3 September 2026 · 4 min read · By Junaid, founder
The line between an AI that answers your phone and an AI that practises medicine
Every AI receptionist being sold into aesthetics clinics right now — and there are several, from roughly £297 to £599 a month as standalone add-ons — makes some version of the same promise. It answers calls, books appointments, never misses an enquiry. Almost none of them say, in public, what it's built to refuse. In a clinic, that's the more important half.
Where regulators actually draw the line
The CQC's guidance on AI in health and care is about human oversight, transparency, accountability and monitoring of what the AI produces. It isn't a ban. It's a requirement that a human can be shown to be answerable. The MHRA separately distinguishes administrative AI from AI that performs a medical function: software that schedules, drafts or summarises sits in a different regulatory category from software that diagnoses or recommends a treatment. Products get into real difficulty when they blur the two.
What that means for a phone call in an injectables clinic
Sooner or later, the assistant will take a call from someone describing a complication, someone in a real emergency, or someone asking a clinical question it has no business answering. What it does in that exact moment is the entire safety case. It has to be tested, not claimed.
- It says what it is. Callers hear that they're speaking with an AI assistant and that the call may be transcribed, before anything else.
- Emergencies are caught before the model is. A 999-emergency detection pass runs on every call ahead of any AI response.
- Complications go to a person. If a message describes one, the caller gets your fixed acknowledgement and a named human is alerted. No model is involved.
- It never gives clinical advice, never diagnoses, never recommends a treatment, and never says a toxin brand name aloud.
- It never takes card details and never alters a signed consent record.
- Every reply is a draft until you approve it, unless you choose otherwise for plain booking confirmations.
- It's proven, not promised. A conversation evaluation harness runs on every release to show that the emergency, complication and human-request paths never invoke the model.
Ask every AI vendor this: "What happens when a caller says she can't see properly after her treatment?" If the answer involves the AI answering, walk away.
Appointbox's scribe and voice receptionist are built and awaiting configuration — no channel is described as live until it is configured, deployed and measured. What each system publishes about its AI.