Denied in 1.2 Seconds: How Payers Automate Claim Denials with AI, and the Defense a Practice Can Build

The Sixty-Second Version

In 2022, one of the country’s largest insurers ran claims through an automated review system. In two months it denied more than 300,000 of them. The doctors signing those denials spent an average of 1.2 seconds on each one. ProPublica reported it in March 2023. The figure now sits in a federal lawsuit, and the court has already let the core claims move forward.

Maybe you’ve felt the shift. Denials arrive faster than they used to. The letters read generic. No reviewer is named. You are not imagining it. Claim review moved from people to machines, and most practices still answer machine output by hand.

The defense has three parts. First, read the engine’s patterns in your own claim history. A machine that denies by rule leaves the rule in your data. Second, stop the claims it always kills before they go out the door. Third, hold every payer to the new rules that say a named human must own the denial.

This guide shows what changed, with sources. Then it walks the five builds that turn a practice from target into opponent. Each build ends with a check you can run on your own data this week.

The Clock Changed

For most of the history of medical billing, a denied claim meant a person looked at it. Maybe briefly. Maybe unfairly. But a human opened the file, applied a policy, and made a call. Everything ran at human speed. Your practice and the payer were on the same clock.

That clock is gone. Large payers now push claims and prior-auth requests through automated review at volumes no staff could touch. The system flags mismatches between codes, diagnoses, and preset rules in seconds.

Is it “artificial intelligence”? It doesn’t matter. Lawyers who work these cases point out that much of the damage comes from plain rules engines, simple logic run at enormous scale. The label is a distraction. The clock speed is the event.

You can usually feel it before you can prove it. The denial comes back fast, sometimes too fast for a file a clinician supposedly read. The language echoes your codes, never your notes. No reviewer signs it. One of those could be a coincidence. All three, across a quarter of denials, is a pipeline.

The profession has noticed. The American Medical Association surveyed a thousand physicians in December 2024. 61% said they worry that health plan AI is driving prior-auth denials up. 75% said denials have risen over the past five years. The average physician now spends 13 hours a week on prior auth, four in ten practices employ staff who do nothing else, and 93% of physicians said the process delays care.

That’s what doctors report. Here’s what’s documented.

The Record

Everything below is public: news investigations, court filings, rulings, and government documents. Where a company disputes the story, its side is included. The fact that matters to your practice survives every rebuttal.

Cigna and PxDx

In March 2023, ProPublica reported that Cigna used a system called PxDx to check claims against preset rules and reject the mismatches in bulk. Staff doctors signed off in batches. Over two months in 2022, by ProPublica’s count, the system denied more than 300,000 requests for payment. The reviewing doctors averaged 1.2 seconds per claim.

A class action followed, Kisting-Leung v. Cigna. The plan documents promised review by a medical director. The suit says handing that decision to a machine broke the promise. Cigna argued it was within its rights. The court disagreed, called Cigna’s reading of its own plan an abuse of discretion, and let the core claims continue.

Cigna’s side: the company says PxDx is not AI, just a sorting technology the industry has used for years. It says the system covers about fifty low-cost tests and procedures, and that most claims it reviews get paid automatically. As of early 2026, Cigna had not said it retired the system.

UnitedHealth and nH Predict

In November 2023, families of two deceased Medicare Advantage members sued UnitedHealth Group over nH Predict, an algorithm that projects how long post-acute care should last. The complaint draws on a STAT News investigation. It says employees were expected to keep nursing stays within 1% of the model’s projection, and were “disciplined and terminated” for straying. It also says the model’s calls were overturned at very high rates when patients managed to appeal.

UnitedHealth disputes all of it. In February 2025, the court threw out five of seven counts and kept the rest alive. For an algorithm case, that’s further than most expected it to get.

Beyond two companies

Humana faced a parallel suit over the same model in late 2023. In 2022, the HHS Office of Inspector General found Medicare Advantage plans denying prior-auth requests for services that met Medicare’s own coverage rules. That report is part of why Washington started paying attention.

Now the government runs its own version. On January 1, 2026, CMS launched the WISeR model, an algorithm-assisted prior-auth pilot in New Jersey, Ohio, Oklahoma, Texas, Arizona, and Washington. Providers there must get prior auth for the targeted services or face pre-payment review. CMS says licensed clinicians make every non-payment call. In March 2026, the Electronic Frontier Foundation sued CMS over an unanswered records request and says the model’s denial rates run too high. That fight is live. The direction of travel is not.

