What AI should and should not do in a clinical note
AI is genuinely good at turning speech into a structured note and genuinely unfit to attest that you reviewed a medical history. The line between those two is the most important design decision in any AI charting tool — here is where we draw it, and why.
The interesting question about AI in clinical documentation is not whether it works. It plainly does, at the specific job of turning a spoken account of an appointment into an organised note. The question is which parts of a chart note it should be allowed near — because a note contains two kinds of sentence, and they have very different tolerances for a machine.
What are the two kinds of sentence in a note?
Observations, and attestations. An observation reports something about the patient: the tissue looked like this, the fit was that, the patient reported soreness on the left. An attestation reports something about you: that you reviewed the medical history, that you obtained consent and explained the alternatives, that you checked the appliance before it left.
Observations can be dictated and shaped by software, because you said them and the software is arranging them. Attestations cannot be produced by anything except the clinician, because they are claims about what a specific person actually did — and their entire value in a record, months later, is that someone stood behind them.
What is AI genuinely good at here?
The tedious middle: turning what you said into what the chart needs to look like. That is not a small thing — it is most of the after-hours work in a denture practice:
- Structuring dictation. You describe an exam in the order you did it; the note comes back in the order a note is supposed to be in.
- Consistent terminology. The same finding described the same way on Tuesday as on Friday, which is what makes a chart readable a year later.
- First drafts of correspondence. A referral letter or consult report is largely a re-presentation of information already in the chart, for a different reader — a task where a draft you edit is enormously faster than a blank page.
- Reading assistance on radiographs. As a second look that flags what to examine, feeding your judgement rather than replacing it.
None of these require the model to know anything you did not tell it. That is the common property of the safe tasks.
Where does it go wrong?
A model asked to produce a complete-looking note will complete it. That is what the technology does. Ask for a chart entry and give it eighty percent of an appointment, and it will supply a fluent, correctly formatted, entirely plausible remaining twenty percent — drawn from what usually appears in notes like that, not from your appointment.
This is dangerous precisely because it does not look dangerous. There is no garbled sentence to catch your eye. The invented line is the most normal-sounding one on the page, and it will still be there in three years when someone reads the chart back to establish what happened. A record's only job is to be true; fluency is not a proxy for it.
The mitigation is structural rather than a matter of prompting more carefully: do not let the model assert anything it was not told. Attestations are collected as explicit confirmations from the clinician, not generated as text. Gaps stay gaps, visibly, so a human fills them.
Does a clinician still have to read every note?
Yes, and any tool suggesting otherwise is selling the wrong thing. The time saving is real and it comes from not typing — not from not reading. Reviewing a good draft takes a fraction of the time composing one does, and that is the whole honest pitch.
The software should make that structurally true rather than merely advisable. In DentureFlowPro, AI output is a draft until a clinician acts on it; denturist notes lock on entry, notes written by other staff enter a review queue and are locked by the denturist, and the note locks on the clinician's action rather than the model's. Corrections append and nothing is backdated. Authorship stays with the person who signs.
When does the AI run at all?
Only when someone asks it to, on the record in front of them. There is no background processing sweeping your database, which matters for two reasons: it bounds what is ever sent for processing to what a staff member deliberately submitted, and it keeps the audit story simple. Ours runs on Google Cloud Vertex AI in the same Montréal region as the rest of your data, under a signed enterprise agreement — patient information is not used to train models, ours or anyone else's.
The short version
Let AI do the transcription, the structure and the first draft. Never let it do the attesting. Keep the signature — and the lock — attached to a person, and keep the model's output labelled as a draft until that person acts.
How this works in the product is described on our privacy and PHIPA page, and you can watch it happen in the demo videos.
Quick answers
Can AI write a clinical note for a denturist?
AI can draft one. It can structure dictation, apply terminology consistently and lay a note out in the expected format. What it cannot do is attest — state that a history was reviewed or consent explained — because those are claims about what a clinician did, and only the clinician knows.
Is an AI-generated chart note legally acceptable?
The note is acceptable because a clinician reviewed and signed it, not because of how the draft was produced. Authorship rests with the person who signs. Software should therefore treat AI output as a draft until a clinician acts on it, and lock the note on that action.
What is the risk of AI in clinical documentation?
The main one is a plausible sentence nobody said. A model asked to produce a complete-looking note will fill gaps with what usually appears there — which reads perfectly and can be false. The mitigation is to never let a model assert something it was not told.
Does AI charting save a denturist time?
It saves the transcription and formatting time, which is most of the tedium — dictating an exam and getting a structured note back rather than typing it after the patient leaves. It does not remove the review step, and a tool that promises it does is promising the wrong thing.
Does DentureFlowPro let AI sign or lock a note?
No. AI output is a draft; a clinician reviews and signs, and the note locks on their action rather than the model's. Clinical attestations — that the medical and dental history were reviewed, that consent was obtained and explained — are confirmed by the clinician and never asserted on their behalf.
See how it actually works
Every claim on this page is set out in detail on our compliance page — or watch the software do the work in the demo videos.