Can an AI agent give medical advice?

No. Diagnosis, treatment recommendations and reassurance that a symptom is safe to ignore belong to clinicians. An AI agent can share approved health information, help people book the right appointment, collect symptoms for a clinician and send anyone describing an emergency straight to urgent care. Personal clinical advice stays with a clinician, who is accountable for it.

What counts as medical advice?

Medical advice is a recommendation about a specific person's health: what they probably have, what they should take, whether they need to be seen. "Paracetamol and ibuprofen can be taken together by most adults" is information. "You can take both for your headache" is advice, because it assumes facts about the person that the agent does not know.

The boundary is not always visible to the person asking. Someone typing "is this rash serious" wants a judgement about their own body. Whatever the agent says, they will treat it as an answer to that question.

There is a regulatory dimension too. In many markets, software that diagnoses or recommends treatment can fall within medical device rules, which bring their own approval process. An agent that drifts from information into diagnosis can change its regulatory status without anyone deciding to.

Why is bad medical advice irreversible?

The harm happens in the time between the advice and the correction. A person told their chest pain is probably muscular stays at home. A person told a drug combination is fine takes it. By the time anyone reviews the conversation, the event has passed.

That is the point made in why autonomy in regulated industries is a question about reversal: judge an action by whether its consequences can be undone. A conversation log can be corrected; a delayed hospital visit cannot.

What can the agent do, and what stays with a clinician?

The agent can:

• answer questions from approved patient information, quoting the source;
• explain how to prepare for a procedure, using the clinic's own instructions;
• book, move and cancel appointments, and choose the appointment type from a set protocol;
• collect symptoms using an approved question set and pass them to a clinician;
• recognise red-flag symptoms and tell the person to call emergency services or attend urgent care at once;
• send repeat prescription requests to a clinician for approval.

A clinician keeps:

• any diagnosis, or any view on what is probably wrong;
• any recommendation to start, stop or change a medicine;
• any judgement that someone does not need to be seen;
• interpreting test results for the patient;
• every conversation where the person is distressed, confused or at risk.

The same split applies to legal advice and financial advice: information can be automated, personal recommendations cannot.

A worked example

Take a group of outpatient clinics running an agent on its patient messaging line. A patient writes that they started a new blood pressure tablet last week and now feel dizzy when they stand up. They ask whether they should stop taking it.

The agent does not answer the question. It asks the approved follow-up questions: any fainting, chest pain or breathlessness. The answers are no. It tells the patient that a clinician will review the message the same day, gives the clinic's urgent number in case symptoms change, and sends the clinician the medicine, the start date and the answers.

Had the patient reported fainting, the agent would have told them to seek urgent care immediately and flagged the case for follow-up.

What controls make it safe?

• Approved sources only. The agent answers from patient information the clinical team has signed off, not from whatever sits on a network folder. If you are wiring it up, read how to give an AI agent access to shared drives before connecting anything.
• A fixed escalation list. Red-flag symptoms are defined by clinicians and tested before go-live. The agent may escalate on its own; it may never de-escalate.
• Certified refusals. Test the agent on real messages from the clinic's history and check that it refuses advice questions as reliably as it answers information ones.
• Data limits. Health data is special-category data under GDPR and protected health information under HIPAA in the United States. The agent sees what the task needs, and nothing more.
• Start with administration. Appointment booking is a better first use than symptom handling; see how to choose the first process to give an AI agent.