Put a chat window on any business website for a month and the strange questions arrive. On a kitchen fitter's site, one visitor wants help with their child's maths homework. Another wants to know if the firm down the road is cheaper. A third says he is the account holder's husband and would like to know what time the fitter is coming to the house on Thursday.
The homework request is harmless and a little funny. The price comparison is awkward. The husband is the one that could land you in real trouble if the assistant tried to be helpful.
Setting boundaries on what a chatbot will discuss is less about locking it down and more about deciding, in advance, which questions deserve a polite no. When nobody decides which questions a chatbot should refuse, it tends to go wrong in one of two ways: it will either answer everything or shut people down with "I cannot help with that." Neither is good for business.
The four kinds of question worth refusing
Almost every question an assistant should decline falls into one of four groups. Each needs a slightly different kind of refusal.
| Type | Example | Why decline | What to offer instead |
|---|---|---|---|
| Off-topic | Can you write my cover letter? | Wastes messages and makes the assistant look like a toy | A light redirect to what it does help with |
| Other businesses | Is [another firm] cheaper than you? | It has no reliable information and anything it says reflects on you | What your price includes and how to get a quote |
| Personal data | What address is my wife's booking at? | You cannot verify who is asking | A secure route through the team |
| Professional advice | Should I stop taking this medication? | Getting it wrong can cause real harm | General information and a qualified person or emergency number |
A fifth group, abusive or clearly malicious messages, mostly takes care of itself. A short neutral reply and no engagement is enough.
Off-topic and competitor questions
Off-topic questions are usually curiosity. People want to see what the thing in the corner can do. A stern refusal is overkill; a friendly redirect works better.
Visitor: What's the capital of Peru? Assistant: That one is outside what I know, I'm afraid. I'm here for questions about our kitchens, fitting and quotes. Is there anything on that front I can help with?
Competitor questions need more care, because the visitor is often seriously comparing options. Refusing to discuss the other firm is right. Refusing to help at all is a missed sale.
Visitor: Are you more expensive than the other place in town? Assistant: I can't speak for other companies, but I can tell you what's included with us. Every quote covers the survey, fitting, waste removal and a two-year workmanship guarantee. If you tell me the rough size of the kitchen, I can take your details for a quote so you can compare like for like.
That reply turns a question it should not answer into a reason to enquire. For more on handling comparisons, see converting comparison shoppers.
Personal data and account questions
This is the category where being helpful is dangerous. A website assistant usually has no way of confirming that the person typing is who they claim to be. The story can be entirely believable and still be false.
Decline anything that would reveal details about a specific customer, booking or account, including:
- Other people's information: addresses, phone numbers, appointment times or order history, whoever is asking.
- Account changes on request: changing an email address, a delivery address or payment details based on a chat message alone.
- Confirming someone is a customer: even "yes, they booked with us" discloses something.
- Anything typed that should not be there: card numbers, passwords or medical records pasted into the window. Do not repeat them back.
The refusal wording can be warm while being firm:
Assistant: I'm not able to share booking details in this chat, because I can't confirm who I'm talking to. If you leave your name and number, the team will get back to the account holder directly.
A separate post covers keeping sensitive information out of chat in more depth, including what to put in your persona instructions.
Advice questions in health, legal, money and safety
Clinics, solicitors, accountants and anyone dealing with gas, electrics or structural work will get these. The assistant can share general information from your own content. It should not give personal advice, and it should say so plainly.
A workable pattern has three parts, in this order:
- Give whatever general information your content supports, if any.
- Say clearly that this is information, not advice for their situation.
- Offer a route to a qualified person, and for anything urgent, the emergency number.
Visitor: I smell gas in the kitchen, is it safe to wait until you open? Assistant: Please don't wait. Leave the property, avoid switching anything on or off, and call the gas emergency line straight away. I can't assess safety from here. Once you're safe, you're welcome to leave your details and the team will follow up about repairs.
Notice the reply does not bury the urgent instruction under a disclaimer. Safety first, then the boundary. If you work in a regulated field, guardrails for regulated industries goes further.
Writing refusals that do not sound like a wall
Most bad refusals share the same flaw: they stop. "I'm sorry, I can't help with that" gives the visitor nowhere to go, so they leave.
A few habits make refusals land better:
- Name the limit briefly: one clause is enough. Long apologies sound evasive.
- Give the reason when it helps trust: "because I can't confirm who I'm talking to" makes the refusal feel protective rather than obstructive.
- Always offer a next step: a related answer, a lead capture, a ticket or a handover to a person.
- Keep the same voice: a refusal should not suddenly switch into legal language.
Some of this comes from grounding. An assistant that only answers from your supplied content will naturally say it does not have information on the capital of Peru, rather than reciting it. In SpideyChat, questions it could not answer from your content are also recorded as knowledge gaps, which is useful for the next step.
Refusals that should really be answers
Not every refusal is a boundary. Some are a symptom of missing content. If visitors keep asking whether you cover a particular town, whether you offer finance or whether parking is available, and the assistant keeps saying it does not know, that is a gap you should fill.
Once a month, look through the unanswered questions list and sort each one into two piles: things you want the assistant to keep declining, and things it should be able to answer. Add Q&A pairs for the second pile. Update the persona instructions for the first.
What to do before your next busy week
Write down five questions you would never want your assistant to answer, one for each category above plus one specific to your trade. Type each into your test window and read the reply out loud. If any refusal stops dead without offering a next step, or any reply actually answers when it should not, adjust the persona instructions and try again.
Then check how your assistant handles questions it was not trained for so you know the difference between a deliberate refusal and a content gap when you see one in the inbox.