Short answer: no. Longer answer: the job is changing faster than it has in thirty years, and the coders who treat AI as a threat will lose ground to the ones who treat it as a tool.
01 / The caseA morning at OHSU
At Oregon Health & Science University Hospital (576 beds, the only Level I trauma center in the state), radiology coders used to face a backlog they couldn't out-hire their way through. The hospital was expanding. Volumes were climbing. The market for credentialed coders was, and is, tight enough that you can't simply post the job and wait.
So OHSU did something a growing number of health systems are now doing: it routed its radiology charts through an AI engine first. Charts the engine could handle confidently went straight to billing. Charts it couldn't (the messy ones, the unusual ones, the ones with documentation gaps) got routed to a human coder.
The early numbers, reported by the hospital's revenue cycle director Tammy Bickle, are striking. Coder workload dropped 28 percent. The denial rate on autonomously coded radiology cases came in roughly 70 percent lower than on the manually coded ones. The automated denial rate sat at 0.33 percent.
02 / The numbersThe numbers everyone is circling
OHSU + CodaMetrix, radiology coding pilot:
That is the data point everyone in revenue-cycle leadership is now circling. And it's the data point driving the conversation behind every "Is manual coding dying?" question I get asked.
Here's my honest take.
03 / DefinitionsWhat "autonomous coding" actually is, and what it isn't
Autonomous medical coding is software that reads a clinical encounter (a radiology report, an ED visit note, an op-report) and assigns ICD-10-CM, CPT, and HCPCS codes without a human in the loop on that specific chart. The serious systems (Nym, CodaMetrix, a handful of others) do two things at once: they use clinical language understanding to interpret the documentation, and they layer on rules that map findings to the correct codes per current payer guidance.
What it isn't:
- It is not "AI suggests, coder approves." That's computer-assisted coding (CAC), and it has been around for fifteen years.
- It is not magic. Vendors that publish 95-percent-plus first-pass accuracy figures are quoting those numbers on the charts the system chose to handle. The rest still go to humans.
- It is not a replacement for a department. It's a triage layer that handles the high-volume, well-documented, narrow-scope encounters (radiology, pathology, ED visits, certain outpatient surgeries) and routes the rest to people.
Autonomous coding is the difference between a human reviewing every chart and a human reviewing only the charts that need a human.
04 / ContextWhy this is happening now, not five years ago
Three things converged.
The denial spiral.
Initial claim denial rates climbed from roughly 10.2% a few years ago to 11.8% in 2024, and 41% of providers now report denial rates above 10%, up from 30% in 2022. Reworking a single denied claim costs between $25 and $181 depending on the system and the payer. Outpatient denials rose 14% year over year per MDaudit's 2025 data. Medicare Advantage denials are up 22.4% in average dollar amount. The math on prevention got better than the math on cleanup, fast.
Sources: Experian Health, State of Claims 2025; MDaudit 2025 Benchmark Report; Kodiak Solutions 2025 Hospital Revenue Cycle Benchmark.
The coder shortage.
Credentialed coders are not graduating at the rate volume is growing. Bickle at OHSU said outright that hiring more coders wasn't an option. Not in their market. Most CFOs have heard a version of that sentence in the last 18 months.
The technology actually works now.
A 2024 Mount Sinai study comparing retrieval-augmented LLMs to human ED coders on primary-diagnosis ICD-10-CM assignment found the gap had narrowed dramatically from earlier-generation models. Nym reports Geisinger Health drove its coding-related denial rate to under 0.1% on encounters routed through their engine. Inova went after ED coding with the same approach. These aren't pilots being quietly killed; they're production deployments being expanded.
AI doesn't fix bad documentation. It just exposes it, faster.
05 / Failure modesWhere AI breaks
Anyone selling you autonomous coding without naming the failure modes is selling you something.
- Clinical nuance. The kind a coder builds up over a decade: knowing that the ED physician at your level-I trauma center always documents a certain way and doesn't mean what the words literally say; knowing that a particular surgeon's "exploration" is documented language for the billable procedure she just did. Algorithms don't pick that up from training data alone.
- Edge cases. Inpatient DRG coding involves principal-diagnosis sequencing decisions that can move a case across an entire MS-DRG boundary, with five-figure reimbursement consequences. An AI that's 95% accurate on radiology is not 95% accurate on a complex inpatient case with multiple comorbidities and a sequencing question.
- Documentation gaps. If the physician didn't document it, no system can code it. AI doesn't fix bad documentation. It just exposes it, faster.
