Automated medical coding: how AI makes clinical coding more reliable

The essentials in 30 seconds

Introduction

Coding stays and procedures is one of the most invisible tasks in a hospital, and one of the heaviest. Every hospital stay has to be translated into diagnosis and procedure codes so that activity can be described and funded, and this translation is done from clinical reports that are often long, incomplete or written entirely in free text. It is meticulous work, it happens after the care itself, and it rarely gets the attention it deserves, even though a large part of a hospital's funding depends on it.

For years the answer has been to add coding staff, train them again and again, and run retrospective quality checks. It works, but it is slow, it is costly, and it leaves a structural gap between what was actually done at the bedside and what ends up being coded. Automated coding assistance promises to narrow that gap, and this is exactly where artificial intelligence starts to be genuinely useful in a hospital, provided we are honest about what it can and cannot do.

Galeon has been building an AI-native electronic health record (EHR, in French DPI) alongside caregivers since 2016. It is used in 19 hospitals, including 2 university hospitals (CHU), across more than 3 million patient records, and it is hosted on HDS-certified infrastructure aligned with ISO 27001:2022. Coding assistance inside the care workflow is therefore not a slide for us, it is something we design and operate. That vantage point is also why we are cautious: the value of coding AI is real, but it is entirely conditional.

One thing to remember: AI coding assistance only pays off when the data it reads is structured, and when a human keeps the final say.

What exactly is automated medical coding?

Automated coding is, above all, coding assistance. From the clinical report of a stay, a language model proposes the most likely codes: principal diagnosis, associated diagnoses, comorbidities, procedures. Crucially, each suggestion is backed by the passage of text that justifies it, so the professional can see why a code is proposed and accept, correct or reject it in one glance.

The goal is not to code in the professional's place. It is to spare them from starting from a blank screen, and to flag what a tired eye, at the end of a long list of stays, might have missed. A forgotten comorbidity is not a detail: it changes how the complexity of a stay is described, and therefore how it is funded. Coding assistance is, in that sense, as much about completeness as it is about speed.

It is worth being precise about vocabulary. In France, coding feeds the PMSI (the medicalised information system) and, through it, activity-based funding (T2A). The codes are not an administrative formality bolted onto care: they are the language in which the hospital's activity is described to the outside world. Getting them right is getting the description of the work right.

How does the AI actually read a clinical report?

The engine behind coding assistance is natural language processing: the ability to read free text the way a human would, and to map what it finds onto a coding system. Concretely, the model detects the clinical entities mentioned in the report (a diagnosis, a procedure, a complication), understands their context (is this condition present, absent, suspected, or part of the patient's history?), and links them to the corresponding codes.

Context is where it gets hard, and where quality is won or lost. "No sign of infection" and "infection confirmed" contain the same word but mean opposite things for coding. A good model has to handle negation, uncertainty, temporality and family history without turning a ruled-out diagnosis into a billed one. This is exactly why human validation is not a nicety but a safeguard: the professional is there to catch the cases where the machine misread the nuance.

The other decisive factor is the raw material. A model reading a well-structured report, where diagnoses and procedures are captured cleanly at the point of entry, has an easy job. A model asked to mine a wall of unstructured free text, dictated in a hurry and never reviewed, has a hard one, and its suggestions will reflect that. The intelligence is not only in the algorithm; it is in the data the algorithm is fed.

Why is manual coding so costly?

Because it is time-consuming, repetitive, and entirely dependent on the quality of the report. A mis-coded procedure is revenue that is simply lost. A forgotten diagnosis is case complexity that is never reflected, so the stay looks lighter than it really was. Across a whole hospital, over tens of thousands of stays a year, these small gaps add up into a significant amount of time and funding.

There is also a hidden cost that rarely appears in a budget line: the mental load. Coding is often done under time pressure, sometimes weeks after the stay, from memory or from a report that has to be reread in full. It pulls skilled professionals away from higher-value work, and it is a frequent source of friction between clinical teams and the medical information department (DIM). Making coding faster and more reliable is not only a financial question; it is also a working-conditions question.

Where does AI coding assistance help the most?

Coding AI is not uniformly useful. It pays off fastest in specific, well-bounded situations, and it is honest to name them rather than promise a blanket revolution.

1. High-volume, repetitive stays

On standardised, frequent stays, suggestions are reliable and the time saved is immediate, because the model has effectively seen thousands of similar reports. This is where assistance earns its keep first.

2. Complex stays with many comorbidities

Here the value is less about speed and more about completeness: the model surfaces associated diagnoses that are easy to forget, each justified by the report. The professional still decides, but starts from a fuller picture.

3. Catching omissions before billing

Used as a second pass, coding assistance behaves like a checklist: it flags a documented diagnosis that was not coded, or a coded item that the report does not support. It reduces both under-coding and over-coding.

4. Onboarding and continuity

For a new coder, or when a team is short-staffed, suggestions backed by the source text shorten the learning curve and keep coding consistent, instead of depending entirely on one person's expertise.

5. Feeding cleaner casemix data

More consistent coding means more reliable casemix analysis downstream, which matters for steering activity and for dialogue with the funder. Better input, better analytics.

What does structured data change? Everything.