Now put that record next to how a practice answers it.

The Asymmetry, and the Gift

Here is the fight as it stands. The payer’s engine reviews thousands of claims at once. Its rules update quietly, in one place. Each decision costs the payer almost nothing, and your appeal is a cost it already planned for. Your practice works the other way. Claims go out one at a time, denials come back one at a time, and appeals happen when somebody has the hours. The AMA’s 13-hours-a-week number is what that feels like from the inside.

You cannot out-staff this. Another biller adds addition. Their engine added a zero.

And every year the gap grows, and every year it looks like your team slipping. Sloppier claims. Slower follow-up. A billing department that used to keep up and somehow can’t. That reading is wrong. Your people didn’t slow down. The other side changed clock speed, and nobody sent your practice the memo.

Now the part the doom coverage always misses.

A machine that denies by rule cannot help but reveal the rule. People are inconsistent, so human review scatters. A rules engine does the same thing every time the same conditions appear. Repetition is exactly what grouping exposes. Every denial the engine sends you is a logged output of its own logic. Your claim history is a running record of the machine’s behavior, aimed at your practice, sitting in a file you already own. Group it by payer and procedure, and the engine’s preferences turn into a ranked table.

We ran that table. Five months of remittance data, one multi-provider practice. More than 3,000 denied lines spread across nearly 280 payer-and-code pairings. The top ten pairings held close to 80 percent of every denial. The worst single pairing held more than half, by itself.

One payer in this dataset sells a simple promise to its members: no deductible, no copay. That promise has a cost, and the cost lands on the provider side of the ledger. A plan with no member cost-sharing pays the full allowed amount on every visit. It has one big lever left to control what it spends. That lever is deciding claims.

Across those five months, this payer denied at a rate the rest of the payer mix never approached. Almost all of it hit a single add-on code. More than 90 percent of everything billed on that code came back denied, over $230,000 in charges, with the same stated reason on line after line: the documentation does not support the service.

The zero-pay lines tell the rest. Over those five months, this payer zeroed out more than $285,000 in billed charges, more lines than every other payer combined. The rest of the payer mix zeroed out a nearly identical amount, which makes the comparison unusually clean. On the rest of the mix, more than 40 percent of those dollars moved to patients as deductible, copay, or coinsurance, where they stayed collectable. More than half the lines kept some patient balance. On this payer, less than 7 percent moved, and only one line in seventeen kept any patient balance at all. The rest, more than a quarter million dollars, was never paid by the plan and can’t be billed to the member. The member’s free care was not free. The providers financed it.

Run the scoreboard. The practice is out more than a quarter million dollars, money the plan didn’t pay and the rules say the member can’t be billed for. The member got care and paid nothing, and in sixteen of every seventeen zero-paid lines, no bill ever existed. From inside the plan, nothing happened. The member has no idea the visit was denied. The denial was invisible to the one person whose plan issued it.

And the payer? The premium was collected months before the visit. The claims expense was avoided at the moment of decision. The brand promise, free care, stays intact because the cost of keeping it landed on someone with no seat at the table. Payers call this kind of denial documentation integrity, a case-by-case judgment that the record didn’t support the service. A judgment that lands on line after line of a single code, at a rate above nine in ten, is a policy applied at scale.

The losses land at different speeds, and that gap is the whole game. The practice’s loss books now, on one balance sheet. The member’s loss comes later, when practices quietly stop taking that plan, and it comes so slowly that nobody connects it back to the denials. The payer’s risk, a network adequacy complaint to a regulator someday, comes last and lands softest. Every month those speeds stay different, one side collects.

Here is the gift inside the bad news. The payer’s automation is a one-way mirror only as long as you never group your denials. The moment you do, the mirror works both ways. There is one dataset the engine can never see better than you can: the complete record of what it has done to your practice.

Reading the pattern is step one. Here is the build list.

The Defense: Five Builds

Every build has the same shape. Make the machine’s behavior visible. Decide in advance who acts when it shows up. Let your own system do the watching, so the defense survives turnover and busy seasons. None of this needs new software or a data hire. It needs the claim history you already have, arranged to answer questions nobody assigned.

