The figure now showing up in every RCM vendor pitch deck, "autonomous coding cuts denials by 70 percent," has a specific origin. Here's where the number actually came from, and what it takes to reproduce it.
01 / The numberThe 70% number, where it actually came from
Oregon Health & Science University Hospital, working with CodaMetrix on radiology coding, reported a roughly 70% drop in coding-related denials on autonomously coded cases compared to manually coded ones. Coder workload dropped 28 percent. The automated denial rate landed at 0.33%.
That is a real, audited result. It is not a marketing average. It is the outcome of one health system, one service line, one vendor, one workflow, run for long enough to mean something.
Geisinger Health, working with Nym, drove coding-related denials below 0.1% on the encounters it routed through autonomous coding. Inova Health ran the same pattern in its emergency departments. Different vendors, different service lines, different outcomes. But the direction is the same.
The number that matters is not 70. It's the question underneath: why is this working, and where will it keep working?
02 / ContextWhy denials are getting worse, not better
Before we get to the mechanism, the context. Initial claim denial rates rose to 11.8% in 2024, up from 10.2% a few years earlier. Forty-one percent of providers now report denial rates above 10%, up from 30% in 2022 and 38% in 2024.
It is getting worse on the dollar side too. MDaudit's 2025 network data, drawn from 1.2 million providers and 4,500 facilities, showed average outpatient claim denial amounts rising 14% year over year. Average Medicare Advantage denial dollars rose 22.4% to roughly $1,000 per claim. Outpatient coding denials specifically rose 26% from 2024 to 2025. Hospital revenue leakage rose roughly 25% in 2025 even as payment speed improved. A tell that providers are getting paid faster on what they collect, but losing more of what they should have collected.
Initial denial rate: 11.8% in 2024, up from 10.2%. 41% of providers now report >10% denial rates. Average MA denial dollar: ~$1,000 (up 22.4%). Outpatient coding denials: up 26% YoY. Hospital revenue leakage: up ~25% in 2025.
Why the squeeze? Three reinforcing trends:
- Payers are running their own AI on incoming claims, scrutinizing medical necessity and documentation at a depth that wasn't economically possible five years ago.
- Payer-specific rules are proliferating. Medicare Advantage plans, in particular, layer their own utilization-management rules on top of CMS guidance.
- Documentation requirements have moved faster than physician workflows. Clinical notes that used to be acceptable now generate denials they wouldn't have generated in 2019.
In other words: providers are bringing manual coding to a fight that's increasingly automated on the other side.
03 / Root causeWhere the denials actually come from
Most coding-related denials trace to a small number of root causes:
Distribution of coding-related denial drivers based on ASP-RCM client benchmarks across high-volume outpatient settings, 2025.
Manual coding, even by an experienced coder, hits all five at scale. There is no human who can hold every payer's rule library in working memory across thousands of encounters a day.
This is precisely what an autonomous coding platform is built to address. Not all of it. Some of it.
AI applies rules consistently, at speed, with memory. That's the mechanism.
04 / MechanismWhy the 70% works where it works
Look closely at what an autonomous coding system actually does between encounter and submission, and you can see why the denial rate moves.
- Real-time payer rule application. The system loads each payer's current rule set (coverage policies, prior auth requirements, documentation requirements, reimbursement guidelines) and applies them to every chart. A coder might know fifty rules cold. A platform applies five thousand consistently.
- Pre-submission scrubbing. Before a claim leaves the building, the engine flags missing documentation, inconsistent code combinations, modifier issues, and medical-necessity mismatches. They get fixed upstream of the payer.
- Continuous learning from denials. When a claim is denied, the platform doesn't just rework it. It updates so future similar claims don't repeat the error. Manual coding doesn't have that feedback loop unless the organization builds one explicitly.
- Volume consistency. Tuesday morning's chart is coded the same way as Friday afternoon's. Coder fatigue is a real, measurable contributor to error rates in busy departments, and the engine doesn't get tired.
- Compliance posture. The good systems generate a transparent audit trail for every code assigned, which matters when the OIG comes knocking. Nym, for instance, markets explainability (every code with a documented rationale) as a primary feature, not a nice-to-have.
That is the mechanism. It is not "AI is smart." It is AI applies rules consistently, at speed, with memory.
05 / Reality checkWhere 70% becomes 20%, or 5%, or zero
The OHSU result is in radiology. There are reasons.
Radiology reports are highly structured. The documentation is usually complete (you read the image or you don't). The procedure list is bounded. The codes that apply are a relatively narrow set. The same is true for ED encounters, pathology, certain outpatient surgical specialties.
