How an AI scribe works, and where it stops

Recording, transcription, drafting and signing, explained for a doctor deciding whether to trust one.

Four steps

Recording: the consultation is recorded on a phone or tablet, with the patient's consent, and stored encrypted. Transcription: speech becomes text, with the speaker marked, in the language spoken. Drafting: a language model reads the transcript and writes the note in the structure the hospital uses: complaints, examination, diagnosis, prescription, tests, advice. Signing: the doctor reads, corrects and signs. Only then does a record exist.

What the model is given

The transcript, the specialty and the structure of the note. It should not be given the patient's name, number or identifiers; they add nothing to the draft and they should not leave the hospital. Ask any vendor to show you the exact request.

What it is good at

Turning a spoken consultation into a complete, well-ordered note. Catching the medicine, dose and duration the doctor said. Writing the how-to-take lines in the patient's language. Producing the same structure every time, which is what audits and claims need.

Where it stops

It does not diagnose, and it should not suggest a treatment the doctor did not mention. It can mishear a drug name or a dose, which is why the review screen shows each line's origin and warns about allergies and missing fields. It does not replace the doctor's judgement; it replaces the doctor's typing.

Questions to ask before trusting one

Can I see which lines the AI wrote and which I changed? What happens to the recording afterwards? Can the signed note be altered later, by anyone? Does it understand my patients' languages, mixed in one conversation? What does it send outside the hospital?

If a vendor cannot answer all five in plain language, keep asking.