AI in Medical Coding and Billing
AI medical coding tools should be judged by whether the codes they suggest are supported by the note and survive audit; in a 2026 systematic review, most automated ICD-10 coding studies had not left the research stage. Medical coding turns the diagnoses and procedures in a clinical note into the ICD-10-CM, ICD-10-PCS, and Current Procedural Terminology (CPT) codes that drive claims, quality measures, and research data. At one academic health system, billing and insurance-related costs after a non-primary care encounter were mostly attributable to medical coding (Tseng et al., 2018). Of the 24 studies of automated ICD-10 coding published from 2020 through 2025 in a 2026 systematic review, 21 (87.5%) remained at the research stage and 3 (12.5%) reached pilot testing, despite promising benchmark performance (Tangudomkit et al., 2026).
After completing this chapter, clinicians, coding leaders, and health system leaders should be able to:
- Describe where AI tools enter the coding and billing workflow, from documentation prompts to claim submission
- Summarize what a 2026 systematic review found about automated ICD-10 coding research, including how few studies (3 of 24) reached pilot testing
- Interpret the two Mount Sinai studies of large language models for medical coding, and what each did and did not measure
- Apply professional guidance on AI-generated documentation queries to a tool a vendor proposes
- Explain why a tool that adds diagnoses can raise audit and payment-integrity risk, using a federal oversight evaluation that did not study AI as a comparator
- List the questions to ask before adopting an AI coding tool
Evidence at a Glance
| Use | Source | Source type and design | Setting | What was measured | Main result | Main caveat |
|---|---|---|---|---|---|---|
| Automated ICD-10 coding research | Tangudomkit et al., 2026 | Peer-reviewed systematic review, preregistered | 24 studies, 2020 to 2025 | Methods, performance reporting, readiness | 87.5% at research stage; 12.5% pilot testing | Preprints eligible; mixed ICD-10 versions; descriptive synthesis |
| Code lookup by language models | Soroush et al., 2024 | Peer-reviewed benchmark | Code descriptions from Mount Sinai records | Exact code match | GPT-4 best; no model above 50% | Descriptions, not clinical notes |
| Coding with retrieval | Klang et al., 2025 | Peer-reviewed retrospective case study | 500 Mount Sinai emergency visits | Masked reviewer preference | GPT-4 code preferred for accuracy 447 vs 277 | Compared with provider codes, not an audit |
| AI-generated documentation queries | AHIMA and ACDIS, 2026 | Professional practice brief; not peer reviewed | All care settings | Compliance standards | Same rules as manual queries | Guidance, not outcome evidence |
| Diagnoses reported only on health risk assessments (non-AI comparator) | HHS OIG, 2024 | Federal oversight evaluation of encounter data; did not study AI | Medicare Advantage, 2022 data | Payments from diagnoses with no other service record | Estimated $7.5 billion for 2023 | Did not check medical records; not an AI study |
| Claims and appeals agents | Bedi et al., 2026 | Preprint benchmark | Simulated systems, synthetic patients | Whole-task completion | Best agent 25.0% of appeals and denials tasks | Not peer reviewed; simulated |
Introduction
Coding sits between documentation and payment. A physician documents the visit; a coder, or the physician in many professional-fee settings, assigns codes; a claim is sent; the payer pays, denies, or audits. A missed diagnosis understates how sick a patient was and can lower payment. A code the note does not support can produce an overpayment that must be returned.
AI tools enter this workflow at several points. Computer-assisted physician documentation (CAPD) prompts a physician during the encounter. Computer-assisted coding (CAC) suggests codes after it for a coder to accept or change. Autonomous coding means software that assigns codes and sends them toward the claim without a coder reviewing each one. Language models have been tested for coding (Soroush et al., 2024; Klang et al., 2025) and for drafting documentation queries (AHIMA and ACDIS, 2026), and claims agents have been tested in simulation (Bedi et al., 2026, preprint). In one costing study, the tasks a physician did for billing included responding to queries from the billing organization, including coding-related questions (Tseng et al., 2018).
Ambient documentation tools and the billing rules for AI-drafted notes are covered in AI-Assisted Clinical Documentation. Prior authorization and other administrative work are covered in Healthcare Administration. Coding itself, documentation integrity (CDI) queries, and the audit risk that coding tools create are covered below.
What the Coding Research Shows
A systematic review of automated ICD-10 coding
A team at Prince of Songkla University in Thailand reviewed studies published from January 1, 2020, through December 31, 2025, that assigned ICD-10 codes from clinical text with machine learning, deep learning, transformer, or large language model methods. The review followed PRISMA 2020, was preregistered, and searched seven databases, including arXiv and Google Scholar, so preprints and conference papers were eligible. Of 257 records identified, 24 studies met the criteria and contributed 296 experimental evaluations. The authors rated study quality high in 7, moderate in 8, and limited by technical or methodological concerns in 9 (Tangudomkit et al., 2026).