- Compliance posture. The OIG, DOJ, and major payers have all signaled clearly that human-in-the-loop validation is required for AI-generated codes in a meaningful share of contexts. "The model said so" is not a defense in a RAC audit.
- Bias and drift. Autonomous coders are trained on historical coding decisions. Historical decisions include historical errors. Without continuous validation, error patterns get baked in.
This is why every credible deployment is hybrid.
06 / WorkforceThe job is changing, and the BLS isn't worried
The U.S. Bureau of Labor Statistics projects medical records specialist employment (the category that includes coders) will grow 9% from 2023 to 2033. More than double the average for all occupations. Read that twice. The agency that tracks workforce trends is forecasting growth, not contraction, in the same decade autonomous coding enters production.
What's changing is the work, not the headcount. AHIMA's recent survey work shows 75% of HIM professionals saying upskilling is now essential. The job description is moving from "code every chart" to:
- Validate the AI's output on the charts it handled
- Code the complex cases the engine sent over
- Audit and appeal: the highest-leverage work in the cycle
- Translate clinical documentation gaps back to physicians, before they cause denials
- Manage payer rule libraries inside the autonomous engine
Mary Beth Haugen of the Haugen Consulting Group put it cleanly in industry talks earlier this year: "AI isn't here to replace us. Coders will need to embrace lifelong learning and adapt to new tools." That's the strategy. Coders who become validators, auditors, and engine stewards are getting promoted. Coders who don't are getting left behind.
07 / PlaybookWhat this means if you run a revenue cycle
I would not, today, deploy autonomous coding across your full encounter mix. I would pick one high-volume, well-documented service line (radiology is the most common starting point for a reason) and run a controlled deployment with parallel human coding for the first 60 to 90 days.
Measure three things:
- Denial rate on AI-coded cases versus the same coders' manually coded cases.
- Total coder hours saved, net of the time spent reviewing the engine's output.
- Routing accuracy: the rate at which the engine correctly hands off uncertain cases to humans rather than confidently coding them itself. This metric matters more than the headline accuracy number.
Watch for the failure modes specific to your payer mix. MA denials are rising fastest in dollar terms; if your payer mix is heavy MA, your validation discipline matters more than your automation rate.
And invest in your coders' new role now. The systems getting OHSU and Geisinger numbers aren't doing it on technology alone. They're doing it because they reorganized their coding teams around what humans are uniquely good at, and let the engine handle what it's uniquely good at.
It isn't "What's the accuracy?" It's "Where does this engine route the charts it isn't sure about, and what does my team do with them?" Get that workflow right, and the rest follows.
08 / Practitioner viewHow ASP-RCM helps health systems get there
At ASP-RCM Solutions we don't sell technology. We run revenue cycles. When clients ask us about autonomous coding, the conversation starts with their denial data, not a vendor demo. Our practice covers:
- AI-assisted medical coding with senior CCS/CPC validators reviewing the routes the engine isn't confident about
- Predictive denial analytics mapped to your specific payer mix and historical denial patterns
- Pre-submission scrubbing against payer-specific rule libraries, refreshed weekly
- Documentation feedback loops back to your physicians, because the engine surfaces the gaps; we close them
- Coder upskilling for clients who want to keep the team and reorganize the workflow
- Compliance audit trails for every AI-assigned code, ready for OIG or RAC review
09 / Business valuePre vs post autonomous coding
The numbers below reflect what production deployments (OHSU, Geisinger, Inova, and ASP-RCM client engagements in the same service-line profile) typically see in radiology, ED, and outpatient procedural coding. Inpatient DRG coding moves more conservatively and is not represented in these ranges.
10 / Final readSo, is manual coding obsolete?
No.
The chart that gets coded entirely by hand? That's becoming rare. The coder who only codes? Also becoming rare.
But the coder who validates, audits, appeals, trains the engine, and bridges clinical documentation back to physicians? That coder is more valuable than they were five years ago, not less. The hospitals figuring this out are pulling away from the ones that aren't.
Manual coding isn't dying. It's getting promoted.
Sources & references
OHSU + CodaMetrix radiology results: Healthcare IT News, May 2025. Experian Health, State of Claims 2025. MDaudit, 2025 Benchmark Report. Kodiak Solutions, 2025 Hospital Revenue Cycle Benchmark. Nym Health, Geisinger and Inova case studies. AHIMA, 2024 Workforce Survey. U.S. Bureau of Labor Statistics, Occupational Outlook for Medical Records Specialists, 2023-33 projections. Mount Sinai Health System, retrieval-augmented LLM ED coding study, 2024.