Why claim denials are a $20 billion problem you can't outsource away

Healthcare providers write off, chase, or slowly recover somewhere north of $20 billion a year because payers deny claims that were often payable in the first place, according to the American Hospital Association. The AHA found 15.7% of Medicare Advantage claims and 13.9% of commercial claims were initially denied. Aptarro's 2025 analysis of Crowe RCA data puts the industry-wide initial denial rate at 11.81% in 2024, up from roughly 10.2% the year before. The trend line only moves one direction.

Here's the detail almost nobody leads with: 54.3% of denied claims get overturned and paid once someone appeals them, per the same AHA report. That number changes what kind of problem this is. Denial management isn't mostly a lost-revenue problem. It's a bandwidth problem. More than half the money is recoverable, and the bottleneck is how fast a human biller can read a denial code, pull the chart, write an appeal, and resubmit before the payer's filing window closes. That's a workflow you can speed up with the right system. It's also exactly the kind of workflow that generic RCM software and Big 4 advisory decks gloss over, because "buy our platform" is a much easier pitch than "your denial patterns are specific enough that no off-the-shelf module will learn them for you."

What is denial management in healthcare billing?

Denial management is the set of processes a provider uses to identify why a claim was denied, decide whether it's worth appealing, correct or document the issue, resubmit within the payer's deadline, and track the outcome. It sits downstream of claim submission and upstream of final adjudication: prior authorization happens before a service is rendered, denial management happens after the claim is filed and rejected. If you're evaluating the pre-service side of this same lifecycle, that's a separate build decision, one we've covered in detail here given the CMS 2027 interoperability deadline reshaping that workflow.

Denial management is often confused with denial prevention, and the distinction matters for where you put your budget. Prevention means catching eligibility problems, missing modifiers, or coding mismatches before the claim ever goes out, the goal is fewer denials. Management means working the denials you already have, the goal is recovering money that's already at risk. Most organizations need both, but they're different systems solving different failure points, and a vendor selling you a "denial management platform" that's really an eligibility-checker at the front end is solving a different problem than the one you called about.

Context for scale: KFF found HealthCare.gov marketplace plans denied 19% of in-network claims and 37% of out-of-network claims in 2024, averaging about 20% overall. AJMC reported a similar 20% denial rate on qualified health plans in 2023, with documented harm to patients whose claims go unappealed, per their February 2025 analysis. A "healthy" clean-claims denial rate benchmark sits at 5 to 10%. If your group is running above that, you're not unusual. You're average, and average is expensive.

Where off-the-shelf RCM software and denial modules stop working

Every enterprise RCM platform, Waystar, athenahealth, Epic's own tooling, ships a denial management module, and every one of them works reasonably well for the denial patterns common across their whole customer base: eligibility mismatches, missing authorizations, duplicate claims. What they don't do well is learn your payer mix. A multi-specialty group billing a regional Blue Cross plan, three Medicare Advantage carriers, and a handful of commercial payers accumulates denial codes, appeal letter formats, and payer-specific quirks that are genuinely idiosyncratic. Enterprise tooling is built to generalize. Your denial patterns don't generalize; they're a function of your specialty mix, your coding habits, and which claims examiners at which payers have which pet peeves this quarter.

This is the gap the current search results for denial management automation quietly skip over. The hospital associations and Big 4 firms correctly frame the size of the problem. The RCM vendors correctly pitch automation as the fix. None of them say the thing that's actually true for a $5M to $50M multi-specialty group, ASC, or physician practice management company: your volume is often too specific for a generic module to prioritize learning your patterns, but too small to justify Epic- or athenahealth-scale enterprise tooling built for hospital systems processing millions of claims a year. You're stuck in the middle, and the middle is where "just buy the add-on module" quietly stops delivering.

We've seen the same structural pattern in accounts receivable automation, where off-the-shelf AR tools work fine until billing complexity (multiple entities, nonstandard terms, exception-heavy collections) outgrows what a generic rules engine can encode. We wrote about that pattern in more depth in our AR build-vs-buy guide, and the diagnostic questions are nearly identical here: how much of your workflow is genuinely standard, and how much is specific enough to your organization that a system has to be trained on your own history to be useful at all.

Build, buy, or augment: how to actually decide

Start with three numbers before you talk to a single vendor: your monthly denial volume, the number of distinct payer-denial-code combinations you see regularly, and your current appeal win rate. Those three numbers tell you almost everything.

• Buy a bolt-on module if your denial volume is low, your payer mix is concentrated in two or three plans, and your denial codes cluster around a handful of repeat issues (eligibility, missing modifiers). A configured off-the-shelf tool will pay for itself faster than a custom build here.

• Hire more billers if your denial volume is genuinely low and your appeal win rate is already strong. Sometimes the honest answer is that automation isn't the bottleneck, headcount is, and no software fixes a staffing gap.

• Build or augment with a custom AI agent if your denial volume is high enough that billers can't keep pace with appeal deadlines, your payer mix is fragmented across five or more plans with different documentation requirements, and your win-rate-on-appeal is meaningfully below that 54.3% industry average, which usually signals appeals aren't getting filed in time, not that they're unwinnable.

This is the same diagnose-before-you-build logic we apply across every RCM engagement: figure out where the actual leakage is happening before deciding what to automate. At a utility client, we found that most of a $1M+ annual revenue leak in billing wasn't an AI problem at all, it was broken process and disconnected systems, and fixing it required six discrete projects, most of them plain automation and integration work, not machine learning. The lesson carries over directly: denial management often looks like an AI problem from the outside and turns out to be partly a workflow-sequencing problem once you actually open it up.

