is to bring the same technology onto the provider side of the wire. Rule-driven validation that reads every clinical packet against the actual payer coverage policy before it leaves the office. Denial decoders that map a CARC code or non-affirmation letter back to the specific policy section that was cited, so staff isn’t guessing at what to fix. Deadline trackers that don’t let a Level 1 redetermination clock run out because a spreadsheet wasn’t updated.

This is the side of the problem we work on at Everi Labs. The asymmetry that defines payer-provider economics in the US is now a software problem on both sides of the wire. The clinics that come through the next decade intact will be the ones that treat it that way.