Inpatient DRG coding is a different problem. Principal-diagnosis sequencing involves judgment calls that the best AI systems still get wrong often enough to require human review on most cases. Behavioral health is different. Complex multi-condition oncology cases are different. Anything that depends on inferring intent from prose, rather than mapping prose to codes, is harder.
The honest framing for any health system evaluating this: autonomous coding will move denial rates dramatically in the service lines where the documentation is structured and the code set is narrow, and it will move them less, or not at all, in the service lines where it isn't. The 70% is a real number for one part of the operation. Promising a system-wide 70% across every service line is not how this technology works.
06 / PlaybookWhat it takes to actually get the result
The systems publishing strong outcomes share four things, and most of them are operational rather than technical.
- Pick the right service line first. Radiology, ED, pathology, and certain outpatient procedural areas are where teams should start. High volume, structured documentation, bounded code set.
- Run parallel for at least 60 days. Code the same charts manually and through the engine, then compare. This is not optional. It's the only way to see where the engine's confidence threshold is calibrated wrong for your population.
- Measure routing accuracy, not just coding accuracy. A 95%-accurate engine that confidently routes the wrong charts to itself is worse than a 90%-accurate engine that knows when to ask for help. Watch the rate at which the engine correctly hands off uncertain cases to humans.
- Reorganize your coders around the new workflow. This is the failure mode for most deployments. Buying the platform and leaving the team structure unchanged produces modest results at best. The teams getting OHSU and Geisinger numbers retrained their coders into validators, exception handlers, and physician-feedback specialists, and built denial-pattern review into their weekly cadence.
- Keep humans on the most expensive cases. Inpatient DRGs, complex specialty cases, anything where a misroute costs five figures. The economics of autonomous coding don't depend on covering 100% of encounters. They depend on covering the right 60 to 70%.
07 / Practitioner viewHow ASP-RCM helps you reproduce these numbers
We don't sell autonomous coding platforms. We run revenue cycles, and we've deployed engines from multiple vendors inside client environments. That's a different conversation than the one the platform vendors will have with you.
Where we add the lift:
- Pre-deployment denial diagnostic: we segment your last 12 months of denials by root cause and service line, so you know which lines will move 60% and which won't move at all
- Vendor-neutral evaluation across CodaMetrix, Nym, Fathom, and others, mapped to your specialty and payer mix
- Parallel-run discipline for the first 60-90 days, with weekly routing accuracy reviews
- Coder reorg into validators, exception handlers, and CDI feedback specialists. The operational change that determines whether you get the OHSU number or a fraction of it
- Payer rule library maintenance against weekly MA and commercial policy updates
- Compliance & audit posture: full code-level rationale trails for OIG, RAC, and payer audits
08 / Business valuePre vs post: the numbers we see in production
The ranges below reflect what production deployments (OHSU, Geisinger, Inova, and ASP-RCM client engagements) typically see in the right-fit service lines (radiology, ED, pathology, outpatient procedural). Inpatient DRG and behavioral health move more conservatively in year one.
09 / Where to startThe realistic outlook
Denial rates in U.S. healthcare are not going down on their own. The macro pressure (payer automation, MA rule complexity, tightening medical-necessity scrutiny) is moving the wrong direction. The providers who do nothing will see denials keep climbing.
Autonomous coding, deployed honestly and operationally, is a real lever. The 70% number is real for radiology at OHSU. Your number, on your service line, with your payer mix and your documentation discipline, will be different. It might be 70%. It might be 45. It might be 25 in year one and 50 in year two as the engine learns your population.
What it won't be, if you do this right, is zero. And in a market where the average MA denial is now north of $1,000 and the cost to rework a single claim runs to $181, even a 25% reduction is a serious number on the bottom line.
The right next step isn't a vendor demo. It's an internal review of your denial data by service line and root cause. Find the lines where denials are concentrated, the documentation is structured, and the dollar amounts are highest. That's your starting point. The technology decision comes after.
Sources & references
OHSU + CodaMetrix: Healthcare IT News, May 2025. Experian Health, State of Claims 2025 (denial rate trends and provider survey data). MDaudit, 2025 Benchmark Report on payer audits and denial dollar amounts. Kodiak Solutions, 2025 Hospital Revenue Cycle Benchmark on revenue leakage. Nym Health, Geisinger and Inova case studies. Becker's Hospital Review on AHIMA workforce signals. CounterForce Health denial-and-appeals analysis.