Most studies (87.5%) remained at the research stage without validation in real-world clinical settings, and pilot clinical testing accounted for 12.5%. Most used single-site datasets, and only a small proportion reported external or cross-domain validation. (Tangudomkit et al., 2026) Source code was often unavailable, and reproducibility was usually partial.
Among studies that reported both, micro-averaged F1 scores, which pool every coding decision, were consistently higher than macro-averaged F1 scores, which weight each code equally; the gap reflects weaker performance on rare codes. When each note had to receive every applicable code from the full code set, large language models showed weaker results than hybrid deep learning and fine-tuned transformer models, although they were still emerging in the evidence base and were often tested in zero-shot, few-shot, or retrieval-assisted settings (Tangudomkit et al., 2026).
The authors concluded that although AI-based coding models show promising benchmark performance, evidence remains insufficient for real-world deployment, and that future work should assess clinical validity, workflow integration, error impact, auditability, and human oversight. (Tangudomkit et al., 2026) The included studies differed in dataset, language, label set, code granularity, and national ICD-10 version, so the review compared them descriptively rather than by meta-analysis.
Language models alone: a code lookup benchmark
At Mount Sinai, researchers gave GPT-3.5, GPT-4, Gemini Pro, and Llama 2 70B the official descriptions of 7,697 ICD-9-CM, 15,950 ICD-10-CM, and 3,673 CPT codes from the health system’s records and asked for the matching code. The models were queried in December 2023. GPT-4 had the highest exact match rate, 45.9% for ICD-9-CM, 33.9% for ICD-10-CM, and 49.8% for CPT. Of GPT-4’s wrong ICD-10-CM answers, 18.5% were codes that do not exist and 12.7% were nonbillable codes (Soroush et al., 2024).
The task was matching a code’s description to the code, not coding from clinical notes, and the study did not test prompt engineering, retrieval, tools, or fine-tuning. Models released after December 2023 were not tested. The full benchmark is discussed in Hospital Operations and Management.
Language models with retrieval: an emergency department case study
A Mount Sinai team then studied 500 randomly selected emergency department visits. Nine language models assigned each visit’s primary ICD-10-CM diagnosis code with help from a retrieval system built on more than 1 million earlier emergency visits. In a masked review by four physicians and two language models, human reviewers favored the retrieval-supported GPT-4 code over the provider-assigned code for accuracy in 447 instances versus 277, and for specificity in 509 versus 181 (P<0.001 for both). These counts are reviewer judgments, not visits. The authors reported that smaller open models also improved with retrieval (Klang et al., 2025). A more specific code is better only when the note supports it.
The comparison was with provider-assigned codes rather than a professional coding audit, and the study measured reviewer preference for the primary diagnosis code, not claim acceptance, payment, or audit outcomes. (Klang et al., 2025) The study suggests that giving a model real, locally used codes to choose from can produce primary diagnosis codes that physician reviewers prefer. It does not show that a model can code whole encounters without review.
AI-Generated Documentation Queries
A query asks a physician to clarify documentation, for example whether findings in the record support a diagnosis that has not been documented. In August 2026, the American Health Information Management Association (AHIMA) and the Association of Clinical Documentation Integrity Specialists (ACDIS) updated their joint practice brief on compliant queries. The brief is professional guidance from the two coding and documentation associations, not peer-reviewed research, and it states that it is not legal or compliance advice (AHIMA and ACDIS, 2026).
The 2026 brief states that technology-generated queries, including prompts, nudges, advisories, and alerts, are subject to the same compliance standards as manually constructed queries, and that organizations should not assume a vendor-supplied or EHR-integrated tool produces compliant queries by default. Its scope names computer-assisted physician documentation, computer-assisted coding, and large language models and generative AI platforms drafting or delivering query language. A communication counts as a query when it presents a physician with a specific diagnosis or documentation option for a specific patient encounter, however the software labels it (AHIMA and ACDIS, 2026).
Under the brief, a technology-generated query must include relevant clinical indicators sourced from the health record and cited with their location. Titles and subject lines visible to the physician must not name a desired diagnosis, reimbursement impact, or quality measure. Answer options must be clinically supported, must include “other” or room for another explanation, and must not use formatting, emphasis, or ordering that steers the answer. Sending a manual follow-up query solely to override a physician’s answer to an automated one, without new clinical indicators, is noncompliant (AHIMA and ACDIS, 2026).
Humans remain responsible: organizational process owners retain accountability for every query delivered to a provider, including those generated through automated or AI-assisted processes, and organizations that deliver such queries without pre-delivery human review should audit the output on a regular basis. (AHIMA and ACDIS, 2026) The brief warns that tools can produce leading or incomplete queries through configuration, training data, or design that prioritizes reimbursement over clinical accuracy.
