An AI receptionist for a clinic, specifically
Front desks lose calls during consultations and after close. What an agent can safely handle in a clinic, and the clinical line it must never cross.
A clinic front desk is interrupted by design. Someone is at the counter, the phone rings, and the phone loses. Then there is lunch, and the hours after close, when a patient with a genuine problem rings and hears nothing at all.
This is a good fit for an AI agent, and also a setting with a bright line it must never cross.
Where the calls actually go missing
Pull a month of call logs from your provider and split unanswered calls by hour. Nearly every clinic finds the same three clusters:
- During consultation blocks, when the desk is with a patient.
- Lunch.
- The two hours after closing, which are often the highest-intent calls of the day.
You cannot staff those without hiring. Measure it first — the number is usually larger than the front desk believes, because they only remember the calls they answered.
What it should handle
- Are you open, and when
- The most common out-of-hours question by a distance.
- Where are you, and parking
- Facts that never change and cannot be got wrong.
- Do you do this treatment
- From a list you wrote. Not inferred.
- What does it cost
- Only if you publish the price. If your pricing depends on an examination, the agent should say exactly that and offer a consultation.
- Book, reschedule, cancel
- The one that pays for the deployment — provided it reads real availability. What has to be true first.
- Take a proper message
- Name, number, what it is about, urgency — on the patient record, not a voicemail nobody transcribes.
The line it must not cross
No clinical advice. Not triage, not "does this sound serious", not whether to take a painkiller, not whether to come in tonight. Write these as explicit refusals, because the failure mode of a helpful model is being helpful about exactly this.
Two specific behaviours to configure deliberately:
- Emergencies: detect, redirect to the right number immediately, in the first sentence, and stop. Do not attempt to assess.
- Anything about an existing treatment or medication: take details, promise a clinician, do not engage.
How to make refusal the easy path rather than something the model works around.
Patient data is the sensitive part
A caller will volunteer symptoms whether you ask or not. That means health information is now in a transcript, and possibly a recording.
- Decide a retention period for recordings and transcripts, and write it down. "Forever" is a decision you have made by not making one.
- Say the call is recorded, at the start, if it is.
- Keep the knowledge base to facts about the clinic — never anything about individual patients.
- Ask your vendor the nine questions in what your AI agent does with customer data, and get the answers in the contract.
Say it is an agent
Disclose in the opening line. Patients are more forgiving of an AI that says so and hands over than of one that pretended. It also stops people describing symptoms to something that cannot help.
A sensible first month
- Point it only at calls that currently ring out — after hours and overflow. The baseline is silence, so the bar is low.
- Give it four jobs: hours, location, take a message, book from real availability.
- Read every transcript for two weeks. You will learn which questions your front desk answers fifty times a day.
- Only then widen the scope, and never into clinical territory.
The realistic outcome is not a replaced receptionist. It is the same receptionist no longer losing the calls they were never going to reach. Bring your call log to a demo.