Autonomous medical coding is the use of artificial intelligence to read clinical documentation and produce final billable codes without requiring human review of every encounter. Accuracy is the defining question of the category, and the methodology used to measure it matters as much as the number itself.
The accuracy of autonomous medical coding is the single most-asked question by healthcare organizations evaluating the technology. The methodology used to measure that accuracy matters as much as the number itself. Numbers quoted by vendors without an audit methodology behind them are not comparable.
What does "accuracy" mean in autonomous medical coding?
There is no shared definition. "Accuracy" in autonomous coding is one of the most loosely used numbers in healthcare AI, and the looseness is structural to the category. Three distinct measurements get conflated under the single word: accuracy on the subset of charts a system attempts to code, the automation rate (the share of total charts the system codes end-to-end versus the share it routes to a human), and accuracy at the encounter level across the full chart population. A vendor reporting "98% accuracy at 90% automation" is reporting accuracy on the 90% of charts the system was confident enough to code on its own. The accuracy of the 10% it routed to humans is not measured, and whether the confidence calibration on that 10% is correct is also not measured.
The market-defining analysis of the current state of the category is KLAS Research's August 2025 report "Autonomous Coding 2025: A Promising Start for an Early Market." It is the first KLAS analysis to formally recognize autonomous coding as a distinct healthcare technology segment, and it documents that production deployments today are concentrated in two specialty settings: radiology and emergency department coding. These are specialties where clinical documentation is comparatively structured and where the distribution of codes used is comparatively narrow. The high automation rates that vendors publish today should be read with that context. They are typically high in narrow specialty footprints, not across the full mix of work a multi-specialty health system or physician group actually produces.
A methodologically rigorous accuracy figure requires three properties to be comparable across vendors. First, it must be measured on a representative sample of the encounter population the system is sold to handle, in the proportions those encounters actually appear in production. Second, the comparison standard must be panel-adjudicated source documentation, not a single human coder. Two credentialed coders independently coding the same chart only agree about 82% of the time at billing-grade specificity (Peng et al., 2018), and a number measured against a single human carries that noise floor. Third, the methodology and sample size must be published. The difference between a 5,000-chart audit and a 50-chart vendor demonstration is the difference between a population estimate and an anecdote.
Anything short of those three properties is a marketing number. It may correlate with real performance. It does not measure it.
How is accuracy measured?
The methodology that survives scrutiny is the credentialed-panel audit. The structure is straightforward. A statistically representative sample of charts is drawn from the encounter population. Each chart is independently coded by two or more credentialed coders, typically Certified Coding Specialists (CCS) for facility coding and Certified Professional Coders (CPC) for professional services. The autonomous coding system codes the same charts in parallel. Disagreements between the AI output and the human coders, and between the human coders themselves, are then adjudicated. Each disagreement is investigated against the source clinical documentation, current ICD-10-CM and CPT guidelines, the most recent AHA Coding Clinic guidance, the CMS National Correct Coding Initiative (NCCI) edits in effect for the date of service, and any payer-specific rules that apply.
The output is two numbers, and both matter. The initial agreement rate is the share of charts where the AI's code matched the human coder before adjudication. The post-adjudication consensus rate is the share of charts where the AI's code matched the panel-adjudicated ground truth after disagreements were investigated. The post-adjudication rate is the system's actual error rate once human coder errors are removed from the comparison. The initial agreement rate is informative because it approximates what a customer would see if they ran a casual audit using a single in-house reviewer.
This is the methodology AccuCode has been audited under for the past eighteen months across tens of thousands of encounters. The audits were conducted by two independent organizations. MedAxiom, the American College of Cardiology's cardiovascular collaborative, audited cardiovascular surgery and evaluation-and-management (E/M) codes. Professional Consulting Services (PCS), Arkansas's largest third-party medical billing firm, audited E/M and broader specialty coverage. The MedAxiom audit was the gating test for the partnership that became AccuCode CV. It preceded the partnership rather than followed it. Both audits were led personally by the most senior credentialed coders at each organization, who put their professional reputations and their organizations' continued financial responsibility for hundreds of provider customers on the line in the result.
A related test that is useful but is not a substitute is the credentialing exam itself. The AAPC Certified Professional Coder (CPC) examination requires a 70% score to pass; credentialed coders who pass routinely score in the 75 to 85 percent range. In spring 2025, AccuCode's coding engine was given a copy of the CPC exam with no preprocessing and no contextual hints, and scored 100%. The CPC exam is closed-form and time-limited; it is not the same test as production coding against a hospital's clinical documentation. But the score establishes that the system's working knowledge of the codes themselves, and of the rules that govern their selection, is at the ceiling of what the credential is designed to measure.
How does AI coding accuracy compare to human inter-coder reliability?
The human baseline is best established in the peer-reviewed literature. The most rigorously cited recent study is Peng et al., 2018, published in the International Journal of Population Data Science. The study audited 1,636 emergency department records sampled from 11 hospitals in Alberta, with two credentialed coders independently coding each chart. Agreement was 86.5% at 3-digit ICD-10 (Cohen's kappa 0.86) and 82.2% at 4-digit (billing) specificity (kappa 0.82). Independent replication in other populations and chart types produces figures in similar ranges, with variation driven by chart complexity, specificity of comparison, and the credentialing level of the reviewers.
