Best Practices· 5 min read

Guardrails for Regulated Industries: What Your Assistant Must Not Advise On

How clinics, law firms and financial businesses can use an AI assistant for information, not advice, with boundary wording you can copy.


A physiotherapy clinic's assistant is asked whether a sharp knee pain after running needs rest or ice. A family law firm's assistant is asked whether a partner can take the children abroad without consent. A mortgage broker's assistant is asked whether now is a good time to fix.

Each question is reasonable. For an AI assistant in regulated industries, each is also a request for advice about one person's situation. In health, legal and financial services, that belongs to a qualified human who can ask follow-up questions, check records and carry professional responsibility for the answer.

The assistant can still be very useful. It just needs a clear line, written down, between giving information and giving advice, and boundary wording that leaves visitors feeling helped rather than stonewalled.

This is general guidance, not legal or compliance advice. Your regulator and your professional body set the actual rules, so confirm your obligations with them.

Information versus advice

The simplest test: information is true for everyone who asks; advice depends on who is asking. "We offer 45-minute initial assessments for £65" is information. "You should book an assessment because that sounds like a ligament issue" is advice.

Sector Fine for the assistant Needs a qualified person
Health clinics Services, fees, opening hours, how to register, what to bring, parking Symptoms, diagnosis, treatment choice, medication, test results
Dental Treatments offered, price ranges, NHS or private options, emergency slots policy Whether a tooth needs treatment, pain management, suitability for a procedure
Law firms Practice areas, consultation fees, what happens at a first meeting, documents to bring Whether someone has a case, deadlines on their matter, what they should do next
Accountants Services, fixed fees, deadlines published by the tax authority, what records to gather How to treat a specific income, whether to register, what they owe
Financial advice and brokers Types of products handled, how fees work, the advice process Product recommendations, timing decisions, affordability judgements
Veterinary Opening hours, vaccination packages, out-of-hours arrangements Whether an animal is ill, dosage, whether to wait until morning

The right-hand column is not banned from the conversation. It is where the assistant should stop and pass the enquiry to a person.

The words that should trigger the boundary

Visitors rarely announce that they want advice. The signal is usually in the phrasing, and it helps to spell these out in the assistant's instructions:

When any of these appear, the assistant's job changes from answering to routing: acknowledge, explain briefly why a person needs to handle it, and capture the details.

Boundary wording that does not feel like a brush-off

A bare "I cannot give advice" is technically correct and quite irritating. A good boundary reply says what it cannot do and who can, then makes the next step easy.

Visitor: I've had a sharp pain on the outside of my knee since my run on Sunday. Should I rest it or keep training? Assistant: I'm not able to assess injuries, and it would not be fair to guess about your knee. One of our physiotherapists can look at it properly. Initial assessments are 45 minutes. Would you like me to pass on a request with your availability? If the pain is severe, or you cannot put weight on it, please contact NHS 111 or your GP rather than waiting for us.

Visitor: Do I have grounds to contest my father's will? Assistant: That depends on details a solicitor would need to go through with you, so I can't answer it here. Our contentious probate team offers an initial call. I can take your name, a number and a line about the situation, and someone from the team will be in touch to arrange it.

The prices and services in these examples are illustrations. Your assistant should only quote what is in your content.

Emergencies come first

For health, care and safety businesses, one rule overrides everything: if a message suggests immediate danger, the first thing the assistant says is where to get urgent help. In the UK that means 999 for emergencies and NHS 111 for urgent medical advice. Use the right numbers for where your patients are.

It should not try to triage, reassure or counsel. Write this instruction in explicit terms, and add a Q&A pair so the wording is fixed. Then test it with messages that describe chest pain, a child who has swallowed something and a pet that has collapsed. Our guide to after-hours emergency triage covers the routing side in more depth.

Building the guardrails in

Guardrails live in several places, and it is worth using all of them:

  1. The business description. State plainly: "You provide general information about our services. You never assess symptoms, give legal opinions or recommend financial products. You pass those questions to the team."
  2. Q&A pairs for boundary topics. Write the exact reply for "can you tell me if I have a case" or "what dose should I give".
  3. Content hygiene. Remove blog posts that read like personal advice ("if you have X, do Y") from what the assistant is taught, or they may be quoted as if they were tailored guidance.
  4. Handover. Switch on lead capture and transfer to a human so the boundary leads somewhere useful. See when a chatbot should hand over to a human.
  5. Test with awkward questions. Ten real-sounding advice requests, before launch and after every major content change.

For sector-specific detail, our posts on law firm intake and dental practices go further.

What to do before switching it on

Sit down with whoever holds compliance responsibility in your business, even if that is you on a Tuesday afternoon, and fill in a version of the table above for your own services. Agree the right-hand column in writing.

Then turn that column into instructions and Q&A pairs, write the emergency reply, and test it with a colleague trying their hardest to get advice out of it. Only go live once it fails politely every time.

Frequently asked questions

Can a medical practice use a website chatbot at all?
Many do, for opening hours, services, fees, registration steps and appointment requests. It should not assess symptoms or suggest treatment, and urgent matters must go to a clinician or emergency number.
Does adding a disclaimer make chatbot advice safe?
No. A disclaimer helps set expectations, but the assistant should still avoid giving personal advice. Check your regulator's guidance and speak to your own compliance adviser.
Can a law firm's assistant tell someone if they have a case?
It should not. It can explain what areas the firm handles, what an initial consultation involves and what to bring, then capture the enquiry for a solicitor.
What should the assistant do if someone mentions self-harm or an emergency?
It should give emergency contact details straight away and not attempt to counsel them. Put that instruction in writing and test it before going live.

Keep reading

Guardrails for Regulated Industries: What Your Assistant Must Not Advise On · SpideyChat