Artificial intelligence
Models draft. Clinicians decide.
The dangerous failure in clinical AI is not the obvious error — it is the plausible one, arriving at the end of a long day. Everything below is designed for the tired clinician rather than the careful one.
01Principles
Six rules that constrain what we ship.
These are structural, not aspirational. Each one is enforced in code rather than in a policy document.
Drafts, never decisions
A model can propose a note, a plan or a recall. Only a clinician can accept it. There is no path in our software where an AI output reaches a patient without a human having read it first.
Grounded in the practice’s own data
Clinical assistance retrieves from the record in front of the clinician, not from an open web index. If the answer is not in the chart or in a cited clinical reference, the model says so rather than filling the gap.
Every output carries its receipts
Suggestions ship with the source they were drawn from and a confidence indicator. A clinician can open the underlying note, radiograph or guideline in one tap.
Disagreement is the product
When a clinician edits or rejects a draft, we capture what changed. Those corrections drive our evaluation set — the fastest way to improve a clinical model is to make being wrong visible and cheap to fix.
Scoped to what it should know
Models run under the same authorisation as the person using them. An assistant querying the record cannot surface data the assistant’s own account is not permitted to open.
Measured against clinicians
We evaluate on held-out cases reviewed by practising dentists, and we publish what the model is not good at. Capability claims without an error rate are marketing, not evidence.
02Guardrails
Four things our AI is not permitted to do.
Stated plainly so a clinician can hold us to them, and so a procurement team can put them in a contract.
No autonomous writes
A model output never reaches a chart, a patient or a bill without a clinician explicitly accepting it.
No training on identifiable data
Evaluation sets are de-identified and used only under agreement with the custodian clinic.
No answers without sources
If the answer is not in the record or a cited reference, the model says it does not know.
No privilege escalation
Models run under the permissions of the person using them, enforced at the data layer.
03Applications
Where intelligence actually helps in a clinic day.
Chosen because they remove work a clinician resents, not because they demonstrate a capability.
Ambient charting
Turn a spoken consultation into a structured clinical note against the right teeth and surfaces, ready for the clinician to correct and sign.
Treatment plan drafting
Propose staged plans with costs from the practice’s own fee schedule, based on findings already charted.
Recall intelligence
Predict which patients are due, which are likely to lapse, and which message and time actually gets them back in the chair.
No-show forecasting
Score each booking so the clinic can double-book intelligently instead of losing an hour of chair time.
Patient-language search
Map “tooth broke while eating, hurts with cold” to the right procedure, price band and available clinician.
Document understanding
Read uploaded radiographs, lab reports and old paper case sheets into structured, searchable record entries.
04Evaluation
How we know whether it works.
Every clinical model is measured on held-out cases reviewed by practising dentists. When a clinician edits or rejects a draft in the product, we capture the difference — those corrections are the most valuable signal the system produces, because they are a working clinician telling us precisely where the model is wrong, in context.
We report accuracy alongside failure modes. A capability claim without an error rate is marketing, and we would rather publish the cases we still get wrong than describe the ones we get right.
If accepting a suggestion is one tap and correcting it is a form, you have not built a review step. You have built a nudge toward agreement.
We publish what the models get wrong.
Our methodology, evaluation design and known failure modes are written up on the blog. If you are a researcher or clinician who wants to review them, we will send you the full set.