The fastest way to make an AI receptionist impressive in a demo is to let it improvise: reassure the patient, guess at the problem, offer a booking. The fastest way to hurt someone is exactly the same list. This product is built around a safety contract — a short set of things the AI is never allowed to do, enforced in code rather than in a prompt.

The contract

  • Never diagnoses or prescribes.No “sounds like an abscess”, no medication suggestions, no “that's probably nothing serious”.
  • Never books appointments. It can propose a slot; only reception confirms it in the practice management system.
  • Never claims to be a person. The first SMS says a receptionist will review the conversation.
  • Never improvises emergency wording. Red-flag replies are fixed text, written by clinical reviewers.

Red flags are a rule, not a model

When a patient describes facial swelling, difficulty swallowing or breathing, uncontrolled bleeding, or trauma, the response must not depend on what a language model felt like generating that day. So it doesn't: a deterministic keyword rule runs before any AI call, and when it fires, the patient receives one fixed message directing them to urgent help and NHS 111, while reception is alerted in parallel. The AI cannot edit that wording. It also can't suppress it — the rule sits in front of the model, not behind it.

If you ever see an appointment-slot suggestion on a red-flag conversation, that is a bug, and the dashboard says exactly that. Boundaries you can verify beat promises you can't.

Layers, because any one layer fails

  • Legal compliance first. STOP and START keywords are honoured before anything else — before triage, before any reply logic.
  • Deterministic rules handle red flags and opt-outs without a model in the loop.
  • Structured, validated AI output.When the model does classify a message, its output is schema-checked; if it can't produce a valid classification, a deterministic fallback takes over rather than guessing.
  • A human in every loop. Reception confirms every action — bookings, callbacks, escalations.
  • A kill switch. AI replies can be disabled clinic-wide in under a minute, per conversation instantly, and nothing is deleted when you pause.
  • An audit trail. Every state change is recorded, so what happened is never a matter of memory.

Why sell the boundaries?

Because in UK dentistry the boundaries are the product. A tool that texts patients on your behalf is acting in your name, under your GDC obligations and your patients' trust. The honest pitch is not “the AI handles everything” — it's that the AI handles the chasing, the structure, and the speed, while your team keeps every decision. You can see all of it, including the red-flag flow, in the live demo — it runs on fixture data, so nothing you type there touches a real patient record.