AI Medical Coding: Accuracy, Errors and the Role of AI

Last updated: October 21, 2026

Editorial note: This article draws on peer-reviewed studies published in NEJM AI, federal data from the Centers for Medicare & Medicaid Services (CMS), an industry survey by Experian Health, and compliance guidance from the HHS Office of Inspector General (OIG). Figures supplied by software vendors are labeled as vendor-reported. Helathora has no commercial relationship with any coding software company mentioned here.

6.55%Medicare fee-for-service improper payment rate, FY 2025

$28.83BEstimated improper Medicare payments in the same period

41%Providers reporting claim denial rates of 10% or more (2025)

<50%Accuracy of general chatbots asked to produce codes on their own (2024 study)

Every hospital visit, lab test and surgery ends up as a string of codes. ICD-10 codes describe the diagnosis, CPT and HCPCS codes describe the service, and together they decide what an insurer pays, what a public health agency counts and what shows up in a patient’s record for years. It is meticulous work, it is chronically understaffed, and small slips are expensive. That is why medical coding has become one of the most heavily marketed uses of artificial intelligence in healthcare administration.

The real question is not whether software can assign codes. It already does, in many hospitals. The question is whether it makes fewer mistakes than the current process, and which mistakes it introduces that humans would not.

This article focuses mainly on the United States, where most of the published research, payment data and enforcement actions come from. A short section at the end explains how the picture differs elsewhere.

Quick answer: Yes, AI can reduce some administrative coding errors, but only in specific designs. General-purpose chatbots that “guess” codes performed poorly in testing. Systems that look up official code descriptions and similar past cases before choosing a code performed much better, in one study even outperforming physician-assigned codes. Every serious deployment still keeps trained coders reviewing uncertain or high-value charts, and compliance guidance increasingly stresses controls to catch unsupported codes, whoever or whatever produced them.

Where Coding Errors Actually Come From

It helps to separate coding errors from the wider pool of billing problems. CMS measures payment accuracy each year through its Comprehensive Error Rate Testing (CERT) program, which reviews a random sample of Medicare fee-for-service claims against coverage, coding and payment rules. CMS estimated that 6.55% of Medicare fee-for-service payments were improper in fiscal year 2025, representing $28.83 billion, down from 7.66% the year before. The underlying claims review covered July 1, 2023, through June 30, 2024.

How to read this figure: The 6.55% is not a medical coding error rate, and improper does not mean fraudulent. The estimate includes payments affected by insufficient documentation, coverage and medical necessity requirements, coding issues and other payment errors. It measures payment compliance, not the performance of any coding software.

On the provider side, Experian Health’s 2025 State of Claims survey of 250 revenue-cycle professionals asked what drives claim denials. The answers show that coding is a real but not dominant factor: front-end data problems are bigger.

Most frequently reported causes of claim denials (2025 survey)

Missing or inaccurate claim data

50%

Prior authorization

35%

Registration data errors

32%

Code inaccuracy

24%

Service not covered

23%

Share of 250 surveyed revenue-cycle professionals naming each factor; respondents could select more than one. These are survey responses, not an analysis of all medical claims. Source: Experian Health, State of Claims 2025, as reported by Becker’s Hospital Review.

This matters for expectations. A perfect coding engine would still leave most denials untouched if the patient’s insurance details were wrong at check-in. AI coding tools address one link in the chain, not the whole revenue cycle.

Three Generations of Coding Software

“AI coding” covers very different technologies. Lumping them together is the main reason headlines swing between “AI will replace coders” and “AI is useless at coding.”