Upcoding, Audit Risk, and Payment Integrity
A tool that finds more diagnoses or more specific codes can raise revenue, and it raises audit risk when the documentation does not support the code. In the code lookup benchmark, which used code descriptions without notes, wrong answers tended to be either too general or falsely specific (Soroush et al., 2024).
Federal oversight of Medicare Advantage shows what is at stake when diagnoses are added from sources other than a patient’s ongoing care. The evaluation below did not study AI and is included as a comparator. In Medicare Advantage, CMS pays plans more for enrollees whose diagnoses fall into payment categories called hierarchical condition categories (HCCs), a process called risk adjustment. A chart review is a plan’s retrospective review of the medical record to find diagnoses a provider did not submit. The HHS Office of Inspector General (OIG) found that diagnoses reported only on health risk assessments and chart reviews linked to them, and on no other 2022 service record, resulted in an estimated $7.5 billion in risk-adjusted payments for 2023, for 1.7 million enrollees. In-home assessments and their linked chart reviews accounted for $4.7 billion, or 63%, of that total (HHS OIG, 2024).
In describing chart reviews, OIG noted that vendors hired to review charts may employ staff with clinical or coding experience or may use artificial intelligence software, and that reviewers may add diagnoses they believe were missed. The evaluation did not measure AI use. OIG did not determine whether the diagnoses were supported by the medical record, and its concern was that either the payments were improper or enrollees did not receive needed care for the conditions reported. CMS concurred with one of OIG’s three recommendations. The same evaluation reports that, for fiscal year 2023, CMS identified $12.7 billion in net overpayments that resulted from plan-submitted diagnoses not supported by the medical record (HHS OIG, 2024).
Software that adds diagnoses after the visit does the same kind of work that federal oversight bodies already scrutinize, and its output will be judged by whether the note supports each code. Policy proposals on AI-driven coding intensity are discussed in AI-Assisted Clinical Documentation, and payer use of algorithms in Healthcare Policy and AI Governance.
Claims, Denials, and Appeals Agents
In a preprint benchmark of 135 administrative tasks in four simulated environments with synthetic patient data, the best agent, a Claude Opus 4.6 computer-use agent, completed 36.3% of tasks from start to finish and 25.0% of appeals and denials tasks; a run counted as successful only if every subtask was completed (Bedi et al., 2026, preprint). No source reviewed here measured denial or appeal outcomes from a deployed coding or claims agent. Evaluating such agents is covered in Evaluating AI Agents in Healthcare.
Evaluating an AI Coding Tool
The questions below adapt Evaluating AI Clinical Decision Support Systems to coding tools.
| Question | A good answer | Why it matters |
|---|---|---|
| 1. What exactly does the tool do? | Suggests codes, assigns them without review, drafts queries, or submits claims, stated separately | Each use carries different risk and different rules |
| 2. Was it tested on notes like ours? | Agreement with a blinded audit by certified coders on a sample of the organization’s own notes, by setting | Most studies used single-site datasets (Tangudomkit et al., 2026) |
| 3. How does it do on rare codes? | Results reported for rare and common codes separately | Systems did worse on rare codes in the review (Tangudomkit et al., 2026) |
| 4. Does every code cite its support? | Each suggestion links to the text that supports it | The review called for auditability and human oversight (Tangudomkit et al., 2026) |
| 5. Can it return a code that does not exist? | Output restricted to codes valid and billable on the date of service | 18.5% of GPT-4’s wrong ICD-10-CM answers were nonexistent codes (Soroush et al., 2024) |
| 6. Are its queries compliant? | Queries meet the AHIMA and ACDIS standards and are audited | The brief holds technology-generated queries to the manual standard (AHIMA and ACDIS, 2026) |
| 7. What happens to denials and audits? | Denial, rebill, and audit-finding rates tracked against a comparison period | The studies reviewed here measured code agreement or preference, not payment outcomes |
| 8. Who signs off? | A named coder or physician reviews each code before it is billed, with sampled audits | Process owners remain accountable for every query delivered, including AI-assisted ones (AHIMA and ACDIS, 2026) |
| 9. How is a rise in coded severity explained? | A review of any rise in coded severity or risk scores after go-live | Diagnoses reported only on health risk assessments and chart reviews drew federal scrutiny (HHS OIG, 2024) |
Check Your Understanding
A vendor proposes that its AI assign codes for emergency department visits without coder review, citing high accuracy on a public benchmark.
Question: What should the coding director ask first?
