What GAO Says About AI Medical Coding and Coder Oversight

What does GAO's spotlight say about coder work and oversight?

The Government Accountability Office’s *Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)* says coder-reviewed code suggestions are widespread, while autonomous code assignment is emerging. Its warning that such systems could replace coders describes a possible capability, not measured displacement: the spotlight gives no employment outcomes or personal job-loss estimate. A U.S. medical coder should therefore decide what to do next by checking how their own employer allocates coding, review, exception handling, and final sign-off. The evidence below separates what tools can do, where workflows use them, and what is known about work consequences. For a career decision, the key local facts are the tasks assigned, the review path, and the responsibility retained by staff.

Sources: Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)

What exactly changes between a code suggestion and code assignment?

The workflow distinction begins before coding. In *Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)*, an ambient scribe listens during a clinical encounter and drafts a note for the clinician to review; it supports documentation, while the clinician remains responsible for the note. A separate post-visit coding tool processes the record and proposes billing or diagnosis codes for a coder to review. GAO describes those reviewed suggestions as widespread. Generative or agentic systems that assign codes without that coder handoff represent the emerging step: the tool moves from proposing a candidate to producing an assignment. These are different points in a workflow, with different human checks.

| Work step | System handoff described by GAO | Human check visible in the account | What the description leaves open | | --- | --- | --- | --- | | Encounter documentation | Ambient scribe drafts a note from the visit | Clinician reviews the note | Whether a given site uses a scribe or how often it changes later coding | | Post-visit code selection | Tool suggests codes from the record | Coder reviews suggestions | How much review, correction, or exception work each case requires | | Autonomous assignment | Generative or agentic system assigns codes | No coder review is inherent in the described capability | Whether a particular employer has adopted it, with what safeguards, or at what scale |

The exposed step is code selection from documentation, but a medical records specialist’s task inventory is broader. O*NET’s *29-2072.00 Medical Records Specialists* lists reviewing records for completeness, maintaining confidentiality, retrieving records for authorized users, checking compliance with regulations, and resolving unclear or incomplete information. Those duties point to work around the code itself: locating the relevant record, noticing a missing detail, clarifying conflicting information, and maintaining compliant handling. The inventory is useful for asking where a local workflow has handoffs and exceptions; it does not say how much time any task takes or assign an automation score to it.

An illustrative workflow makes the boundary concrete, without claiming to describe a studied workplace: a coder receives a chart, checks that the documentation is complete, reviews a suggested code, flags an unclear diagnosis for clarification, and records the final disposition under the employer’s process. A tool might assist with the suggestion while leaving the clarification and disposition steps to staff—or a local design might distribute those duties differently. O*NET names relevant kinds of work, but it cannot establish that every coder performs every listed task, or that code assignment is a small share of an individual’s role.

That distinction matters in both directions. An occupation-wide list can make coding appear to be just one element among several, yet a particular job may center almost entirely on assigning codes. Conversely, a title such as medical coder does not tell you whether the job also includes record retrieval, completeness checks, compliance work, or clarification. The task list supplies no task-hour shares, site-level workflow map, or measure of which duties a tool currently handles. To understand a specific role, compare the steps in your own recent cases with the actual system handoffs and identify who reviews proposals, resolves exceptions, and records the final decision.

Sources: Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116); 29-2072.00 Medical Records Specialists

How much adoption evidence is actually available?

The available evidence supports the conclusion that AI is being used in clinical documentation and coding-related work, but it does not establish what share of medical coders or facilities use coding tools. The Government Accountability Office’s *Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)* describes coder-reviewed code suggestions as “widespread.” That is a qualitative characterization: the two-page spotlight gives no coder sample, facility denominator, adoption rate, or time series behind the term. It indicates that GAO regards reviewed suggestions as established in practice, but it cannot be converted into a percentage of the coding workforce or a count of hospitals.

A separate measurement comes from *AMA Augmented Intelligence Research: 2026 Physician AI Survey*. The American Medical Association reported that 28% of surveyed physicians in 2026, compared with 21% in 2024, said they used AI for a combined category of billing-code, medical-chart, or visit-note documentation. The 2026 survey covered 1,692 U.S. physicians and was fielded from January 15 to February 2, 2026. The increase is a useful directional signal that physicians report more AI use in work connected to documentation and coding. It is self-reported physician use, however, rather than observed system logs or a census of organizations.

The category boundary matters. Because billing-code work is grouped with chart and visit-note documentation, the 28% figure does not isolate code selection, code review, or autonomous assignment. It also cannot tell us whether a respondent personally used a tool, whether a whole department adopted one, how many claims passed through it, or how often AI-generated codes were accepted. A physician’s report of documentation use is not interchangeable with a coder’s report of coding-tool use: the occupations may encounter different tools and different workflow steps. Nor does a respondent percentage describe facility prevalence; one hospital could have many users and another none, without the survey result revealing that distribution.