The peer-reviewed reality is that two credentialed coders, working from the same documentation under controlled conditions, agree on roughly four out of five charts at the level of specificity that determines billing. The remaining one in five is structural noise: different but defensible interpretations of the documentation, codes selected at different levels of specificity, and genuine disagreements that adjudication is required to resolve. This figure is the structural ceiling on any AI system trained against human-coded historical data. A model trained on past human-coded examples cannot reliably exceed the noisy ground truth it was trained against. A vendor claiming "96% accuracy against human coders" is implicitly claiming to be roughly equivalent to an average human coder. On a methodologically rigorous comparison against panel-adjudicated source documentation, that number would likely come down materially.
Human inter-coder reliability at billing specificity, contrasted with AccuCode's published audit results. Cardiovascular surgery is the floor of AccuCode's specialty audit; every other specialty produces higher scores on both measures.
Sources: AAPC CPC certification passing threshold (70%); Peng et al. 2018, IJPDS (82.2% inter-coder agreement at 4-digit ICD-10, kappa 0.82); AccuCode independent audits by MedAxiom and PCS, 2024-2025.
AccuCode's published audit results are not trained against human-coded data, which is why they exceed the human-baseline ceiling. The architecture reads clinical documentation and reasons from what is documented, then produces output that conforms to the current ICD-10-CM, CPT, and HCPCS code sets, current NCCI edits, and current payer-specific rules. There is no pattern-matching against historical human-coded examples, and there is no fine-tuning on customer data. The MedAxiom cardiovascular surgery audit produced 95.6% initial agreement with the credentialed coding panel and 99.1% post-adjudication consensus against panel-adjudicated source documentation. The MedAxiom and PCS E/M audit produced 98.4% initial agreement and 99.6% post-adjudication consensus. Cardiovascular surgery is the floor of AccuCode's specialty audit, among the most complex specialties in the codebook for bundling and sequencing. Every other specialty the system covers produces higher scores than cardiovascular surgery on both measures.
The operational implication is the one most often missed in vendor evaluations. If a system's initial agreement with a single human coder is materially higher than the rate at which two credentialed human coders agree with each other, then routing every chart through a single human reviewer in production does not increase accuracy. It decreases it. The human reviewer is, in that dimension, a less accurate measurement than the AI is. The appropriate role of human review in a high-accuracy autonomous system is sampling and methodology: periodic credentialed-panel audits to verify continued performance, and structured review of the small set of cases the system flags as low-confidence. Not per-chart verification of every encounter.
Quick answers to the questions buyers ask most often about this topic.
What is autonomous medical coding?
Autonomous medical coding is the use of artificial intelligence to read clinical documentation and produce final billable codes (ICD-10-CM diagnoses, CPT and HCPCS procedures, and modifiers) without per-chart human review. It is distinct from computer-assisted coding (CAC), which presents suggested codes to a human coder for verification, and from documentation-assist tools that surface gaps or clinical indicators. An autonomous system produces final codes ready for claim submission.
How is autonomous coding accuracy verified?
The methodology that survives scrutiny is the credentialed-panel audit against source documentation. A representative sample of charts is independently coded by two or more credentialed coders. The autonomous system codes the same charts in parallel. Disagreements are adjudicated against the source documentation, current ICD-10-CM and CPT guidelines, current AHA Coding Clinic guidance, current NCCI edits, and applicable payer rules. The published outputs are the initial agreement rate and the post-adjudication consensus rate. Vendor-quoted accuracy not measured this way is not directly comparable.
Can autonomous coding handle inpatient encounters?
The architectural complexity differs materially between settings. Outpatient and professional-services coding operates on smaller documentation footprints with narrower code distributions. Inpatient facility coding involves DRG assignment, MS-DRG and APR-DRG logic, present-on-admission indicators, multiple procedures sequenced across an admission, and substantially longer clinical narratives. Most autonomous coding systems are stronger in one setting than the other. The KLAS 2025 Autonomous Coding report documents broader current production deployment in outpatient and emergency department settings than in inpatient facility coding.
Sources cited
- Peng, M., Eastwood, C., Boxill, A., Jolley, R.J., Rutherford, L., Carlson, K., Dean, S., Quan, H. (2018). Coding reliability and agreement of International Classification of Disease, 10th revision (ICD-10) codes in emergency department data. International Journal of Population Data Science, 3(1):445. DOI: 10.23889/ijpds.v3i1.445. Inter-coder agreement of 86.5% (3-digit) and 82.2% (4-digit ICD-10) between credentialed coders, with Cohen's kappa of 0.86 and 0.82 respectively. ijpds.org/article/view/445
- KLAS Research. Autonomous Coding 2025: A Promising Start for an Early Market. August 2025. The first KLAS report to formally recognize autonomous coding as a distinct healthcare technology segment. klasresearch.com/segment/autonomous-coding
- American Academy of Professional Coders (AAPC). Certified Professional Coder (CPC) examination content outline, format, and 70% passing threshold. aapc.com/certification/cpc
- American Hospital Association. AHA Coding Clinic for ICD-10-CM/PCS. Quarterly official coding guidance referenced by all U.S. coders and AI coding systems. ahaonlinestore.com
- Centers for Medicare & Medicaid Services. National Correct Coding Initiative (NCCI) policy manual and procedure-to-procedure (PTP) edits. cms.gov/medicare/coding-billing/ncci-edits
- MedAxiom (American College of Cardiology). Independent audit of AccuCode AI cardiovascular surgery and E/M coding accuracy, conducted by Nicole F. Knight, LPN, CPC, CCS-P (EVP, Revenue Cycle Solutions) and team, 2024-2025. medaxiom.com
- Professional Consulting Services (PCS). Independent audit of AccuCode AI E/M and multi-specialty coding accuracy, conducted by Scott Roper, MBA, CPC and Tracye Enis, CPC, 2024-2025.