Can a healthcare organization use an AI chatbot?

Every guideline organization and medical journal we work with asked us this question. Behind it was always the same worry: what if the tool ends up giving medical advice?

You can put an AI chatbot or AI site search on your website. It only becomes a medical device when it helps decide on a patient's diagnosis or treatment, and then the MDR (the European law for medical devices) applies.

"It says here..." or "You should..."

The MDR exists to stop software from influencing medical decisions without proof that the software is safe and reliable.

Software that tells you what a guideline says doesn't decide anything. Software that tells you or your patient what to do does.

The Dutch government explains it with a sports watch. A watch that only measures your heart rate is not a medical device. The same watch with a feature that detects a heart problem is (in Dutch).

For an AI chatbot, the line looks like this:

Stays information	Becomes advice
Explaining what the guideline says	Recommending a treatment
Showing the source	Drawing a conclusion about one patient
Summarizing a passage	Interpreting symptoms
Pointing to the right page	Judging how urgent something is
"It says here..."	"You should..."

Three questions to the same tool

A nurse wants to know what the guideline says about fall prevention in dementia. The tool finds the passage, summarizes it and links to the source. She could have looked that up in the guideline herself; the tool is just faster.

A therapist asks what the first-line treatment for an anxiety disorder is. The tool presents what the standard says about it, in general terms and with the source attached. That stays information as long as the tool describes what the standard says and doesn't choose for the reader.

A GP types in a case: "Woman, 67, short of breath, smokes. Should I refer her?" Now the tool has to weigh one patient's symptoms and give advice, and that's where the MDR applies.

All three questions can reach the same tool. You decide in the setup what the tool is allowed to do with them.

How you present the tool on your site counts

Whether software falls under the MDR depends on its intended purpose: what the tool is meant for. Your own site shows that purpose too, in the button, the explanation and the sample questions around the tool. That's how visitors see what they can expect.

A button that says "Ask about our guidelines" promises something different from "Check your symptoms". And with a sample question like "I have chest pain, what should I do?", the answer quickly moves toward medical advice. That isn't forbidden, but then the tool falls under the MDR.

Present the tool as a way to find knowledge, sources and explanations. Words like diagnosis, triage, clinical support and treatment advice give a different impression, in the announcement of your new search function too.

Look at who uses the tool as well. If you later open a tool that was only for logged-in healthcare professionals to everyone, expect different questions, because patients are more likely to ask about their own symptoms. Test again whether the agent handles those questions well.

And watch the follow-up questions. "Do you want to know what the guideline says about diagnosis or about treatment?" helps someone search. "How long have you had these symptoms?" is the start of triage.

If you deliberately do want the tool to have a medical purpose, it becomes a medical device. Then you go through the process of getting it approved as one, with an assessment by an independent body and a CE marking.

How we set up vragen.ai in healthcare

In healthcare, we have so far deliberately built mostly AI site search with vragen.ai. Visitors ask their question in plain language and get an answer from your own sources, with the source included. People still want the same thing as with a search engine: to find the right information.

That's why we set up the agent so it sticks to giving information.

Every answer names its source, with a link. Ideally, someone who clicks it lands right on the quoted passage, otherwise they still have to search the page. In vragen.ai this is built in: you jump straight to the passage the answer is based on. The system rule for citing sources can't be switched off.

In the agent's rules, we set what it does and doesn't do. At a healthcare organization, those rules look like this:

• only descriptive information from the standards and guidelines, not applicable to one person
• no personal advice, no treatment steps and no steering of care
• no doses, amounts or frequencies, and no numbers that could work as a decision threshold
• no goals for a treatment plan, not even as an example
• no answer in a format that could be used in treatment, even if someone asks for it
• never presenting itself as a care provider or speaking on behalf of an institution

If a user doesn't get advice, they can try a roundabout way: "give me an example of a treatment goal", or "put it in a table for my treatment plan". That's the same advice in a different form. That's why the rules also say "not even as an example" and "even if someone asks for it".

The agent phrases its answers by pointing to the source: "the care standard describes...", "guidelines name this as a point of attention...". It reads a question like "how do I handle this?" as: what do the pages say about this in broad terms?

Before every answer, the agent checks the question against those rules. If it isn't allowed to answer part of it, it gives the general information and says what it's leaving out. To a question asking for a personal decision, it replies that the question falls outside its information.