Approach	How it works	Who decides the final code	Main error risk
Computer-assisted coding (CAC)	Natural language processing highlights terms in the note and suggests codes	A human coder reviews every chart	Coder fatigue and over-trust in suggestions
General chatbot, no tools	A large language model is asked to produce codes from memory	Not suitable for billing	Invented or imprecise codes
Retrieval-augmented model	The model describes the diagnosis, then checks official code lists and similar past cases before choosing	Human review of flagged or uncertain cases	Missed secondary codes; limited testing so far
Autonomous coding	Software finalizes high-confidence encounters and routes the rest to coders	Software for qualifying charts; humans for exceptions and audits	Systematic errors repeated at scale

What the Evidence Shows

General-purpose LLMs: limitations in medical coding benchmarks

In April 2024, researchers at the Icahn School of Medicine at Mount Sinai published a benchmark in NEJM AI using more than 27,000 unique diagnosis and procedure codes from a year of routine care. GPT-4, GPT-3.5, Gemini Pro and Llama 2 were given the official description of each code and asked to return the matching code. All of them stayed below 50% exact-match accuracy, and they often produced codes that were imprecise or did not exist. GPT-4 performed best, but not well enough to be trusted with billing.

It is important to be precise about what was tested. This was a code-lookup task: going from a short description to the right code. The study did not ask the models to read full medical records or to build every code needed for a real insurance claim, which are different and usually harder tasks. The result shows that these models cannot be relied on to recall codes from memory; it does not mean they get more than half of all real charts wrong.

Why this happens: Codes are long, precise alphanumeric strings with tens of thousands of near neighbors. A language model trained to produce plausible text can produce a plausible-looking code that is one character away from correct, which is still a wrong claim.

The same models with a lookup step: much better

The same research group returned in 2025 with a different design. Instead of asking the model to recall a code, they had it write a plain-language diagnosis first, then retrieve ten candidate entries from a database of more than one million hospital records before choosing. Tested on 500 emergency department visits across nine models, the retrieval-enhanced versions beat the plain models every time and, in many cases, scored better than the codes physicians had entered. Smaller open-source models came close to large commercial ones once they had the lookup step.

These findings should be interpreted within the study’s specific setting. Performance on a limited sample of emergency department encounters does not establish that retrieval-augmented systems are superior to experienced medical coders across inpatient records, specialist encounters, procedure coding or complete insurance claims. Independent validation on representative clinical data is needed before generalizing the results to routine billing.

The authors were careful about scope: the study covered only primary diagnosis codes at emergency department discharge, the system is not approved for billing, and Mount Sinai is piloting it as a suggestion and error-flagging tool inside the health record.

Autonomous coding in hospitals: promising, mostly vendor-reported

Several large health systems now let software finalize routine outpatient charts without a coder touching them. Published results come mostly from vendors. In one case study published by Solventum (formerly 3M Health Information Systems), a health system reported that after building trust by reviewing every auto-assigned code, it stopped reviewing about 95% of high-confidence codes, and around 15% of complete code sets needed no review at all. These are useful signals, but they are not independent trials, and they do not report error rates measured by outside auditors.

ICD-10-CMCPT / HCPCSComputer-assisted codingRetrieval-augmented generationAutonomous codingClaim denials

The Risk of Errors at Scale: Why Auditing Matters

A tired coder makes scattered mistakes. A misconfigured algorithm makes the same mistake on every chart that matches its pattern. That difference is now drawing regulatory attention.

In February 2026, the OIG published compliance guidance for Medicare Advantage that names, as a potentially abusive practice, prompting physicians through electronic record systems, including prompts generated by AI, to add diagnoses that patients did not have or that did not affect their care. The same period brought two large Medicare Advantage settlements: Kaiser Permanente affiliates agreed to pay $556 million in January 2026 and Aetna $117.7 million in March 2026, resolving allegations involving inaccurate or unsupported diagnosis codes.

These settlements concerned allegations about diagnosis coding and reimbursement practices; they were not findings that AI coding software caused the misconduct, and settlements are not determinations of liability. They are relevant here because they illustrate the financial and compliance consequences of submitting unsupported diagnosis codes, regardless of whether coding decisions are made by humans or assisted by software.

Regulatory guidance highlights the importance of effective compliance programs, accurate diagnosis reporting and controls to identify unsupported coding. The specific legal and audit obligations depend on the organization, the payment program and the applicable rules. In practice, many health law firms advise clients to run periodic audits comparing software-generated codes with the underlying documentation.

For patients, too: A wrong diagnosis code can follow someone into future insurance decisions or clinical records. Patients who spot an unfamiliar diagnosis on an explanation of benefits can ask their provider’s billing office for a coding review.

