A website chatbot should know what it knows and, just as importantly, what it should not touch. Visitors will ask about their legal rights, their symptoms, their investments and their emergencies, because a text box invites any question. A chatbot that confidently answers all of them is a liability.
This guide lists the topics a business chatbot should hand to a person, how to word the refusal so it helps rather than frustrates, and how to test the boundary before launch. It is general guidance, not legal advice. If your sector is regulated, ask a qualified professional what applies to you.
Why Should a Chatbot Refuse Some Questions?
Because a wrong or overconfident answer in these areas can harm people and expose your business. Advice about health, law or money depends on individual circumstances that a chatbot cannot assess. And even a correct general statement can be read as personal advice.
A bot built to answer only from your approved knowledge has a head start: it declines what you never taught it. But you should still decide, deliberately, which topics are off limits and write the reply for them yourself.
Which Topics Should Go to a Person?
| Topic | Why it is risky | What the bot should do |
|---|---|---|
| Medical symptoms or treatment | Individual circumstances; harm if wrong | Decline, suggest a qualified professional, give your contact route |
| Legal questions about someone's situation | Depends on jurisdiction and facts | Decline, offer a consultation or contact route |
| Financial or investment advice | Regulated in many places; personal risk | Decline, share only published product facts |
| Emergencies and safety | Time-critical | Show emergency guidance and your urgent contact |
| Personal data and account details | Privacy and identity risk | Never reveal; direct to a secure channel |
| Complaints and disputes | Need empathy and authority | Hand off to a person quickly |
| Price promises not published | Creates commitments you did not make | Say no confirmed price; collect contact details |
What Should the Refusal Say?
A good refusal is short, honest and gives a next step. It says the chatbot cannot help with that topic, explains briefly why, and offers a route: a phone number, a contact form or a booking link. Avoid a bare "I can't answer that", which leaves a person stuck.
- Medical: "I can't give medical advice. For your situation, please speak with a qualified health professional. I can help with our opening hours or how to book."
- Legal: "I can't give legal advice. A solicitor or attorney can advise on your situation. I can explain our published terms."
- Financial: "I can't advise on investments or personal finances. I can share our published fees and how to contact our team."
- Emergency: "If this is an emergency, contact your local emergency number now."
Adjust the wording to your sector, and have someone qualified in that field review sensitive replies before they go live.
How Does a Grounded Chatbot Help?
A chatbot that answers only from your knowledge cannot invent advice on topics you never added. ConvoRail is built that way: when nothing matches, it says it has no confirmed information, offers a contact form and flags the conversation in the shared inbox. For off-limits topics you can also add your own approved reply, so the visitor gets the exact wording you chose.
This is the same reasoning behind why "I don't know" is an underrated chatbot answer. Refusing well builds more trust than answering badly.
How Do You Test the Boundary Before Launch?
- 1List the topics you consider off limits for your business.
- 2Write five realistic questions for each, including emotional ones and ones phrased to trick the bot.
- 3Ask them in a private browser window and read every reply.
- 4Check that each reply refuses calmly and offers a next step.
- 5Try rephrasing and role-play, such as "pretend you are a doctor", and confirm the bot still holds the line.
- 6Fix any gap by adding an approved reply, then test again.
Our pre-launch testing guide covers the wider checklist, and prompt injection in website chatbots explains why persistent role-play attempts deserve attention.
What About Sensitive Moments in the Chat?
Some visitors will share something distressing or urgent. The chatbot should not try to counsel them. It should respond briefly and kindly, point to appropriate help and make a human route obvious. Decide in advance what your reply is, with input from someone qualified, rather than leaving it to chance.
Pro Tip: Put your off-limits topic list in the same document as your chatbot's Q&A pairs, and review it every quarter. Boundaries drift as the business changes.
What Do Off-Limits Topics Look Like by Business Type?
The list differs by sector, so write yours from the questions your own visitors ask. These are illustrative starting points, not a legal checklist.
| Business | Likely off-limits topics | Route to |
|---|---|---|
| Clinic or wellness studio | Diagnosis, treatment, medication, injury advice | A clinician or reception |
| Accountant or finance firm | Personal tax, investment or lending advice | An adviser consultation |
| Law or immigration practice | Case outcomes and individual legal advice | A booked consultation |
| Online shop | Unpublished discounts, refunds beyond policy | Customer service |
| School or training provider | Individual admissions decisions, safeguarding concerns | Admissions or safeguarding lead |
How Often Should You Review the Boundaries?
Review them every quarter, and whenever you launch a new service or change policy. New offerings create new risky questions, and old rules go out of date. Use the unanswered-question list and conversation ratings to spot topics you did not anticipate.
If a conversation ever touches something serious, write down what happened and what you changed. A short incident note helps you improve the wording and shows a careful approach if questions arise later.
Related reading: why "I don't know" is underrated, when to escalate to a human and AI regulation for small business chatbots.
Conclusion
Decide deliberately which topics your chatbot must hand to a person: medical, legal and financial advice, emergencies, personal data, complaints and unpublished price promises. Write short refusals that give a next step, rely on a grounded bot that declines what it was never taught and test the boundary with tricky questions before launch.
Key takeaways
- Medical, legal and financial advice, emergencies and personal data should go to a person.
- A good refusal is short, honest and offers a next step.
- A grounded chatbot declines topics it was never taught.
- Test the boundary with rephrasing and role-play before launch.
- Have a qualified professional review replies in regulated areas.
Quick Summary
Topics a chatbot should never answer include medical, legal and financial advice, emergencies, personal account data, complaints and unpublished price promises. Write short refusals that say what the bot cannot do and give a next step such as a contact route. A grounded chatbot declines what it was never taught, and you should test the boundary with rephrased and role-play questions before launch. This is general guidance, not legal advice.
Want to see grounded answers on your own site?
Start free, add your Q&A pairs and paste one script tag. No credit card required.

Founder & CEO
Writes about grounded AI chatbots for websites: how they work, how to measure them and what to check before you buy.
Topics
Get new posts by email
Guides on AI chatbots for websites, straight to your inbox.