What a custom AI denial-management agent actually does

In practice, a working denial agent does four things, and none of them require replacing your billing staff.

First, it categorizes denials by root cause automatically, reading the remittance advice and denial code, pulling the relevant claim and clinical documentation, and sorting denials into buckets: missing authorization, coding mismatch, medical necessity, eligibility, timely filing. Second, it learns your specific payer patterns over time, which carriers reject which code combinations, which documentation formats satisfy which plans on appeal, so the system gets sharper at your organization's actual denial mix rather than a generalized industry pattern. Third, it drafts the appeal letter itself, pulling the correct clinical language, citing the relevant policy, and attaching supporting documentation, so a human reviews and sends rather than writes from scratch. Fourth, and this is the part vendors underplay, it flags exceptions and low-confidence cases for a human instead of guessing, which is the difference between a system billers trust and one they quietly route around.

That human-in-the-loop pattern is the same one we built for a medical-legal operations client where document intake, appointment management, and email-to-case assignment now run on autopilot and only surface exceptions when confidence is low or data is missing, which took roughly $300K a year of manual work off the table (see the Preferred Med Network case study). Denial appeals are structurally similar: high document volume, payer-specific language requirements, and a real cost to false confidence. If you're also evaluating document-heavy healthcare automation upstream of billing, our guide to medical chronology software covers the same build-vs-outsource-vs-buy decision for legal record summarization, a closely related problem shape.

Compliance and data: what to ask before AI ever touches a claim

The compliance question isn't whether AI can read a denial letter, every foundation model can do that. It's where your PHI goes once it's in the pipeline. Any vendor or build partner should be able to answer three questions clearly: does claims and clinical data ever leave your infrastructure, is any of it used to train a shared model that other customers' data also touches, and what's the data retention policy on anything sent to a third-party API.

For claims and clinical data specifically, the safest architecture runs open-source language models self-hosted on your own infrastructure or a HIPAA-compliant private cloud, with zero data retention on any external call. That removes the "our data trained someone else's model" risk entirely, which matters more in denial management than in most AI use cases, because appeal letters routinely include diagnosis codes, treatment notes, and other PHI that has no business sitting in a third-party vendor's logs. This is the same self-hosted, zero-retention posture we run for regulated clients generally, and it's worth putting in writing in any vendor contract, not just assuming it from a SOC 2 badge.

What this actually costs and how long it takes

Realistic ranges, based on comparable healthcare document-automation and billing-automation engagements: a narrow denial-categorization and appeal-drafting agent for a single-specialty or small multi-specialty group typically runs in the range of a focused 8 to 14 week build, not a year-long platform rollout. That's roughly the same timeline as the credentialing automation we built for a healthcare staffing firm that cut a 120-day process to 30, detailed in this breakdown of that engagement, and it's consistent with the three projects (14, 10, and 4 weeks) that saved a separate healthcare staffing client roughly $310K a year across invoicing, applicant screening, and performance tracking.

The number to actually track before scaling anything is your denial-to-cash cycle time and your appeal win rate on the cases the system touches, not "number of denials processed." A system that processes twice as many denials but doesn't move your win rate or your days-to-resolution isn't solving the problem, it's just producing faster paperwork. That's the fake-productivity trap: pilots that look busy without moving the number that pays for the pilot. Prove the win-rate and cash-cycle improvement on a narrow slice of your denial volume, usually your top three or four payer-code combinations, before expanding scope.

If you're working through this decision, this is exactly what our Discovery phase maps out before any code gets written, and we're happy to compare notes on where your denial volume and payer mix actually land.

Frequently asked questions

What is denial management in healthcare billing?

Denial management is the process of identifying why a submitted claim was rejected by a payer, determining whether it's worth appealing, correcting or documenting the issue, resubmitting within the payer's filing deadline, and tracking the outcome. It happens after claim submission, distinct from prior authorization, which happens before a service is rendered.

How much do claim denials actually cost hospitals and medical groups each year?

The AHA frames denials as a roughly $20 billion problem for providers industry-wide, with 15.7% of Medicare Advantage claims and 13.9% of commercial claims initially denied. Aptarro's 2025 data puts the overall initial denial rate at 11.81%, up from about 10.2% the year prior, so the trend is worsening, not improving.

What's the difference between denial management and denial prevention?

Denial prevention happens before submission: catching eligibility gaps, coding errors, or missing authorizations so fewer claims get denied in the first place. Denial management happens after a denial arrives: categorizing the cause, deciding whether to appeal, and recovering money already at risk. Most organizations need investment in both, but they're separate systems solving different points of failure.

Should a mid-size medical group build custom AI denial management or buy an RCM add-on?

Buy an off-the-shelf module if your denial volume is low and concentrated in a few payers with repetitive causes. Build or augment with a custom agent if your payer mix is fragmented across five or more plans, your denial volume outpaces staff bandwidth for appeals, and your appeal win rate sits well below the industry's roughly 54% average, which usually signals a speed problem, not an unwinnable one.

How does AI-driven denial management actually work day to day?

A working system reads incoming remittance advice, categorizes each denial by root cause, pulls the relevant claim and clinical documentation, drafts an appeal letter in the payer's expected format, and routes anything low-confidence to a human biller for review before it's sent. Over time it sharpens against your organization's specific payer patterns rather than generic industry rules.