A Checklist for Evaluating an AI Coding Tool

Accuracy is reported at the code level and the full-encounter level, measured against an independent audit rather than the vendor’s own sample.
The share of charts finalized without review is reported next to the accuracy figure; one number means little without the other.
Every code links back to the sentence in the note that supports it.
Low-confidence, high-reimbursement and unusual cases route to a credentialed coder automatically.
A random sample of auto-coded charts is re-reviewed on a fixed schedule.
Coding patterns are compared against payer and CMS benchmarks to catch drift toward higher-paying codes.
Annual code-set updates are applied and tested before the October 1 ICD-10 changeover.

Related reading on Helathora

Limitations

The strongest independent evidence comes from a single health system and, in the 2025 study, from primary emergency department diagnoses only. Inpatient coding, procedure coding and specialty charts are harder and less studied. Real-world autonomous coding results are largely published by the companies selling the software. Some figures come from summaries and press releases rather than the full reports, and coding rules and payment data are updated every year.

What AI Medical Coding Means Outside the United States

The US system is unusual: codes such as ICD-10-CM, CPT and HCPCS feed directly into claims sent to Medicare, Medicaid and private insurers, so a coding error often becomes a payment error. Many other countries use their own versions of the ICD classification and different procedure systems, and coding in public hospitals is mainly used for activity statistics, hospital funding and health planning rather than individual insurance claims. In Spain, for example, hospital discharge coding feeds the national minimum basic data set (CMBD) and is carried out within each regional health service.

The general lessons still travel well: models that guess codes from memory are unreliable, lookup-based designs perform better, and human review of uncertain cases remains essential. The specific figures on denials, improper payments and legal settlements in this article, however, apply to the US context and should not be read as universal.

The Bottom Line

AI can reduce administrative coding errors when it is built to check official sources rather than guess, and when humans stay in charge of the cases where mistakes are costly. It does little for denials caused by registration or authorization problems, and it can multiply errors if nobody audits it. The organizations seeing benefits treat it as a fast first pass and an error detector, not as a replacement for coding expertise.

Frequently Asked Questions

Will AI replace medical coders?

Current evidence points to a changing role rather than replacement. Software handles routine, high-confidence charts, while coders focus on complex cases, audits and appeals. All published deployments keep human review for exceptions.

Can I use ChatGPT or a similar chatbot to code medical records?

Not on its own. The 2024 Mount Sinai benchmark found that general chatbots asked to produce codes directly were below 50% accurate and sometimes invented codes. Accuracy improved substantially only when the model was connected to official code lists and past cases.

Does AI coding lead to upcoding?

It can if a system is tuned toward higher-paying codes or prompts clinicians to add unsupported diagnoses. Federal compliance guidance for Medicare Advantage published in 2026 specifically flags AI-generated prompts of this kind, which is why regular audits of a tool’s output are widely recommended.

Is autonomous coding approved by regulators?

There is no specific federal approval process for coding software in the United States. The organization submitting the claim remains responsible for its accuracy, whatever tool produced the codes.

Official Sources

Soroush A, et al. Large Language Models Are Poor Medical Coders — Benchmarking of Medical Code Querying. NEJM AI, 2024.
Klang E, et al. Assessing Retrieval-Augmented Large Language Models for Medical Coding. NEJM AI, 2025.
Centers for Medicare & Medicaid Services. Comprehensive Error Rate Testing (CERT).
Experian Health. 2025 State of Claims survey, September 2025.
Becker’s Hospital Review. What’s to blame for claim denials? (summary of the Experian Health report), 2025.
HHS Office of Inspector General. Medicare Advantage Industry Compliance Program Guidance, 2026.
U.S. Department of Justice. Kaiser Permanente affiliates settlement and Aetna settlement, 2026.
Solventum. Autonomous medical coding journey (vendor case study).

Disclaimer: This article is for general information only and does not constitute medical, legal, billing or financial advice. Coding rules change every year and differ between payers and countries. Healthcare organizations should consult certified coding professionals and legal counsel before adopting or changing coding software. Helathora is not affiliated with any company or product mentioned.