One boundary first. Everything here happens before and around the claim, upstream of claims follow-up. Your billing team, and your RCM service if you use one, own the claim once it exists: submission, correction, appeal, aging. They’re good at that, and nothing here replaces an hour of it. These builds work the territory that was never anyone’s job, which is exactly why the engine has had it uncontested.

Build One: Pairing Intelligence

The mechanism. Automated denial concentrates. Engines apply rules to combinations: this payer plus this code, sometimes plus a diagnosis or modifier. A combination that trips the logic trips it every time. Spread across months of denials worked one at a time, the concentration is invisible. Computed, it’s a list.

What to build. A running denial rate for every payer-and-code pairing in your book, from your own claim history, over the trailing six to twelve months. Denials divided by submissions, ranked. Version one is a spreadsheet. The durable version refreshes itself from your system’s data. If your practice runs AdvancedMD, every record this build needs is already in your claim history.

Check yours this week. Export ninety days of denials. Group by payer and code, sort by count, and divide denials by submissions for the top rows. The number at the top of that table is a fact about your practice. The payer’s engine has known it for years. You’re seeing it for the first time.

Build Two: Pre-Submission Gates

The mechanism. Once you can see which pairings die, sending them out unchanged stops being billing and starts being donation. Every doomed claim costs staff time going out, rework coming back, and weeks of delay on money that was never coming by that route.

What to build. A threshold and a routing rule. A pairing over the line gets held before submission and routed by cause. Missing auth? Prior-auth step. Code and note mismatch? Documentation review. The payer’s policy itself? Contract conversation. The gate ends the ritual of resubmitting dead claims, and turns each one into the action that can actually pay.

Check yours this week. Take the worst pairing from Build One. Hold this week’s instances for a ten-minute review before they go out. Log what you catch. One pairing, one week, usually settles whether the gate earns a permanent place.

Build Three: Complete Documentation Before the Charge Exists

The mechanism. Audit engines feed on gaps between what was billed and what was written down. An unsigned note. A missing note. A code the note doesn’t support. These are the cheapest kills an automated auditor has, because a machine can spot the mismatch and the denial is nearly appeal-proof. Loose documentation leaks revenue on its own, and it hands the engines their favorite target at the same time.

What to build. A completeness gate of your own: no charge exists until a finished, signed note exists behind it. Enforce it with daily visibility, never with asking clinicians to remember. We walk this build in the revenue integrity guide, where it serves the internal chain. Here it doubles as armor. Every encounter that reaches the payer fully documented is one less free denial you hand the machine.

Check yours this week. Pull your last ten denials that cite documentation, medical records, or coding support. For each one, check whether the note was complete and signed when the claim went out. The ones that weren’t are the denials you built yourself.

Build Four: The Payer Ledger

The mechanism. Payers keep detailed files on your behavior. Almost no practice keeps a file on theirs. Denial rates over time. How long they take to respond. Every downcode. Every dispute and how it ended. One at a time, these are annoyances, worked and forgotten. Accumulated, they’re evidence. Appeals and contract negotiations run on evidence.

What to build. One file per payer, fed automatically from your claims data: monthly denial rate, days to respond, every downcode with the original and altered code, every dispute with its category and outcome. Two uses. In appeals, you stop arguing single claims and attach the pattern instead, which changes the conversation. In renegotiation, you walk in with a year of the payer’s own behavior, dated and counted, next to a contract that promised something else.

The zero-deductible payer from the Asymmetry section is the argument for this build. A ledger turns that pattern into dated, counted evidence the week it forms, instead of a discovery someone pieces together months later.

Check yours this week. Count last quarter’s downcoded claims from your largest payer. If getting that count takes more than ten minutes, the ten minutes are the finding. They have a complete file on you. You have nothing on them.

Build Five: Drift Watch

The mechanism. Engines get updated. Rules tighten, code policies shift, and a pairing that cleared for two years starts dying in a month. Nothing announces the change except your denial letters. A defense built once and left alone is a snapshot of last year’s war.

What to build. Build One, on a schedule. The pairing table refreshes monthly. Movement past a set band, a rate jumping or a clean pairing entering the top ten, fires an alert to a named owner with a pre-decided next step. The machine watches. The person acts. Payer changes stop arriving as a bad quarter you diagnose afterward and start arriving as a flagged row you handle the week it moves.

