Can an AI Scribe Assign ICD-10 Codes? What to Check Before You Trust One

Yes, most AI scribes can produce ICD-10 codes. Some produce CPT codes too, and a suggested visit level.

That is not the useful question. The useful question is what the scribe does when the session it just transcribed did not actually establish a diagnosis — because that is most sessions, and because the code goes on a claim with your name on it, not the vendor's.

The failure mode nobody demos

A patient describes low mood, poor sleep and a stressful few months. You are thinking about it. You want to see them again in two weeks before you commit to anything. The note itself is straightforward — whichever format you use for that kind of visit handles it fine.

A scribe that codes off the transcript will read those symptoms and produce a depressive disorder code, because that is what the words map to. It is not hallucinating — the symptoms are genuinely there. It has simply skipped the part where a clinician decides.

Now the code exists in your note. It is plausible, it is specific, and if you are reviewing twenty notes at the end of a clinic it is very easy to leave in. From there it reaches the claim, the patient's record, and every future clinician who reads that record and sees a diagnosis that was never actually made.

The problem is not accuracy. The problem is sequence.

The four questions worth asking a vendor

1. What does it do when the session had no confirmed diagnosis?

Ask for a demo on a genuinely ambiguous consultation, not a clean one. Watch whether a code appears anyway.

2. Does anything else key off the diagnosis?

Coding is rarely alone. Clinical guideline references, patient education material and treatment suggestions are usually generated from the same diagnosis. If the diagnosis was inferred rather than confirmed, everything downstream inherits that inference — and a patient handout for a condition nobody diagnosed is a worse problem than a wrong code.

3. Is the visit level derived from what is actually documented?

A CPT visit level should reflect the complexity in the note. If it is derived from session length alone, it will drift from what the documentation supports, and that is precisely what an audit looks for.

4. Can you see and change it before it goes anywhere?

A suggestion you can edit is a tool. A code that is already attached to the encounter by the time you see it is a liability.

Diagnosis-gated coding

There is a straightforward design answer to all four: generate nothing downstream of a diagnosis until a clinician has confirmed the diagnosis.

That is how Cognivolt is built. The note, the mental state examination and the structured risk assessment are all generated from the session as it runs. Differential considerations are offered — ranked, with the evidence for each — because a differential is a prompt to think, not a conclusion.

But ICD-10 and CPT codes, clinical guideline references and patient psychoeducation are locked until you confirm the diagnosis yourself. Not a checkbox you can leave ticked from last time; the confirmation belongs to that session.

If a consultation ends without a diagnosis — and plenty do — you get the note, and no codes. Which is the correct output for that visit.

What this costs you

Honestly: a click, and a few seconds.

What it buys is that no code exists in your records that a clinician did not put there — and that the guideline reference and the patient handout are for the condition you actually diagnosed, not one the software inferred from symptom words.

The checks that are not the model's opinion

One more thing worth asking about, because it is adjacent and often conflated.

Allergy and drug–drug interaction checks should not be generated by a language model. A model asked "does this interact with that" will produce a fluent, confident answer, and it will sometimes be wrong in both directions — inventing an interaction that does not exist, and missing one that does.

In Cognivolt those checks are lookups against fixed, versioned reference tables. The result carries the table version, how many drug pairs were compared, and any drug it did not recognise. That last one matters: a check that silently ignores a drug it could not parse is worse than no check, because it reads as clearance.

It is not a licensed exhaustive database and does not claim to be. Absence of a finding is not clearance, and the software says so on every result.

The short version

An AI scribe assigning ICD-10 codes is normal and useful. An AI scribe assigning them before a clinician has confirmed the diagnosis is a documentation problem wearing a productivity label — and the clinician, not the vendor, signs the claim.

Ask what it does with an ambiguous session. That one question separates the two.

Cognivolt generates nothing downstream of a diagnosis until you confirm it — across sixteen specialties, from psychiatry and psychology to cardiology, dermatology and general practice.

Try it free for 14 days — no credit card required.