The two sources therefore answer adjacent questions rather than measuring the same population. GAO offers a qualitative assessment specifically about coder-reviewed suggestions but no denominator. The AMA provides a quantified comparison over time for physician-reported use across a broader combined documentation category, but does not isolate coders or facilities. Read together, they make real use in related workflows credible while leaving employer-level prevalence unanswered. The survey remains informative as a signal of reported uptake; keeping its limits visible prevents that signal from being misreported as the share of coders, hospitals, claims, or AI-coded cases. Neither source establishes how much staffing, review time, or employment changed as a result of adoption. Those outcomes require evidence about a specific employer’s workflow and workforce, which these adoption descriptions do not provide.

Sources: Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116); AMA Augmented Intelligence Research: 2026 Physician AI Survey

What did the Taiwan live-workflow trial actually test?

*Evaluating real-world deployment of an HL7-CDA-aligned LLM for ICD-10-CM coding* reports a 13-week pragmatic randomized pilot at Kaohsiung Medical University Hospital in Taiwan. The live trial analyzed ten certified coding specialists and 10,688 coded summaries. It compared manual coding with three human-in-the-loop AI-assisted workflows: specialists used system suggestions as part of the coding process rather than handing final assignment to an autonomous system. Across the three assisted workflows, the study reported faster completion than manual coding, with accuracy maintained in the live comparison. The result is a measured time-and-accuracy comparison in an operating hospital workflow, not a demonstration that the software can replace the people who reviewed its output.

The counterfactual was manual coding by specialists in the same study setting. That makes the comparison more informative about workflow time than a vendor demonstration or a standalone model benchmark: the reported advantage concerns coding work performed with human specialists involved, compared with the study’s manual process. The study’s 10,688 summaries provide the volume of coded material evaluated, while the ten analyzed specialists describe the participating workforce. Those figures are not interchangeable. A large number of summaries does not turn ten specialists at one center into a representative sample of coders, and the trial does not estimate how every specialist would perform under another organization’s rules, records, or software.

The reported maintained accuracy also needs to be read within that comparison. It means the authors reported no loss of accuracy for the AI-assisted live workflows relative to the study’s manual process under their evaluation; it is not a universal accuracy guarantee, and it does not show that every proposed code was correct or required no correction. The trial tested a specific implementation aligned with HL7-CDA and ICD-10-CM in one medical center over 13 weeks. Its design supports a real workflow finding, but the institution, documentation environment, coding rules, implementation, and participant group bound what can be carried to another workplace. Randomization within this implementation strengthens the comparison between its assisted and manual conditions; it does not remove those limits on generalizability.

For an individual coder, the practical implication is narrower than a claim about headcount. Faster assisted coding could leave more of a shift for checking suggestions, resolving exceptions, or handling other duties if the employer assigns time saved that way. That is an inference about possible task reallocation, not an outcome measured by *Evaluating real-world deployment of an HL7-CDA-aligned LLM for ICD-10-CM coding*. The reported time advantage does not tell us whether the hospital reduced staffing, increased throughput, changed workloads, or kept staffing and used the time elsewhere. The trial therefore shows that human-in-the-loop assistance can speed a particular coding workflow while reporting maintained accuracy in that setting. It does not test autonomous replacement or establish a staffing reduction, and another institution would need to assess its own workflow before assuming the same result.

Sources: Evaluating real-world deployment of an HL7-CDA-aligned LLM for ICD-10-CM coding

What did the separate TMMH validation reveal about transfer?

The separate TMMH evaluation in *Evaluating real-world deployment of an HL7-CDA-aligned LLM for ICD-10-CM coding* shows that transfer was uneven, not uniformly absent. In this retrospective external validation, the researchers evaluated models on a distinct hospital dataset without additional retraining for that validation. Performance fell across the evaluated models, but the size of the decline depended on model architecture. The result therefore gives a coder and an implementation team a more specific warning than “AI does not transfer”: a result from one institution may change at another, and some model families retained useful performance in this test.

The contrast was especially sharp for the non-decoder BERT model. Its reported precision went from 0.875 at the source hospital to 0.137 on TMMH. Precision concerns the proportion of model-selected codes that were correct under the study’s evaluation; the large change signals that, in the external dataset, a proposed-code stream from this model would require much more scrutiny to avoid incorrect selections. It does not mean that every code was wrong, nor does it describe the performance of every BERT system in every coding department. It is one model’s result under the paper’s cross-institution comparison, using the study’s data and evaluation setup.

The decoder results complicate any blanket claim of failure. The paper reports F-scores of 0.60 to 0.75 for the evaluated 7B decoder models on TMMH, and BioMistral reached a full-code F-score of 0.620. An F-score combines precision and recall, so it summarizes a different balance of correct proposed codes and codes found than precision alone; the values should not be compared as though they were the same metric. These scores are not a guarantee of production readiness, and they do not erase the declines relative to the source setting. They do show that some models retained measurable coding performance on this second site rather than collapsing to no useful signal.

