AI chat for clinics: what it should answer, and what it must never touch
Patients ask the same twenty questions before they ever reach a doctor. Here is how a clinic can answer them with AI chat without drifting into medical advice or leaking patient data.
MagizAI Team5 min read
Reception in a busy clinic is interrupted work. A patient at the desk, two on hold, and a WhatsApp message asking whether the doctor is in on Saturday. Most of those interruptions are not medical. They are opening hours, parking, insurance, what to bring, how to reschedule.
An AI assistant can hold that line. The important part is deciding, on purpose, where it stops.
Start with the questions that are not medical
Before anything clever, write down what your front desk repeats every day. For most practices it looks like this:
- Opening hours, including the exception for public holidays
- Where to park, which entrance, which floor
- Which insurers you work with, and what a self paying visit costs
- How to book, change or cancel an appointment
- What to bring to a first visit
- How long results usually take, and how they are sent
- Which doctor covers which area
None of that is a diagnosis. All of it is on your website somewhere, or in a receptionist's head, and it is what most people are actually messaging about.
The line that matters: information, not advice
A patient will eventually type their symptoms. That is the moment your setup is tested.
The right behaviour is narrow and boring. Acknowledge, do not interpret. Offer the next step, which is usually booking or speaking to a person. Escalate anything that sounds urgent to a human straight away, with the chat and the patient's message attached.
In MagizAI this is set per bot: instructions in plain words, a topic boundary that keeps it to clinic information, and a handoff that takes over when a person is needed. You can write "never suggest a diagnosis, a medicine or a dose, and never estimate how serious something is" and the assistant follows it in every language it replies in.
Test it before launch. Type three symptom messages and one panicked one. If the assistant offers advice, tighten the instructions and test again.
Patient data: collect less, keep less
A clinic chat should not become a second, unprotected medical record.
Three habits keep it clean.
Ask only what the next step needs. A booking request needs a name, a contact, and a preferred time. It does not need a history.
Do not invite details you cannot protect. If the pre chat form asks "describe your problem", people will write things you now have to store. Ask for the department instead.
Know where messages live and who can read them. In MagizAI, conversations sit in your workspace, roles decide who can open them, and a workspace can export or erase a visitor's data when someone asks. If your practice has a retention rule, apply it here as well as in the record system.
Local law decides the rest, and it varies more than vendors admit. Talk to whoever handles your data protection before you turn chat on, not after.
Bookings without pretending to be the calendar
The tempting build is an assistant that books directly into the practice diary. It is also the one that goes wrong loudly, because a double booked slot is a person standing in your waiting room.
Something calmer works better. The assistant collects the request, confirms the details back, and hands it to the desk to place. Patients get an immediate answer, staff keep control of the diary, and nobody is turned away because two systems disagreed.
If your booking system is reachable over an API, you can go further later. Start with the safe version and see how many requests arrive after hours.
Languages, without hiring for them
Clinics serve whoever walks in. A patient who writes in Arabic should not get English back.
MagizAI detects the language of each message and replies in it, including right to left layout where that applies. For a practice in a mixed neighbourhood this is often the single most noticed improvement, more than speed.
After hours is where it earns its keep
Nobody calls a clinic at eleven at night to chat. They message because they are worried, or because that is when they finally have a minute.
An assistant that answers the hours question, offers the booking form and says clearly that a person will reply in the morning turns a missed call into a booked slot. For anything urgent it should point to your emergency instructions immediately, in the first reply, without a menu.
Write that emergency text yourself. Do not let a model improvise it.
What good looks like after a month
Watch four things rather than a dashboard.
Messages answered without a person. These should be the hours, parking, insurance and admin questions.
Handoffs. A rising number is not a failure. It means people are finding the chat and the assistant knows its limits.
Bookings started after hours. This is the number that pays for the tool.
Anything the assistant got wrong. Read them. Each one is either a missing page on your site or a rule you have not written yet.
A sensible order to set it up
- Put your real clinic information into the knowledge base: hours, addresses, insurers, prices, preparation notes.
- Write the boundary rules: no advice, no diagnosis, no dosage, escalate anything urgent.
- Turn on handoff to your team during working hours.
- Add the pre chat form with the fewest fields you can live with.
- Run it on one page for a week, read every conversation, then put it everywhere.
None of this needs a project team. It needs an afternoon and someone who knows what reception actually gets asked.