Answer: Ask for agreement with a blinded audit by certified coders on a sample of the department’s own notes, with missed and unsupported codes reported separately for rare and common codes. In a systematic review, 87.5% of automated ICD-10 coding studies remained at the research stage and most used single-site data (Tangudomkit et al., 2026). The best Mount Sinai emergency department result was reviewer preference for the primary diagnosis code, not an audit of every code on the claim (Klang et al., 2025). Start with suggestions that a coder accepts or changes, and track denials and audit findings against the period before go-live.
A hospitalist receives an EHR prompt that reads “Consider documenting acute blood loss anemia” with a single checkbox to accept it. The diagnosis is not yet documented.
Question: Is this a query, and is it compliant?
Answer: It presents a specific diagnosis for a specific patient, so under the 2026 AHIMA and ACDIS brief it is a query, whatever the software calls it. It fails two of the brief’s standards: its only visible line, which serves as its title, names the desired diagnosis, which the brief bars in titles and subject lines, and a single accept box works as a yes/no query, a format the brief says should not be used to obtain documentation of a new, previously undocumented diagnosis in any setting. A compliant version cites the clinical indicators and their location in the record, offers clinically supported options including “other”, and does not steer the answer through wording, emphasis, or a single option (AHIMA and ACDIS, 2026). The hospitalist should document only what the clinical picture supports and report the prompt to the documentation team for review.
Six months after a coding tool goes live, a medical group’s average risk score has risen and revenue is up.
Question: What should the compliance officer check?
Answer: In a sample audit, check whether added diagnoses are supported by the note and linked to care. In a federal evaluation that did not study AI, diagnoses reported only on health risk assessments and linked chart reviews, with no other service record, generated an estimated $7.5 billion in Medicare Advantage payments for 2023. OIG did not determine whether they were supported, and it raised concern that either the payments were improper or patients did not receive needed care (HHS OIG, 2024). A rise can reflect better capture of real conditions or unsupported codes; only the audit tells them apart.
Key Takeaways
Can AI do medical coding accurately?
The evidence reviewed here does not show it. A 2026 systematic review of 24 automated ICD-10 coding studies found that 87.5% remained at the research stage and concluded that, despite promising benchmark performance, evidence remains insufficient for real-world deployment (Tangudomkit et al., 2026). In a December 2023 benchmark that used code descriptions rather than clinical notes, the best language model, GPT-4, returned the exact ICD-10-CM code for 33.9% (Soroush et al., 2024).
Is there evidence for autonomous coding?
Not from the sources reviewed here. None evaluated a commercial product assigning codes without coder review in routine use, and in a 2026 systematic review 87.5% of automated ICD-10 coding studies remained at the research stage (Tangudomkit et al., 2026). An organization considering autonomous coding can measure agreement with a blinded audit by certified coders on its own notes before turning off review.
Do large language models code better with retrieval?
In one Mount Sinai case study of 500 emergency department visits, physician reviewers preferred the retrieval-supported GPT-4 primary diagnosis code over the provider-assigned code for accuracy in 447 judgments versus 277, and smaller open models also improved with retrieval (Klang et al., 2025). The study measured reviewer preference for one code per visit, not payment or audit outcomes.
Are AI-generated documentation queries allowed?
Under professional guidance, yes, if they meet the standards for manual queries. The 2026 AHIMA and ACDIS practice brief states that technology-generated queries, including those drafted by large language models, must meet every element of compliant query practice, and that humans remain responsible for every query delivered (AHIMA and ACDIS, 2026).
Does AI coding increase audit risk?
It can when a tool adds codes the note does not support. In a Medicare Advantage evaluation that did not study AI, diagnoses reported only on health risk assessments and linked chart reviews generated an estimated $7.5 billion in payments for 2023; OIG did not determine whether those diagnoses were supported or measure AI use (HHS OIG, 2024).
Who is responsible for an AI-generated documentation query?
The organization that delivers it. The AHIMA and ACDIS brief states that process owners retain accountability for every query delivered, including those generated through AI-assisted processes (AHIMA and ACDIS, 2026). Billing responsibility for AI-drafted notes is covered in AI-Assisted Clinical Documentation.
Conclusion
The evidence for AI in medical coding is early: a systematic review in which most studies never reached clinical testing, one benchmark and one retrieval case study from a single health system, professional guidance on queries, and a preprint benchmark of agents. Physicians and coding leaders can ask of each tool whether every code it suggests is supported by the note, how it performs on the organization’s own records, and what happens to denials and audits after go-live.
- Healthcare Administration: billing costs, patient messages, and prior authorization
- Hospital Operations and Management: the revenue cycle and operational tools
- AI-Assisted Clinical Documentation: billing rules for AI-drafted notes
- Healthcare Policy and AI Governance: payer algorithms and prior authorization rules
- Evaluating AI Agents in Healthcare: testing agents that work through portals
The Physician AI Handbook is free to read and cite, with no third-party advertising. If this chapter informed your work, you can support the handbook.