The datasets differed in more than institutional name. The paper describes TMMH as having a different documentation style and a different degree of section completeness, alongside different institutional code distributions. The authors discuss documentation variability as a plausible reason that a model trained or evaluated in one environment may behave differently in another. That explanation fits the observed change, but the comparison does not isolate documentation format as its cause: case mix and code distribution also differed, and the validation was not an experiment that changed only one documentation feature while holding all else constant. The result establishes a transfer problem in this comparison, not a causal estimate for each source of it.

For a workplace deciding whether to use a coding model, the practical inference is to treat outside performance as a starting hypothesis and check it against the local record and code mix. That does not require assuming that every new site will reproduce the BERT decline, or that a stronger retained score will carry over unchanged. A local evaluation can examine whether the model’s errors cluster in particular codes, documentation patterns, or incomplete sections and whether the intended review workflow can catch them. The National Academies’ *An Artificial Intelligence Code of Conduct for Health and Medicine: Essential Guidance for Aligned Action, Chapter 2 Background* offers only a general implementation frame here: capability and downstream effects depend on deployment context. The coding-specific evidence is the study’s contrast itself. It weakens both easy conclusions—that one successful site proves broad transfer, and that another institution necessarily makes the system unusable—and makes the intended setting part of the performance claim.

Sources: Evaluating real-world deployment of an HL7-CDA-aligned LLM for ICD-10-CM coding; An Artificial Intelligence Code of Conduct for Health and Medicine: Essential Guidance for Aligned Action, Chapter 2 Background

What does a performance headline establish—and what can it not show?

The Government Accountability Office’s *Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)* reports that one developer said a deployment in a five-hospital health system achieved greater than 95% code accuracy and reduced annual emergency-department coding costs by more than $1 million. This is a concrete report of operational use and a claimed result worth investigating. The attribution matters: GAO is reporting what the developer told it, not independently certifying the performance or savings. The spotlight supplies no method details that would let readers reproduce the figures or compare them directly with another system.

“Accuracy” is not self-explanatory. To interpret the percentage, a coder would need to know its denominator and unit: was the measure calculated per code, per encounter, or across a different set of decisions? They would also need the rule for what counted as correct, how partial matches or multiple codes were treated, and whether the score included all cases or only a selected subset. Without those details, two systems can both advertise 95% while measuring different tasks and error types. A high aggregate figure can also conceal a lower score in a consequential or less common case category. These are questions about the meaning of the reported number, not evidence that the developer’s number is false.

The cost claim leaves a separate set of unknowns. The spotlight names annual emergency-department coding costs and a five-hospital system, but does not describe the baseline period, which expenses were counted, how implementation and ongoing system costs were handled, or whether volume and case mix changed. Nor does a lower cost say what happened to staffing. Savings could reflect changed labor hours, greater throughput, avoided outsourcing, a shifted budget, or another accounting choice; the reported figure does not identify the mechanism or establish that coder positions were eliminated. Performance, cost accounting, and employment outcomes are distinct claims, and evidence for one does not settle the others.

As editorial reasoning—not a GAO checklist or a validated assessment—a local coder or manager could ask: What cases and decisions form the denominator? Which errors count, and how are exceptions reported? What manual or prior-system baseline is used? Which labor, software, implementation, and rework costs enter the comparison? Did staffing, overtime, outsourcing, or case volume change over the same period? The answers would make a headline more interpretable in that workplace, even if some information is unavailable. A named five-hospital deployment is still a useful lead for inquiry; the missing method details limit what can be generalized, without disproving the reported claim. A useful comparison would also state whether the system was assisting a coder or making assignments without that review.

Sources: Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)

What does accountable review require when a result is challenged?

Accountable review requires a way to reconstruct what happened in a disputed case: the source documentation available to the system, the code or other result it proposed, any correction made by a person, and the final handoff or decision. That capacity is auditability. It matters beyond a headline accuracy rate because an aggregate score cannot explain a particular disputed result or show who had the opportunity and responsibility to correct it. This is a practical interpretation of the record-retention problem described in the Government Accountability Office’s *Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116)*, not a universal technical specification.

GAO reports that independent accuracy studies of AI coding tools are few and says some tools may not retain recordings or transcripts that outside evaluators would need to assess performance. If the underlying encounter or documentation trail is unavailable, a reviewer may be unable to determine whether a questionable code followed from missing source material, a tool error, a later correction, or a handoff that failed. The point is not that every tool lacks records; GAO identifies this as a concern about some tools. Nor does it establish that any particular dispute could be resolved by a transcript alone. It shows why the ability to inspect relevant inputs and actions can affect whether a performance claim or an individual exception is independently examinable.