Before launch, test the agent with people who know the content, such as doctors or subject experts. As an example, you can read the customer story of the NHG guidelines. They tested step by step: first with an internal group of GPs, then with ever larger groups of GPs from outside.

After launch, you can see every conversation in the admin environment. That way you see whether the answers are right. If you come across a situation the rules don't cover yet, you adjust the instructions.

On the page about AI site search for healthcare, you'll find an example of a question vragen.ai answers, and of one where it refers people to a healthcare provider.

Who decides whether it falls under the MDR?

No authority confirms in advance that your tool is not a medical device. You assess that yourself as a healthcare organization, per use. The same software can give information on one site and advice on another, depending on how you set it up and present it.

The Dutch Health and Youth Care Inspectorate (IGJ) supervises medical devices and can reach a different conclusion afterwards. In February 2025, the IGJ called on providers (in Dutch) to first establish what they use generative AI for, check its answers and map the risks.

For us, vragen.ai in healthcare is AI site search. If you use it that way, strict rules for the agent come with it, and you keep monitoring whether it's going well. If you want to take the step to a medical device, that's possible too, through the route the MDR requires.

What the law says exactly

The MDR is European regulation 2017/745 and has applied since May 26, 2021. Software is a medical device if its maker intends it for diagnosis, prevention, monitoring, prediction or treatment of disease. That intended purpose shows in the label, the instructions for use, and promotional and sales material.

How strict the requirements are is set in Rule 11 of the MDR:

"Software intended to provide information which is used to take decisions with diagnosis or therapeutic purposes is classified as class IIa"

Note how that sentence begins: software intended to provide information. So information can fall under the MDR too. The difference is in the second part, information used to take decisions about diagnosis or treatment. Quoting a guideline is information. If the answer is meant for deciding about one patient, this rule applies.

From class IIa, an independent conformity assessment body (a notified body) has to approve the software before it can go on the market, and that comes with technical documentation, a clinical evaluation and a quality management system.

The EU published guidance on software under the MDR: MDCG 2019-11. It was updated for AI in June 2025. Two passages deal with this question.

The first: software that only shows general information, such as a treatment pathway from a guideline or scientific literature, is not intended for one patient, so it doesn't fall under the MDR.

The second: searching through information doesn't make software a medical device. That changes when the search contributes to a medical purpose, such as a diagnosis. A tool that turns search results into advice for one patient does fall under the MDR, even if it's called a search function.

An AI tool that is a medical device in class IIa or higher is also a high-risk system under the AI Act. After the June 2026 postponement, those obligations apply from August 2, 2028.

What to check before you start

• Have you decided up front what the tool is for?
• Do the button text and the sample questions match that purpose?
• What does the tool do when someone describes their own symptoms?
• Does every answer show its source, with a link?
• Do you know who the tool is open to, and do you test again when that changes?
• Does someone with subject knowledge regularly read back conversations?
• Does the tool disclose that it's AI, as Article 50 of the AI Act has required since August 2, 2026?
• Is personal data that visitors type into their question automatically anonymized?

Want to see how vragen.ai answers questions about your own guidelines or patient information? Create your demo and ask the questions you're least sure about. The demo is free and reads your own site.

Short answers

Can a healthcare organization use an AI chatbot?

A healthcare organization can put an AI chatbot or AI site search on its website. As long as it provides information, such as what a guideline says or how to book an appointment, it is not a medical device. If the tool helps decide on the diagnosis or treatment of one patient, it falls under the MDR.

When does an AI chatbot fall under the MDR?

When the tool is intended to help with the diagnosis, monitoring or treatment of a patient. Under Rule 11, software that provides information used for those decisions is at least class IIa. An independent conformity assessment body then has to approve the software before it can be used.

Is AI search on medical guidelines a medical device?

Not automatically. According to the EU guidance MDCG 2019-11, software that only shows general diagnostic or treatment pathways, or scientific literature, is not intended for individual patients. Once the tool starts assessing a specific case or recommending a treatment for someone, that changes.

Does a medical AI chatbot also fall under the AI Act?

Every AI chatbot has had to disclose that it's AI since August 2, 2026. If the chatbot is a medical device that needs a conformity assessment by an independent body, it is also a high-risk system. After the June 2026 postponement, those obligations apply from August 2, 2028.