Jacob Nardini
← All articles

Denials Are Approaching 20%. Most Are Never Appealed.

May 2026 · Healthcare

The initial claim denial rate across the largest provider-side benchmarks climbed to roughly 12% in 2024, and it's still rising. In some environments it's far worse: an analysis of federal marketplace data found insurers denied about 19% of in-network claims in 2024 — nearly one in five, the highest in years. Behind those averages are categories that run hotter still, with requests-for-information denials on commercial plans north of 30%.

Now hold that next to a stranger number. When patients on those marketplace plans were denied, they appealed less than 1% of the time. Provider revenue cycle teams do better than that, but not nearly as well as they should — not because the appeals would lose, but because nobody has the hours. For a mid-market health system or physician group, that's a large, recoverable revenue stream being quietly written off for lack of capacity. And the part that should genuinely irritate any CFO: a great many of those denials are wrong.

Why denials are the cleanest ROI story in healthcare AI

Most healthcare AI pitches measure value in saved hours, and finance teams are right to discount them. Denials work is different, and that difference is the whole reason I point revenue cycle leaders here first: the ROI is denominated in dollars, in numbers the CFO already tracks. Overturn rate, days in accounts receivable, cost to collect, recovered dollars — when an automated appeals workflow recovers claims that previously went unappealed, the value shows up on the income statement, not in a time study.

The scale of the leakage is no longer in dispute. On the largest hospital benchmark, providers lost more than $48 billion in net revenue to final denials and bad debt in 2025 — up roughly 25% in a single year — as the median final-denial rate ticked from 2.5% to 2.7%. That 0.2 points sounds trivial until you remember it's a percentage of every dollar billed. Providers also spend enormous sums fighting denials in the first place: industry surveys put the cost of reworking a single denied claim around $44, and the national tab for adjudicating disputes with payers in the tens of billions — more than half of it spent on claims that should have been paid on first submission. The administrative tax of chasing money you already earned is, itself, one of the largest controllable costs in the building.

The injustice that is also the opportunity

Here's the fact that turns this from a cost-cutting story into a revenue story. When denials are appealed, they're overturned at remarkable rates. A national provider survey found that more than half of private-payer denials — about 54% — were ultimately overturned and paid. An independent study of New York appeals tracked the overturn rate climbing from 38% in 2019 toward 53% by 2025. On Medicare Advantage, overturn rates on appeal have run north of 80%.

Read that plainly: payers are denying enormous volumes of claims they end up paying anyway, once someone pushes back. The denial is often not a considered coverage judgment — it's a filter, and it works because most providers don't have the staff to contest it. That's been made explicit by the controversy over payers' own automation: high-profile litigation has alleged that algorithmic tools were used to deny post-acute and other care at scale, with the denials reversed on appeal the overwhelming majority of the time. Whatever those cases ultimately find, the strategic lesson for a provider is unambiguous. This is an arms race, the other side has automated, and the response can't be a bigger pile of fax cover sheets.

The anatomy of an appeals pipeline

A production denials workflow has four moving parts, and each plays to a genuine strength of modern document AI rather than to a chatbot fantasy:

The economics underneath are stark. Industry data on adjacent transactions shows manual administrative work running a few dollars per touch versus pennies when automated — better than a 90% reduction — and the same logic applies to the research and drafting that swallow a denials analyst's day. The point isn't to remove the analyst; it's to let one analyst work the volume that three used to, and to make sure the denials that are worth fighting actually get fought before the deadline passes.

Not every denial is worth the same fight

A good program doesn't appeal everything; it appeals intelligently, and that starts with knowing your denials. They cluster into a handful of categories, each tied to standardized reason codes on the remittance: prior authorization and precertification, eligibility and coverage, medical necessity, missing or invalid information, and fee-schedule or non-covered-service denials. Those categories behave very differently. A missing-information denial is often a fast, near-automatic resubmission. A medical-necessity denial is a genuine argument that needs clinical evidence. A fee-schedule "denial" may not be a denial at all so much as an underpayment in disguise.