Coding assistance is only as reliable as the data it reads. This is the point most vendor demos quietly skip, and it is the one that decides whether the tool works in real life. An EHR that produces structured, usable data at the point of entry makes coding AI genuinely useful. An EHR that piles up free text makes the same AI fragile, because it is forced to guess.

That is why automated coding is, first and foremost, an EHR question, not an add-on question. If the record captures diagnoses, procedures and observations in a structured way while the clinician works, the model has clean signal to work with. If everything lives in dictated narrative, the model spends its energy reconstructing what should have been captured in the first place. This is the difference between a tool bolted on after the fact and intelligence built into the record itself.

Bolt-on coding tool versus coding built into a smart EHR: what changes?

There are two ways to bring AI to coding, and they are not equivalent. The first is a separate tool that reads exported reports and returns codes. It can help, but it inherits every weakness of the data it is handed, it adds another interface and another export, and its suggestions arrive detached from the moment of care.

The second approach builds coding assistance into the EHR itself. Because the record already holds structured data, the suggestions are better; because they appear inside the workflow the professional already uses, they are actually adopted; and because everything stays in one system, traceability is native rather than reconstructed. The lesson is simple: the quality of coding AI is decided upstream, by the quality of the record, far more than by the cleverness of the model bolted on at the end.

Is automated coding "high risk" under the AI Act?

It depends on the use, and it is worth resisting absolute statements in either direction. Coding assistance that suggests codes for human validation, without making an autonomous decision about a patient's care, does not sit in the same category as an AI that would decide a diagnosis or a treatment on its own. The level of obligation follows the level of autonomy and the potential impact on the person.

What does not change, whatever the classification, are the fundamentals: transparency about what the tool does, traceability of suggestions, human oversight, and protection of the data. A serious coding assistant is designed with those principles from the start, not retrofitted to a regulation after the fact. Treating the AI Act as a design constraint rather than a compliance afterthought is, in practice, what separates a tool you can trust from one you cannot.

What does it change for the DSI, the DIM and the clinician?

For the hospital CIO (DSI)

The main question is not "which coding model is best" but "does our record produce data clean enough for any coding model to work". The DSI's leverage is upstream: structured capture, interoperability, and hosting that meets HDS requirements. Choosing coding AI is, in reality, choosing a data strategy.

For the DIM and coding teams

The day-to-day changes from typing to reviewing. Coding becomes a validation task supported by justified suggestions, which lowers the workload on repetitive stays and frees time for the complex ones that genuinely need human judgement. The expertise of the DIM is not made redundant; it is redirected to where it matters.

For the clinician

The best coding assistance is the one the clinician barely notices, because it draws on data captured naturally during care rather than demanding extra entry. The clearer and more structured the report, the better the suggestions, which quietly turns good documentation into a shared benefit rather than a chore.

Limits and challenges to be aware of

• Garbage in, garbage out: on a poor report, suggestions are poor. No model compensates for missing information.
• No blind automation: human validation stays essential, especially on complex or atypical stays.
• Regulatory framework: coding responsibility and traceability must stay clear, and data protection is non-negotiable.
• Change management: the tool does not replace upskilling coding teams; adoption is a people question as much as a technical one.
• Honest expectations: the aim is accurate coding, not maximised coding. A tool sold on revenue uplift alone is a warning sign.

FAQ

Is coding assistance reliable enough to trust?
It speeds up and secures coding, but human validation remains the rule. Reliability comes from the pair of human and machine, not from the machine alone, and that is by design rather than a limitation.

Do we need to switch software to benefit from it?
Not necessarily, but the gain is proportional to how well the EHR structures data. A record built to capture structured data at the point of entry will always get more out of coding AI than one built around free text.

Does it code more just to earn more?
No, and it should not. The goal is coding that accurately reflects the real activity and complexity of each stay, which means correcting both under-coding and over-coding. Accuracy, not aggression.

Who is responsible for the codes?
The hospital and the professionals who validate them, exactly as today. The AI proposes; the human decides and signs off; responsibility stays where it has always been.

Where does it start paying off?
On high-volume, repetitive stays, where suggestions save the most review time, and on complex stays, where it helps avoid omissions. Those two ends of the spectrum are where the value shows up first.

What about patient data security?
Coding assistance handles sensitive health data, so it must run on infrastructure that meets health-data hosting requirements (in France, HDS certification) and follow strict access and traceability rules. Security is a prerequisite, not an option.

In summary

Automated medical coding is not about replacing humans. It is about giving them time back and making casemix coding more complete and more reliable, so that the description of a hospital's activity finally matches the reality of the care delivered. The suggestions are useful, but they are downstream effects.

The real lever sits upstream: an EHR that produces structured, usable data from the point of entry, with a human who keeps the final say and a hosting model that protects the data. Get that right and coding AI becomes genuinely valuable; get it wrong and no model will save you. That is exactly the logic of an AI-native EHR like Galeon, where the intelligence lives in the record, not in a tool bolted on at the end.

Read next: why clinical AI depends first on data quality.

Sources

• Haute Autorité de Santé (HAS)
• ATIH, medicalised information (PMSI) and activity-based funding
• European Commission, regulatory framework on AI (AI Act)
• Galeon, smart EHR vs traditional EHR