The back office of a hospital runs on codes. Every encounter becomes a set of ICD-10 diagnoses and CPT procedures; those codes become a claim; the claim becomes a payment, a denial, or an appeal. It is repetitive, high-volume, rule-bound work — which is exactly the profile that draws automation. Revenue-cycle and coding agents are among the most widely piloted AI systems in healthcare, and among the easiest to oversell, because a demo that codes ten charts perfectly says very little about the eleven-thousandth. This guide reads the published evidence on what these agents do well, where they fail, and why the failures are a compliance problem before they are a productivity one. As of July 2026.
The money at stake is in the denials
Start with why anyone builds these systems. A KFF analysis of federal Transparency in Coverage data found that insurers denied about 19% of in-network claims on HealthCare.gov marketplace plans in 2024 — roughly 85 million of about 451 million in-network claims — with insurer-level denial rates ranging from 1% to 54% 1. And the appeals machinery barely runs: consumers appealed only a small fraction of denied in-network claims 1. Every unappealed valid denial is revenue a provider earned and never collected, which is precisely the gap a denial-management agent is built to close.
The reasons those claims were denied matter for what an agent can fix. In the 2024 marketplace data, the documented denial reasons were led by unclassified "other" at 36%, then administrative reasons at 25%, lack of prior authorization or referral at 9%, and lack of medical necessity at just 5% 1. Read that distribution carefully: the clinical-judgment category is the smallest. Most denials turn on paperwork, process, and coding — the parts of the workflow an agent is genuinely suited to, and the parts where getting the detail right is the whole job.
Where revenue cycle is already automated
Automation in revenue cycle is not new; it is uneven. The 2024 CAQH Index shows the transactions with mature standards are already highly electronic — claim status inquiry at 81% and claim payment at 78% — while prior authorization lags at about 40% 2. The pattern tells you where the remaining opportunity, and the remaining difficulty, both sit. The easy, structured transactions are largely done. What is left is the judgment-heavy middle: turning free-text notes into correct codes, and turning a denial into a persuasive, accurate appeal. That middle is where large language models enter — and where their limits show.
How good are coding agents, really
The most rigorous public benchmark is unflattering. A 2024 study in NEJM AI tested large language models on medical-code querying and titled its finding plainly: the models are poor medical coders. The strongest system, GPT-4, reached exact-match rates of only 33.9% for ICD-10-CM, 45.9% for ICD-9-CM, and 49.8% for CPT, and the models "frequently produced imprecise or fabricated" codes 3. A clinical LLM that invents a plausible-looking code is doing the coding equivalent of a hallucination — and a fabricated billing code is a more dangerous object than a fabricated sentence, because it flows straight into a claim.
The obvious fix is specialization, and it works — with a catch. A 2025 study in npj Health Systems fine-tuned models on coding data and lifted exact-match accuracy from 3.35% for a base model to 97.48% in a controlled scenario. But when the same fine-tuned system met real-world clinical notes, exact match fell to 69.20% (with a broader category-match of 87.16%) 4. The authors' own conclusion is the sentence to carry out of this section: manual expert review remains essential, especially for complex and rare codes 4. Retrieval-augmented approaches — grounding the model in an actual code set rather than its memory, the technique described in our note on retrieval-augmented generation in healthcare — narrow the gap further, but none of the published work closes it to the point where a code goes out unread.
| Setting | System | Exact-match coding accuracy |
|---|---|---|
| General benchmark, CPT 3 | GPT-4, off the shelf | 49.8% |
| General benchmark, ICD-10-CM 3 | GPT-4, off the shelf | 33.9% |
| Controlled scenario 4 | Fine-tuned model | 97.48% |
| Real-world clinical notes 4 | Fine-tuned model | 69.20% |
The table's lesson is the distance between the second-to-last row and the last one. Lab accuracy is a promise; real-note accuracy is the delivery, and it is thirty points lower. Any coding-agent claim should be read as a question: measured on what — curated cases or the messy documentation your clinicians actually write?
One more distinction decides how much comfort to take from a high number. The same 2025 study reported a category-match of 87.16% alongside its 69.20% exact match on real notes 4. Category match means the agent landed in the right family of codes; exact match means it landed on the billable code. Reimbursement, audits, and medical-necessity edits run on the exact code, so the exact-match figure is the one that predicts revenue and risk — a system that is "usually in the neighbourhood" is a system that is usually submitting a code a coder still has to change. Vendors who lead with the looser metric are describing a research result rather than a billing outcome.