The reason the codes matter is that they're machine-readable — the industry runs on a few hundred standardized claim-adjustment reason codes and well over a thousand remittance remark codes — so an extraction layer can sort the queue by category, dollar value, payer behavior, and time-to-deadline before a human touches it. That sorting is where the leverage is, because the denials picture has shifted. On the largest benchmarks, essentially all of the recent rise in denial rates traces to clinical denials — prior authorization and medical necessity — not the technical, administrative kind. Those are the higher-dollar, higher-overturn fights, and they're exactly the ones a stretched team tends to triage last because they take the most work. An AI layer that assembles the clinical evidence and drafts the argument flips that calculus: the fights most worth having stop being the ones that get dropped at the deadline.

The trap to avoid

Stay on the assistive side of the line, deliberately and visibly. Drafting an appeal that a human reviews and signs is a workflow improvement. Anything that touches a coverage or clinical decision automatically walks into the exact regulatory minefield the payers are getting sued over. That's not a hypothetical: under recent CMS rules, a Medicare Advantage plan can't deny a claim on medical-necessity grounds without a licensed clinician's review, and states have begun codifying that a human — not an algorithm alone — must make the call. CMS is even piloting AI-assisted prior authorization in original Medicare with mandatory human clinical review built in.

The lesson cuts cleanly in the provider's favor. On the provider side, you are assembling evidence, drafting arguments, and prioritizing work — all squarely defensible, all human-supervised. You are not making coverage determinations. Keeping that boundary bright isn't only safer; it's an easier sale to a cautious compliance officer, and in healthcare, compliance is always in the room. Build the human-in-the-loop framing in from day one and it becomes a feature you can point to, not a constraint you have to apologize for.

From recovery to prevention

Recovering denied revenue is the obvious win, but the same infrastructure points at a bigger one: not getting denied in the first place. Every denial that's later overturned represents a claim that should have gone out clean, and the round trip is expensive in ways that don't show up as a single line item. Denied claims sit in accounts receivable while they're worked; receivables that age past 90 days collect at a steep discount to fresh claims, and cost-to-collect — which the best performers hold at or below 2% of net collections — balloons when staff spend their days reworking instead of collecting. The CFO-facing version of the story is that denials don't just cost you the occasional written-off claim; they inflate A/R days and cost-to-collect across the entire book.

And the data you build to fight denials is the same data that prevents them. Once you can see overturn rates and denial reasons by payer and by code, the patterns surface: this payer denies this procedure for documentation four times out of five; that one's eligibility denials spike every January. Feed that intelligence back upstream — into front-end eligibility checks, documentation prompts at the point of order, and cleaner first submissions — and the denial volume itself starts to shrink. The mature version of a denials program is half recovery and half prevention, and the prevention half is where the durable margin lives, because a claim that never denies costs nothing to collect and never ages a day.

The mid-market opening

The reason this is a live opportunity and not a closed one comes down to who has actually deployed AI. The headline that "most hospitals haven't adopted AI" is stale — well over two-thirds of hospitals now use some predictive AI in their EHRs. The real story is the divide underneath it: adoption is heavily concentrated in large, system-affiliated hospitals, while independent and mid-market provider groups lag far behind, by some measures more than two to one. That's precisely the segment a focused denials program serves, and precisely why moving now isn't a technology gamble.

Large systems already have revenue-integrity teams and data scientists. A mid-market system or physician group usually doesn't — which means the organizations that move first on denials aren't taking a frontier risk. They're collecting revenue they already earned while their peers keep writing it off, using technology that's well past the experimental stage. The constraint was never the model. It was having someone scope it to your payer mix, wire it to your remittance data, and earn your denials team's trust.

Where to start

Run a retrospective. Take a representative slice of the denials your team worked over the last year — and, just as importantly, the ones it didn't get to — and put them through an extraction-and-drafting pipeline. Then measure the only things that matter: how accurately it identified the denial reason and the deadline, how often the drafted appeal needed material correction before a human would sign it, and what the recovered-dollar opportunity looks like on the claims that currently age out unworked.

If the numbers don't clear your cost of collection, you've learned that cheaply. But denials work makes its case in dollars on claims you already have records for, against a payer behavior pattern that's only intensifying. That tends to survive the budget scrutiny that kills vaguer AI initiatives — because for once, the business case isn't a promise about productivity. It's revenue you earned, sitting in a queue, waiting for someone to go get it.