An AI chatbot for medical clinics makes sense where reception hears the same questions all day: when does the clinic open, do I need a referral, what should I bring to my first appointment, where can I park. The answer is the same for every patient. Questions about symptoms, medication and test results are different: they need a doctor, and a wrong answer can cause harm. A good assistant knows this line and stays behind it.
In this guide you settle four things: a list of administrative topics, a list of off-limits topics, a fixed sentence about emergencies and the way a chat is handed to reception. The examples use a fictional clinic called "Linden House" with three departments. This text is neither legal nor medical advice.
Which patient questions the assistant can answer
A question suits the assistant when your clinic has written down a single answer to it and it requires neither a judgement about health nor access to one person's records. Opening hours meet that test. "With this kind of pain, should I come in today?" does not.
Scroll the table sideways to see all columns.
| Patient question | Does the assistant answer | Source or next step |
|---|---|---|
| When is the department open? | Yes | Current department schedule on the website |
| Do I need a referral for this department? | Yes, in general | Admission rules written by the clinic |
| What should I bring to my first visit? | Yes | Document list from the patient information |
| Where can I park and how do I get there? | Yes | Directions and parking description |
| How do I book or cancel an appointment? | Yes | Description of booking channels; reception makes the change |
| Are these symptoms serious? | No | Fixed emergency sentence and contact with a doctor |
| What is my test result? | No | Handoff to reception with the patient's consent |
If you cannot point to the paragraph the answer should come from, the question is not ready yet. Moving appointment booking online is a separate topic, covered in the guide on online booking for clinics.
Knowledge only from approved clinic content
The assistant should not draw on the model's general medical knowledge or on random pages. Prepare one "Patient information" document, owned by a named person, and add it to the knowledge base as the main source.
Departments and hours: names, days and opening times. Referrals: which departments require one, without judging whether the patient will get it. Documents: what to bring to a first visit. Booking: reception phone, form, cancellation rules. Directions and parking. What the assistant does not do: "I do not answer questions about symptoms, medication or results." Date of last change at the top.
In Meet AI the assistant answers only from the site's active knowledge base. If it finds no reliable passage, it does not guess; it says it has no information. Knowledge does not refresh on its own, so after editing the document, add it again or refresh the sources manually (a full refresh can be run once every 7 days). Which materials to include and which to leave out is covered in the guide on what data to add to a chatbot's knowledge.
Topics the assistant does not answer
Write the off-limits list before the assistant meets its first patient. Keep these topics out of the knowledge base entirely, because any passage can end up in an answer. The patient document only says that a doctor decides on these matters.
- Assessing symptoms, urgency or whether someone needs to see a doctor.
- Dosage, substitutes and interactions of medicines.
- Interpreting test results and reports from other providers.
- Decisions about referrals, prescriptions or sick notes.
- Doctors' qualifications beyond the clinic's official description.
- Data about specific patients: visits, results, treatment history.
The last two points are also about liability. In the Hamm case described in the guide on who is responsible when a chatbot is wrong, a cosmetic treatment provider's chatbot gave false information about doctors' specialisations. Describe your doctors only from the approved team page.
Emergencies: a fixed sentence instead of a model answer
A patient may write to the chat in a situation that needs immediate help. Do not assume the assistant will always recognise such a message. Emergency numbers should be visible at all times: in the chat greeting, on the contact page and in the patient document.
"If someone's life is in danger, call your local emergency number (112 in the EU). This chat is not for reporting emergencies. If the clinic is closed and your problem cannot wait, use the local out-of-hours medical service." In Poland add 999 and the NFZ night and holiday care; in Germany add 116117 for the on-call medical service.
Poland's NFZ says to call 112 or 999 when life is at risk, and that night and holiday care runs from 6 pm to 8 am on weekdays and around the clock at weekends and on public holidays. The notice says nothing about chats on clinic websites.
116117.de describes 116117 as round-the-clock access to Germany's on-call medical service when a problem cannot wait until the practice opens, and refers life-threatening cases to 112. The page does not cover other countries.
A similar split between urgent and routine matters, only for animals, is covered in the guide on urgent cases at a veterinary clinic.
The greeting: AI, scope and a request not to share health data
Disclosure itself, the assistant's name and the route to a person are covered in the guide on AI chatbot disclosure. A clinic adds a request that patients do not describe symptoms or results in the AI chat.
"Hello, I am the clinic's AI assistant. I help with practical matters: opening hours, referrals, documents and directions. I do not answer medical questions. Please do not share information about your health here. If someone's life is in danger, call your local emergency number (112 in the EU)."
The request has a concrete reason. In Meet AI the conversation text is sent to the model provider, DeepSeek, with every generation, and an administrator accepts a notice about this during installation. Whether that data flow suits a medical practice is for the practice and its data protection officer to decide; the questions to check are collected in the guide on AI in customer service and the GDPR.
Article 9(1) of the GDPR lists data concerning health among the special categories whose processing is prohibited unless an exception in paragraph 2 applies. The provision does not settle whether a particular chat tool may be used at a clinic.
Copilot as a cautious start, then handoff to reception
Meet AI starts in Copilot mode: the patient writes to reception, and the assistant prepares a private draft reply for the staff member only. Nothing is sent automatically. You see the quality of your knowledge without patients receiving a wrong answer directly. The conversation text still goes to the model provider when a draft is generated.
- A few weeks of Copilot.Reception uses the drafts and notes questions where knowledge was missing or a draft drifted into medical territory.
- Fix the document.Add missing administrative topics to the patient information and remove sentences that read like advice.
- Private tests.Ask the questions from the table and the off-limits list. Results show citations of the sources used and have no public effect.
- Decide on a patient-facing mode.Only then consider Autonomous mode or Hybrid mode, where the "Ask the AI assistant" card appears only when nobody on the team is available. In both, the patient chooses that card.
When the assistant has no information, it offers to pass the chat to the team, and the handoff happens only after the patient agrees. Assistant messages carry an AI badge in the widget. When neither the team nor a contact option is available, the assistant does not promise a reply. What the person taking over should see is covered in the guide on handing an AI chat to a person, and a broader set of checks in the guide on testing a chatbot before launch.
Questions from clinics
What if a patient describes symptoms anyway? The administrative document has no answer to such a question, so the assistant should say it has no information and offer contact with the team. Check this beforehand in private tests. Reception then replies with its fixed sentence about seeing a doctor or calling emergency numbers. Do not copy such messages into the assistant's knowledge.
Can the assistant book appointments? A knowledge-based assistant explains how to book, but it does not register patients and cannot see doctors' schedules. If the clinic wants bookings through its website, it needs a separate booking tool.