Why a wrong code is a compliance event
This is the part vendors underplay. In most software, a 69% accuracy rate is a UX problem you paper over with a good interface. In coding, the wrong output is a billing-integrity event. A code the documentation does not support — whether it inflates the bill or simply mismatches the record — is the raw material of an improper payment, and improper payments draw audits, recoupment demands, and, at the extreme, False Claims Act exposure. An agent that "frequently produced imprecise or fabricated" codes 3 is, from a compliance officer's chair, a system that manufactures audit risk at scale unless every output is checked.
That reframes the entire evaluation. The headline question is not "how accurate is the agent?" but "can you prove, for any given claim, what the agent proposed, what a human changed, and why?" An autonomous coder without a per-code audit trail and a human attestation is a liability even at high accuracy, because the errors it does make are the kind regulators investigate. The human in the loop in coding is not there to catch typos; they are there to keep the organization out of a billing-integrity finding.
It also changes what "good" looks like. In most machine-learning problems you optimize for average accuracy; in coding you care disproportionately about the shape of the errors. A model that quietly upcodes — assigning a higher-paying code the record does not support — creates far more exposure than one that undercodes or flags uncertainty, even at the same overall accuracy. The safest coding agents are therefore built to abstain: to route low-confidence cases to a human rather than guess, and to surface the documentation gap rather than paper over it. An agent that always produces a confident code is not more capable; it has simply moved its errors somewhere harder to see. Ask any vendor how their system behaves when the note does not clearly support a code — silence on that question is an answer.
Denial-management agents and the appeal backlog
The other half of revenue cycle is working denials, and here the case is stronger — because the alternative is often nothing at all. With consumers and providers appealing only a sliver of denials 1, an agent that drafts a structured, cited appeal for a denial that would otherwise be written off is capturing value that was being left on the floor. The same drafting quality that makes these systems credible on prior-authorization letters applies here, and the same limit does too: an appeal has to be true and complete, rather than merely fluent, and the clinical facts it asserts have to match the record. The pattern is identical to the one in our prior-authorization agents guide — agents are good at assembling and drafting, and a human still owns the assertion.
The federal backdrop is shifting the workflow underneath all of this. The CMS-0057-F final rule requires payers to give a specific reason for each denial and to report prior-authorization metrics publicly beginning in 2026 6. A specific, machine-readable denial reason is exactly the input a denial-management agent needs to route and rebut efficiently — the rule is, incidentally, making the payer side more legible to the provider side's agents.
A capability-and-limits summary
| Task | What agents can do now | Where a human stays required |
|---|---|---|
| Suggest codes from a note | Draft ICD-10/CPT candidates quickly | Verify each code against the documentation 4 |
| Autonomous coding | Handle clean, common cases in the lab | Own complex and rare codes; keep an audit trail 34 |
| Draft a denial appeal | Produce a structured, cited rebuttal | Confirm every clinical claim is true and supported 1 |
| Reconcile claim status | Read structured payer responses at scale | Sign off on adjustments and write-offs 2 |
| Report metrics | Compile the reporting the rule requires | Attest to accuracy 6 |
How to read this
Three cautions travel with the numbers. First, the coding benchmarks measure code selection on cases, not end-to-end billing accuracy in your system, and the real-note figure 4 is the one that predicts production behaviour — not the controlled one. Second, the denial statistics come from the marketplace segment 1; Medicare, Medicaid, and commercial books differ, so treat 19% as a well-measured example rather than a universal rate. Third, the administrative-burden figure of 13 hours per physician per week 5 explains the demand for automation but says nothing about whether a given agent reduces it safely. As with every claim in this space, the method for appraising the underlying studies is in our guide on how to read an AI validation study, and the broader field numbers sit in our AI in healthcare statistics.
The through-line is simple. Revenue-cycle and coding agents are strongest exactly where the work is voluminous, structured, and currently neglected — status reconciliation, first- draft coding, and appeals that would otherwise never be filed. They are weakest, and most dangerous, at the point where a proposed code or claim becomes a submitted one, because that is where an ordinary error becomes a compliance event. Build for that point: measure on real notes, keep the audit trail, and keep the human who signs.
Sources and method
This guide draws on a KFF analysis of CMS Transparency in Coverage data for the denial and appeal figures 1, the 2024 CAQH Index for transaction-automation adoption 2, a 2024 NEJM AI benchmark of large language models on medical-code querying 3, a 2025 npj Health Systems study on domain-specific fine-tuning for coding 4, the AMA's late-2024 survey for administrative-burden context 5, and the CMS-0057-F final-rule fact sheet for the denial-reason and metrics requirements 6. Every figure is tied to the primary source cited beside it and was checked live as of July 2026. We revisit this page on a 180-day cycle and whenever a new coding-accuracy study or denial dataset is published.