An AI assistant for your clinic:what's allowed in 2026
Updated · 6 min read
An AI assistant on your website can answer patients at midnight, book appointments and talk to a patient from Berlin in German. Since August 2026, though, the AI Act has new rules for chatbots, and from October 28, 2026, Poland's regulator can fine for breaking them. Here's what a clinic's assistant may do, where medical devices begin and how to build one safely. This isn't legal advice.
What an assistant can do with no extra requirements
The MDCG 2019-11 guidance (revised June 2025) says outright that software for admissions, appointment scheduling and billing isn't a medical device in itself. That's where an assistant delivers the most value with the least risk:
- Booking, rescheduling and canceling appointments in the clinic's own system.
- Prices, opening hours, directions, parking, which doctors are available and how to pay.
- How to prepare for a test or procedure, exactly as the clinic describes it: fasting, which documents to bring, how long the visit takes.
- Questions about referrals, insurance and invoices.
- Conversations in the patient's language, which helps with patients from abroad.
- Handing the conversation to the front desk when a question goes beyond these topics.
Where a medical device begins
The intended purpose draws the line, not the technology. If the assistant assesses symptoms, suggests a possible condition or tells a patient whether to see a doctor, it becomes software with a medical purpose. Under Rule 11 of the EU Medical Device Regulation (MDR), that makes it at least a class IIa device: it needs a notified body, CE marking and clinical evidence.
From August 2, 2028, such a device will also be a high-risk AI system under the AI Act. Emergency patient triage systems are listed in Annex III of the AI Act, with obligations from December 2, 2027.
The guidance also notes that natural-language search that contributes to a medical purpose is no longer "simple search." Adding one feature can change how the whole tool is classified. That's why a clinic's assistant should have its scope written down and enforced in code, not just in a description.
The AI Act: patients must know they're talking to AI
Article 50(1) of the AI Act has applied since August 2, 2026. AI systems that talk to people must tell them they're dealing with AI, unless that's obvious. The European Commission reads that exception narrowly. The EU's simplification package (the Digital Omnibus) didn't move this date.
- The notice must be clear and distinguishable, given at the first interaction at the latest, and accessible to people with disabilities.
- The duty sits with the provider of the system. A clinic that puts an assistant into service under its own brand may count as the provider, even if an agency built it.
- Fines go up to €15 million or 3% of worldwide annual turnover. For small and medium-sized companies, the lower of the two applies.
- In Poland, the Commission for the Development and Safety of Artificial Intelligence (KRiBSI), set up by the Act on Artificial Intelligence Systems (Journal of Laws 2026 item 1003), is the regulator. Its penalty provisions apply from October 28, 2026.
Patient data and model providers
Patients tell an assistant about their symptoms even when nobody asks. Those messages are health data, a special category under Article 9 of the GDPR. Plan for it from day one:
- A data processing agreement (Article 28 of the GDPR) with every provider the conversations pass through: the developer, the hosting and the model provider.
- Processing in Europe. OpenAI lets you store and process data in Europe (the eu.api.openai.com endpoint) once the account is approved for zero data retention. Anthropic's own API currently offers only US and global regions, but Claude runs in EU regions through AWS Bedrock and Google Vertex AI. Azure OpenAI has "Data Zone EU" deployments.
- Minimal data: a name, a phone number and a time slot are enough to book a visit. The assistant doesn't need to ask about symptoms or national ID numbers.
- Short retention of conversations, and a privacy policy that says clearly who processes them.
- No advertising pixels in the chat window.
- In 2025, Poland's data protection authority (UODO) criticized a government voicebot for patient registration for not specifying which data it would process, and said the system also had to be assessed under the AI Act.
The clinic is liable for what the assistant says
In 2024, a Canadian tribunal ordered Air Canada to pay damages because the chatbot on its website gave a customer the wrong refund rules. The tribunal called the argument that the chatbot was "a separate legal entity" remarkable. The same logic applies in the EU: what the assistant says, the clinic says.
In Poland, two more rules apply to clinics. The assistant's answers are information about services, so the advertising ban in Article 14 of the Act on Medical Activity covers them: no encouragements, promotions or promised results. And Article 12 of the Code of Medical Ethics requires a doctor who uses AI in treatment to tell the patient, use certified tools and make the final decision personally. EU countries also have until December 9, 2026, to transpose the new Product Liability Directive, which covers software.
How we build an assistant that stays accurate
- It answers only from the clinic's own content: the price list, procedure pages and FAQ. If something isn't there, it says it doesn't know and passes the question to the front desk.
- It has a written list of topics it won't take on: diagnosis, medication doses, reading test results.
- When a patient describes warning signs, it gives the emergency number (112 in the EU) and doesn't discuss the symptoms further.
- It books through the clinic's real calendar, not through a form that someone retypes by hand.
- Before launch it goes through a set of tests with difficult questions, and after launch the clinic reviews a random sample of conversations every week.
- It introduces itself as an AI assistant at the start of the conversation and offers a way to reach a person right away.
Do patients want this?
According to Eurostat, 32.7% of people aged 16-74 in the EU used generative AI in 2025 (22.7% in Poland). Patients are used to chat, but they expect a real answer or a quick way to reach the front desk. An assistant that answers everything with "please call us" does more harm than having none. We build the assistant as part of a clinic website, connected to its booking system.
- Does a chatbot on a clinic's website need CE marking?
- Not if it handles booking, prices and practical information. CE marking as a medical device is needed only once it assesses symptoms, suggests a diagnosis or decides whether a patient should see a doctor.
- How should we tell patients they're talking to AI?
- The simplest way is in the first message and in the chat window's header, for example: "The clinic's AI assistant. You can ask for the front desk at any time." The notice should be visible before the first answer and readable by screen readers.
- Can a clinic's assistant use ChatGPT or Claude?
- Yes, through the API, with a data processing agreement and processing in Europe. Don't paste patient conversations into the regular ChatGPT or Claude apps: those services aren't meant for processing medical data on a clinic's behalf.
- Can the assistant answer questions about medication and procedures?
- It can pass on what the clinic itself publishes: what a procedure involves, how to prepare and what it costs. It shouldn't advise an individual patient on treatment. That's a doctor's job, and assessing symptoms would make the assistant a medical device.
- MCP integration: your ERP, store and CRM in ChatGPT and Claude
- Aesthetic clinic websites in Poland
- Online booking for a private clinic
- ChatGPT product feed: get your products into ChatGPT shopping
- Clinic advertising in Poland: what a website may say
- Your clinic in ChatGPT and Google
- Patients from abroad: a medical tourism website
- Claude connectors: what they are and how to add one
- MCP servers in ChatGPT: how to connect one
- How to connect WooCommerce to ChatGPT in 2026
- OAuth for MCP servers: sign-in in ChatGPT and Claude
- MCP Apps: interface in ChatGPT and Claude
- MCP server security: what to check before launch
- How to test an MCP server in ChatGPT and Claude
- MCP server or API? When you need both
- Hiring a software studio in Poland
- Your Shopify or WooCommerce store in ChatGPT
- What is an MCP server? Explained for online stores
- Claude connector for SaaS: what it takes
- How to sell in ChatGPT in 2026
- MCP server, ChatGPT plugin or Claude connector?
- How to get listed in ChatGPT and Claude
Want an AI assistant on your clinic's website?
Clinic websites