AI medical coding tools read clinical documentation and assign billing codes, with or without a human coder checking the result. Accuracy is strong for narrow, structured specialties and weaker for complex inpatient cases. The larger issue is liability: when an autonomous system gets a code wrong, the hospital or physician who submits the claim still answers for it.
Every coding vendor’s pitch deck has the same slide. A backlog of unbilled charts, a shortage of certified coders, days in accounts receivable creeping up. Then the promise: most charts coded automatically, in minutes, with accuracy “at or above human level.”
Some of that is true. Coding is repetitive, rule-bound and text-heavy, which is exactly the work modern language models are good at. But the slide leaves out the part that matters most to a hospital’s general counsel. A code is not just a label. It is a statement to a payer, often a government payer, that a service happened and was documented to a certain standard. Autonomous coding saves labour, but it does not move the legal responsibility for that statement. The software company writes the code. The hospital signs the claim.
That is the non-obvious point about AI medical coding. The accuracy debate is real, but it is secondary. The primary question is who absorbs the error when a machine that codes at scale is slightly, systematically wrong.
How accurate is AI medical coding?
The honest answer is that it depends heavily on what “AI” and “accurate” mean.
General-purpose language models, used naively, are poor coders. A study from the Icahn School of Medicine at Mount Sinai, published in NEJM AI in April 2024, took more than 27,000 unique diagnosis and procedure codes from a year of routine care, gave models from OpenAI, Google and Meta the official description of each code, and asked them to return the code. None of the models, including GPT-4, reproduced the original codes more than half the time. They often produced codes that were plausible, sometimes codes that did not exist.
That test was unusually hard in one way and unusually easy in another. It stripped away the clinical context a real coder uses, but it also asked for a simple lookup rather than the real job of reading a messy note and deciding what is billable. Its lesson is narrower than the headline: a chatbot is not a coding engine. Commercial autonomous coding products combine language models with code databases, payer rules, edit checks and confidence thresholds, and they route uncertain charts to humans.
Those products perform best where documentation is structured and the code set is small. Radiology reports, pathology, emergency department professional fees and routine outpatient visits are the usual starting points. Complex inpatient stays, where the principal diagnosis, complications and procedure sequence together determine the payment, remain harder, and most hospitals still keep human coders in that loop.
Vendors publish accuracy and automation rates, but those figures are usually self-reported, measured on the vendor’s chosen specialty and compared against the hospital’s own human coders, whose accuracy is itself imperfect. A hospital should treat them as marketing until an independent audit of its own charts says otherwise.
Can AI coding lead to upcoding?
Yes, and it does not need to be designed to.
A coding model trained to capture every documented, billable condition will, almost by definition, raise the average intensity of coded claims when it replaces human coders who were missing some of them. The hospital will call that “capturing appropriate revenue.” A payer will call it upcoding. Sometimes both are right, which is what makes this hard.
The risk rises when AI touches documentation as well as coding. Tools that prompt physicians to add specificity (“consider documenting acute on chronic heart failure”) or ambient scribes that produce fuller notes than a tired doctor would type can both shift coded severity upwards. None of that is fraudulent if the conditions are real and treated. It becomes a problem when the documentation describes more than the care did.
Regulators have seen this pattern before. In 2012, the US Attorney General and the Secretary of Health and Human Services wrote jointly to hospital associations warning that electronic health records were being used to clone notes and inflate billing. Electronic records were sold as an efficiency tool too.
The Medicare Advantage risk adjustment story shows where the scrutiny goes next. In October 2024, the HHS Office of Inspector General estimated that about $7.5 billion in 2023 Medicare Advantage payments were based on diagnoses reported only in health risk assessments and related chart reviews, with no other record of care for those conditions. A month earlier, the Justice Department announced that Cigna would pay $172 million to resolve False Claims Act allegations that it submitted inaccurate and unsupported diagnosis codes from in-home assessments. Neither case involved AI coding. Both show what happens when diagnoses are captured more efficiently than they are treated.
Who is liable when AI coding is wrong?
In the United States, the entity that submits a claim to Medicare or Medicaid certifies that it is accurate. Under the False Claims Act, liability does not require intent to defraud; acting in “reckless disregard” of the truth is enough. Providers also have a legal duty to report and return overpayments within 60 days of identifying them.
Put those rules next to an autonomous coding system and the shape of the risk is clear. If the software systematically assigns a higher-paying code in a certain situation, and the hospital deploys it without auditing, a prosecutor or whistleblower’s lawyer will ask whether ignoring that pattern was reckless. “The vendor said it was accurate” is not a defence written into the statute.
Vendor contracts rarely help much. Many include accuracy guarantees, but these usually mean refunding fees or paying for recoding, not covering repayments, penalties and legal costs, which can be many times larger than the licence. Some vendors offer broader indemnities. Read them closely; the exclusions are where the risk lives.
The same logic applies to physicians. A professional fee coded automatically still goes out under the clinician’s billing number. Most physicians never look at the codes attached to their visits. With autonomous coding, they are even less likely to.