The stakes can reach beyond the coding screen. GAO notes that inaccurate coding could affect medical records and reimbursement, while a tool that identifies more diagnoses could also change payments and overall health spending. The direction and scale of any net effect remain unsettled in the spotlight. Traceable review therefore helps connect a disputed output to the record and the responsible workflow step; it does not, by itself, prove that the output was correct or establish the broader financial effect. Responsibility may involve several people and systems, so a useful record should make the actual handoff understandable rather than imply that one coder controlled every upstream input or downstream use.

In *AHIMA Comments to CMS on the CRUSH RFI*, the American Health Information Management Association advocates human oversight and inspectable evidence about coding-tool activity and accuracy. That document is an interested professional association’s policy submission, not a law, empirical evaluation, or validated oversight standard. It signals a stakeholder position on governance; it cannot establish that one particular review design produces better results. GAO likewise raises the question of suitable oversight without prescribing one level for every workflow. A proportionate local inquiry, offered here as editorial reasoning, is who can inspect a challenged case, see the relevant source and correction history, make or request a correction, and confirm the final disposition. The answers can reveal whether responsibility is traceable in that setting, without assuming every team needs identical logs or review thresholds.

Sources: Science & Tech Spotlight: AI for Medical Notes and Coding (GAO-26-109116); AHIMA Comments to CMS on the CRUSH RFI

How should occupational outlook data affect one coder's choice?

The national outlook is relevant context, but it does not settle an individual medical coder’s prospects. The U.S. Bureau of Labor Statistics’ *Medical Records Specialists, Occupational Outlook Handbook* projects employment in the broader medical records specialist occupation to grow 8% from 2025 to 2035, from 200,700 to 216,300 jobs. That is a projected net increase of 15,600 positions across the occupation. Separately, BLS estimates about 14,000 average annual openings. The annual openings figure includes opportunities created when workers leave the occupation or workforce and need to be replaced, as well as openings associated with growth; it is not another way of stating the net increase.

These measures are meaningful counterevidence to a claim that the national occupation is simply projected to disappear. They describe the expected balance of employment and the annual flow of openings, however, not how work will be distributed among coding specialties, employers, or locations. The profile notes that AI may affect efficiency and demand for medical records specialists but does not quantify AI’s contribution to the projection. A growing occupation can still redesign particular tasks, and national net growth can coexist with a local employer reducing or reorganizing a specific team. The forecast therefore neither makes task change irrelevant nor supplies an individual probability of remaining employed.

For a coder deciding whether to move, use the projection to frame what to investigate locally. Check current openings in reachable locations, the duties attached to those roles, required credentials or experience, and whether the pay range and schedule fit your constraints. Compare those opportunities with the actual direction of your present job: whether coding suggestions are reviewed, what exception or record duties remain, and whether the employer is changing staffing or workload. BLS’s occupation-wide forecast cannot answer those employer-specific questions, and a listing does not establish that a particular candidate will be hired. Separate openings that require credentials you already hold from those that demand additional training, and count the time and cost of that preparation against your available resources. A national opening may be inaccessible if relocation, schedule, caregiving, health, or salary needs make its conditions unsuitable. It can help keep the decision proportionate: investigate realistic in-role or adjacent opportunities before treating either AI task change or national growth as a complete personal forecast.

Sources: Medical Records Specialists, Occupational Outlook Handbook

What should you check before choosing your next move?

When the national account does not answer what is happening in your workplace, take a small, reversible step before committing to a major change. First, map the tasks you performed recently: which involved repeatable documentation or code suggestions, which required resolving exceptions, and where your experience or judgment affected the final work. This is a snapshot for a decision, not a forecast of your job’s future.

Next, ask what is actually changing. Are AI steps being proposed, piloted, or assigned as part of the workflow? For a process question your manager can answer, ask who reviews suggestions, who handles exceptions, and whether the source material, corrections, and final decisions can be traced. Use work artifacts you are allowed to consult and manager-approved questions; confidential records or protected system details are not prerequisites. If you cannot establish how the tool is used or what responsibilities remain, record that as an open question rather than treating a headline as a local finding.

Then compare three feasible paths: deepen a skill in your current role, move to an adjacent role, or make a larger change. Check actual openings and prerequisites against your salary floor, reachable location, training time and cost, health, and family responsibilities. If those facts are missing, gather them before an irreversible resignation, expensive course, or degree commitment. A small course correction can preserve options while you learn more; it does not promise a particular outcome.

The free task checker can help organize task-level change-pressure signals and first actions. The paid career roadmap can compare scenarios against your preferences and constraints. Neither supplies a validated probability of displacement or guarantees employment or income. After this local inquiry, use the checker to clarify your task mix: **Check my task exposure.**

Questions readers ask