If you work in health and care

The sentence that matters most

Software intended for the diagnosis, prevention, monitoring, prediction, prognosis or treatment of disease is regulated as a medical device in the United Kingdom, the European Union, the United States and most other regulated markets — regardless of who built it, whether money changed hands, or whether it is described as a helper.

A general-purpose chatbot is not an approved device. Using one to decide a diagnosis, a triage priority or a dose is using an unapproved device for a clinical purpose, and no amount of care in the prompt changes what it is.

That leaves a large, genuinely useful space on the other side of the line, and the point of this lesson is to be exact about where it runs.

Where it genuinely helps

• Documentation. Drafting a note from a consultation you conducted, which you then read and correct. This is the largest single time cost in most clinical roles.
• Plain-language rewriting. Turning a discharge summary into something a patient and their family can act on, at the reading level they actually have.
• Translation of patient information, checked backwards.
• Summarising a long referral or record for your own orientation before you read the parts that matter.
• Rotas, letters, meeting notes, policy drafts — the administrative half of every clinical job.
• Searching your own guidelines, with the source quoted and the date checked.

Notice that none of these is a clinical decision. All of them are communication and administration around clinical decisions you make.

The failure modes specific to health

Dose and interaction errors. Confident, well-formatted, wrong. Never take a dose, an interaction, a contraindication or a paediatric adjustment from a general model. Use the formulary you already use.

Guidelines that changed. A model's knowledge has a cutoff and its answers carry no date. Clinical guidance changes, and the changed part is exactly what you would be looking it up for.

Propagation. This one is specific to health records and under-discussed. A wrong sentence in a note does not stay in that note — it is copied forward, quoted in the discharge letter, read by the next clinician as history, and becomes part of what everybody believes about the patient. An error in a health record has a much longer half-life than an error in an email.

The defence is not more prompting. It is that you read every note before it is signed, in full, as though a colleague had written it and you were responsible for it — because you are.

Ambient documentation and consent

Tools that listen to a consultation and draft the note are real, in use, and among the most promising applications in the sector.

Two known failure modes: transcription can generate fluent text during silence or noise, and speaker attribution can put the patient's words in the clinician's mouth or the reverse. Both produce a note that reads correctly and says something that was never said.

The evidence on benefit is early and mixed. Some studies report reduced documentation time and improved clinician experience; others find total time roughly unchanged with the work redistributed. It is honest to say that the case is promising and not yet settled, and dishonest to present a pilot's enthusiasm as a result.

Patients should be told that a recording or transcription tool is in use, and the consent lesson in the earlier block applies in full. In many places recording without agreement is not merely discourteous.

Patient data is the strictest case

Health data receives the highest level of protection in essentially every data protection regime: special category data under UK and EU law, protected health information under HIPAA in the United States, with equivalents elsewhere. On top of statute sits a professional duty of confidence that is older than any of it, and in some systems a national data opt-out that patients have exercised.

Pasting a patient's history into a consumer chat account is disclosure to a third party. It does not become acceptable because the name was removed — the redaction lesson explains why a case description is frequently more identifying than a name, and in a small community it certainly is.

The route that works is the organisational one: a tool your trust, board or practice has assessed and contracted for, under the same terms as the record system, with a named information governance owner. If that does not exist yet, the answer for patient material is a local model or nothing.

The free path, which is a real option here

whisper.cpp transcribes consultations on your own laptop with no network connection. A local model through Ollama drafts and rewrites with the audio and the text never leaving the machine. LibreOffice handles the documents. National guidance — NICE, your national formulary, your regulator's standards — is free to read and is the authority a model is not.

For a small practice or a care home without a procurement department, that stack is not a poor substitute. It is the version of this that your professional duties permit.