Check yours this week. Compare this quarter’s top denial pairing to the same quarter last year. If the top of the table changed and nobody can say when or why, drift has been running unwatched.

Every build gets stronger next to what changed in the rulebook this year, because for the first time in this fight, the rules moved toward the practice.

The Rules: What the Law Now Hands You

Those lawsuits set off a regulatory wave. A practice that knows the current rules holds appeal ammunition that didn’t exist when the engines were built. None of this is legal advice. It’s operational use of public rules, and your attorney owns the legal strategy.

Federal, part one. CMS’s 2024 Medicare Advantage rule says medical-necessity decisions must rest on the individual patient and be reviewed by a physician or other qualified professional. A February 2024 CMS FAQ went further: an algorithm alone cannot be the basis for denying admission or ending services.

Federal, part two. A separate CMS rule took effect January 1, 2026. Medicare Advantage denial notices must now cite the specific, current clinical criteria behind the decision. A denial letter with generic language where the rule demands specifics has handed you a deficiency to name in the appeal.

Federal, part three. If you practice in New Jersey, Ohio, Oklahoma, Texas, Arizona, or Washington, the WISeR pilot is part of your reality now. Know its targeted-service list, because the choice it forces is prior auth up front or pre-payment review after.

States. Legislatures moved fast once the record went public. One 2026 tracker counts 134 bills across 37 states, and about 25 states have adopted regulator guidance based on the NAIC’s model bulletin. The enacted laws most useful to a practice:

Laws move fast. Confirm your state’s current status before citing one in writing.

Using it in an appeal. Three moves turn a statute into pressure. Name the law, and state plainly that a denial issued without documented clinician review may not comply with it. Ask in writing for the name and credentials of the clinician who reviewed the denial; most rules require the insurer to answer. Then demand the specific clinical criteria the denial rests on, which the 2026 federal notice rule requires and which generic language fails on its face. Appeals built this way succeed at meaningful rates. Most denials are never appealed at all, and the engines are tuned to that silence.

The two-way street. One law cuts both directions. Indiana’s HB 1271 also bars providers from submitting AI-built claims without review by a person involved in the claim. The wave that started with payer automation is reaching provider automation. The posture it demands, a machine may assemble but a person must own, is how a well-defended practice should run anyway. Automation with no named human owner is the exact design that got the payers sued.

What Defended Looks Like

A defended practice fits in four sentences. The pairing table refreshes monthly, and somebody owns every row that moves. Claims in gated pairings stop at the door and leave through the route that pays. Every encounter reaches the payer complete, signed, and supported, so the cheap kills are gone. And every payer’s behavior piles up in a ledger that turns appeals into pattern arguments and renegotiations into evidence reviews.

The numbers follow. Clean claims hold above 95%. Denial work shrinks from a volume problem to a short exception list. Monthly revenue stops being a surprise and becomes a figure you can state in advance, because nothing between the visit and the deposit moves unwatched, in either direction.

Underneath it all sits one fact worth keeping. This fight is dataset against dataset. About the world, the payer’s dataset dwarfs yours. About your practice, yours is better, complete, and already paid for. The engine sees you as a few thousand rows in a book of business. You can see every decision it has ever made against you. The payers industrialized their half of the argument. Nothing stops a practice from industrializing its own.

And whoever builds this for you owes you three certainties, the same three we hold ourselves to. Found it, with records you can verify line by line. Fixed it, with the dollar figure that moved. Stays fixed, with the watch that catches the drift.

Where to Start

You don’t need a project to find out where you stand. You need one look at the right cut of your own data.

Book a 30-minute look. Bring ninety days of denials, or just bring access and we’ll pull them live. You’ll leave with your three worst payer-and-code pairings, named, and what each one costs you. Want to start smaller? The Practice Cash Scorecard takes three minutes and shows you where you stand against the standard.

This page is for information only and is not legal advice. Consult your attorney before citing statutes or regulations in payer correspondence.

PracticePath is not affiliated with, endorsed by, or sponsored by AdvancedMD. AdvancedMD® is a registered trademark of Global Payments. All references to AdvancedMD are for informational purposes and to identify the software environment our services support.