A worked example: the chart nobody reads
Take a hypothetical 400-bed hospital that bills about 90,000 emergency department professional claims a year. It deploys an autonomous coding tool that codes 85% of charts without human review and sends the rest to coders.
In the first year, coding costs fall and billing lag drops from several days to hours. The share of visits coded at the two highest evaluation and management levels rises by six percentage points. Leadership reads this as the tool capturing complexity that stressed human coders were missing. Some of it probably is.
Now suppose that one percentage point of that shift comes from a single pattern: the model treating documentation of a reviewed old test result as evidence of independent interpretation, which supports a higher level. At an illustrative $60 difference per affected claim, that is roughly 900 claims and $54,000 a year. Small. But it runs silently for three years across every payer before an external audit samples the claims, extrapolates the error rate and asks for repayment across the whole population. By then, the question is no longer what the model got wrong. It is why nobody at the hospital noticed.
The numbers are invented for illustration. The pattern is not: small systematic errors multiplied by volume and time are exactly how coding audits turn into settlements.
Is the compliance risk overstated?
The strongest counterargument is that human coding is not a gold standard either. Human coders disagree with each other, get tired, fall behind and make random errors. Audits of human-coded claims find mistakes in both directions. An AI system that is consistent, logged and auditable could be easier to police than a team of people whose reasoning is never written down.
There is real force in that. A deterministic system that makes the same error every time is also a system whose error can be found, measured and fixed across every chart at once. Random human error cannot. Autonomous coding also creates a complete record of what was coded and why, which a good compliance team can mine.
The weakness is that this advantage exists only if someone uses it. Consistency cuts both ways: an unaudited system that is consistently wrong generates liability faster than any human team could. The counterargument describes what AI coding makes possible, not what most deployments do.
What this means for hospitals, coders and vendors
For hospital leaders, autonomous coding should be bought as a compliance programme with a labour saving attached, not the other way round. That means a baseline audit of human coding before go-live, an independent sample of AI-coded charts every month, specific monitoring of code-level distributions by specialty and payer, and a named person with authority to switch automation off for a code category that drifts. Our earlier piece on the questions hospitals ask too late when buying digital health applies with extra force here, because the downside arrives years after the contract is signed.
For physicians, the change is subtler. As documentation tools, including ambient scribes, write more of the note, clinicians sign text that drives codes they never see. The discussion in Ambient AI Scribes Will Reduce Burnout Only If Nobody Spends the Time They Save is partly a billing story for that reason. If the note says you did it, you need to have done it.
For coders, the job moves from coding charts to auditing machines: reviewing the charts the system is unsure about, sampling the ones it is sure about and spotting drift. That is more skilled work, and hospitals that cut their coding teams to the bone will find they have fired their auditors.
For vendors, the market will reward the ones that make external auditing easy: exportable reasoning for each code, drift reports and contracts that share real risk. Accuracy claims will matter less than evidence that a customer can check them.
Frequently asked questions
Can AI replace medical coders?
AI can already code a large share of simple, high-volume charts without human review, especially in radiology, pathology and emergency medicine. It has not replaced coders for complex inpatient cases, and it creates new work: reviewing low-confidence charts, auditing automated ones and managing denials. The likely outcome is fewer coders doing more specialised, audit-heavy work, not zero coders.
How accurate is ChatGPT at medical coding?
Used on its own, not very. A 2024 NEJM AI study from Mount Sinai found that general-purpose models, including GPT-4, reproduced the correct diagnosis or procedure code less than half the time when given the code’s description. Commercial coding tools add code databases, rules and human review, so general chatbot results should not be read as the accuracy of purpose-built systems.
Is AI-assisted coding considered upcoding?
Not in itself. Coding every condition that is documented and treated is correct coding, even if it raises payments. It becomes upcoding when codes are not supported by the record, or when documentation is inflated to support codes the care did not justify. Regulators look at patterns, so a sudden rise in high-level codes after deployment should trigger an internal audit.
Who is responsible if an AI coding tool makes an error?
In the US, the provider or organisation that submits the claim remains responsible for its accuracy under Medicare and Medicaid rules and the False Claims Act. Vendor contracts may offer limited guarantees or indemnities, but they rarely cover the full cost of repayments and penalties. Providers must also return identified overpayments within 60 days, whatever caused them.
What should hospitals audit after deploying autonomous coding?
At a minimum: a random monthly sample of automatically coded charts reviewed by certified coders, distributions of evaluation and management levels and diagnosis codes compared with pre-deployment baselines, denial and payer audit trends, and any code categories where automation rates changed suddenly. Results should go to compliance leadership, not only to revenue cycle managers.
Autonomous coding is a good idea wrapped around a liability nobody puts on the slide. The hospitals that come out ahead will be the ones that spend part of the labour saving on checking the machine.
Related on the GHC blog: Why Universal Health Coverage Stalls After the Pilot Phase and Paying for Outcomes Rents Behaviour. It Doesn’t Buy It..
If your organisation uses autonomous coding, who reviews the charts the system is confident about? We’d like to hear how you audit it.
