Clinic reception is the most interrupted job in healthcare. Someone is at the desk, two lines are ringing, and the person on hold wants to know whether you take their insurance and whether you have anything on Thursday. Meanwhile the practice's website has a contact form that nobody answers until Tuesday.
An AI chatbot is a good fit for exactly the top layer of that workload: the availability, price, location, parking, insurance and "do you do that treatment" questions that make up most of the volume and none of the clinical risk. It is a bad fit for anything that looks like advice, triage, or a patient record. This guide draws that line clearly, then shows you how to build the safe version.
The line: administrative yes, clinical no
| Ask | Bot handles it? | Why |
|---|---|---|
| Opening hours, address, parking, accessibility | Yes, fully | Static facts, zero risk |
| Treatment prices and what's included | Yes, from your published price list | Cuts your highest-volume call type |
| "Do you take my insurance / are you accepting new patients?" | Yes | Pure eligibility routing |
| Booking, rescheduling requests, cancellations | Capture, and book if calendar-connected | Highest value, still administrative |
| "What does this treatment involve?" | Only from your own written patient information | Safe when quoted, unsafe when improvised |
| "Is this pain normal?" / symptoms / medication | No. Escalate immediately. | Clinical judgement. Never automate it. |
| Anything urgent or emergency-flavoured | No. Show emergency instructions and stop. | Delay is the harm |
Everything in the green half of that table is roughly 80% of your inbound messages. Automating it is a clear win. Everything in the red half should be a hard-coded escape hatch, not a judgement call the model makes on the day.
Say this out loud: don't collect clinical detail
The safest patient chatbot is one that never holds anything sensitive. Design it so that the most detailed thing it ever captures is "wants an appointment about a dental matter, prefers afternoons", plus a name and a contact number. The clinical conversation happens on your systems, with your staff.
Concretely:
- Do not ask for symptoms, diagnoses, medication lists, or history. If a patient volunteers them anyway, the bot should acknowledge briefly and move to booking or escalation without repeating the detail back or asking follow-ups.
- Do not ask for insurance ID numbers, national health numbers, or dates of birth. "Which insurer?" is enough to route. The rest belongs in your practice management system.
- Do not use free-text boxes labelled "tell us about your problem". That is an invitation to paste a medical history into a marketing tool. Use buttons for treatment category instead.
- Be plain about the tool you're using. PopABot is general-purpose business software. It is not sold as a HIPAA-compliant platform and does not offer a business associate agreement, so it should sit strictly in front of your clinical systems, never inside them. If your regulatory position requires a signed BAA or an equivalent processor agreement covering health data, use a platform that offers one and keep the chatbot to purely public information.
That constraint is not a limitation in practice. The volume you want to remove from reception is availability, price, location and insurance questions, and none of it requires patient data.
The escape hatch, and why it must be hard-coded
Every clinic bot needs one non-negotiable path: anything that reads as urgent gets the emergency instructions and a human, immediately, without the bot trying to be helpful first.
Build this as an explicit branch, not as a hopeful line in the prompt. In PopABot's flow builder you define the situation in plain language ("the visitor describes a medical emergency, severe pain, bleeding, or anything time-critical") and point it at a step containing fixed text you wrote: your out-of-hours number, the national emergency number for your country, and a clear "do not wait for a reply in this chat" instruction. The classifier evaluates free text at any point in the flow, so it fires wherever the visitor happens to be. Then the conversation ends. No booking flow, no lead form, no "before you go, can I take your email".
The flow builder is available on every plan, including the free tier, so this does not depend on your subscription level.
Test it before you publish. Type the five most alarming things a patient could type and confirm all five hit the escape hatch. This is the one test worth repeating every time you change the bot.
Booking, the part that pays for itself
The highest-return flow in a clinic is a booking request captured out of hours. Two ways to do it:
- Calendar-connected booking. The bot offers your real free slots with your rules (appointment length by treatment type, no new patients on Friday afternoon, buffer between appointments) and writes the booking straight into Google Calendar. Details on the appointment booking page. Available on Essential and above.
- Request capture. If your practice management system is the source of truth and you don't want a second calendar, the bot captures treatment type, preferred days, name and phone, then submits it as a lead. Reception confirms in the morning. You still caught the patient who was going to book with someone else at 21:00.
Either way, turn on SMS notifications so a new-patient enquiry hits a phone rather than an inbox. New-patient enquiries are the ones with real lifetime value; a two-hour response beats a two-day one consistently.
Price questions are conversions, not annoyances
Private-pay clinics (dental, aesthetics, physio, optical, veterinary) lose a lot of enquiries to price anxiety. People will not ring to ask what a whitening course costs, but they will type it. Publish your price list into the bot's data, let it answer directly, and put the booking prompt immediately after the number. "Composite bonding starts at £X per tooth and takes about an hour. Would you like to book a consultation?" converts far better than a contact form that promises a callback.
If your prices genuinely depend on assessment, say the range and what determines it. A range plus an explanation converts. "Contact us for pricing" does not.
Multilingual matters in healthcare more than most sectors
Patients whose first language is not your local one are exactly the group most likely to avoid phoning a clinic. The bot detects the browser language and replies in it, across roughly 100 languages, from the same English source content you wrote once. For a practice in a mixed-language catchment, this quietly removes a real access barrier and it costs nothing extra.
Setup checklist
- Write the emergency escape hatch first. Before any other content. Test it with five phrasings.
- Load your public information: treatments, prices or price ranges, opening hours, insurers accepted, whether you're taking new patients, parking, access, first-visit expectations.
- Load your own patient information leaflets as text so treatment explanations are quoted from material a clinician approved, not generated.
- Set explicit instructions: no advice, no triage, no symptom discussion, no reassurance about pain, always offer a human.
- Build the booking flow with treatment category as buttons, never as free text.
- Turn on SMS notifications for new-patient enquiries.
- Have a clinician read fifty transcripts in the first fortnight. This is the quality gate that matters; the conversations view makes it a twenty-minute job.
- Update your privacy notice to mention the chat tool, what it stores, and for how long.
Related reading
- AI appointment booking use case
- How to train an AI chatbot on your own business data
- Filtering enquiries before they reach a human
- The 7 best AI chatbots for small businesses in 2026, including which ones offer HIPAA coverage