Website chat should handle the routine part: working out which service the visitor wants, taking the details follow-up needs, and answering from information you've approved. A person should step in for unpublished prices, complaints, anything sensitive and any chat that's stuck, with the conversation so far already in front of them.
Website chatbots for small businesses tend to fail in one of two ways: they're so cautious they're useless, or they make things up. Both leave a transcript in an inbox and a visitor who has to start again. Here's what I'd let chat do on its own, where I'd draw the line, and what a good hand-off looks like.
Chat that only leaves a transcript
A chat widget can sit on the site and still do nothing useful. The visitor types the request. The bot replies with a generic line. The transcript lands in an inbox, and the visitor has to repeat themselves when a person finally reads it.
That's a novelty, not a path. You can't tell whether the bot helped or blocked the conversation, because nothing was captured into the system the team already uses.
What chat should handle first
Chat can take the routine part: which offer the visitor wants, the facts follow-up needs, and answers that come from sources you approved. If the bot doesn't know, it should say so instead of inventing a next step.
- Which offer they are asking about
- Name, work email, and what they want done
- Answers drawn from sources you trust
- A clear stop when the chat is stuck or sensitive
"Sources you trust" is the bit that matters most. In 2024, a Canadian tribunal ordered Air Canada to pay compensation after its website chatbot wrongly told a grieving passenger they could claim a bereavement fare after booking. The airline argued the chatbot was responsible for its own actions; the tribunal called that "a remarkable submission" and said it makes no difference whether the information comes from a static page or a chatbot. It's a Canadian small-claims case, not UK law, but the principle is the one I'd work to: whatever your bot says, you said.
So feed it your actual service list, published prices, opening hours, service areas and FAQs, and nothing else. If a question falls outside that, the right answer is "I'll get someone to confirm that for you."
When a person steps in
A person steps in for sensitive questions and for chats that are stuck. A price you haven't published, an unhappy customer, and anything that needs a judgement call belong with that person. I've written more about the decisions you shouldn't hand over to AI.
Write the hand-off rule before you turn the bot on. The visitor shouldn't retell the whole story. The person should see what was already captured.
Be upfront that it's a bot, and say what happens to the conversation. The ICO's guidance on AI and data protection expects you to be transparent about how you use people's information, and a chat transcript with a name and email in it is personal data like any other enquiry. A line at the start ("I'm an automated assistant. A member of the team will pick up anything I can't answer.") plus a link to your privacy notice covers most of it.
A sample hand-off, start to finish
Here's a hypothetical chat on the site of a 10-person cleaning firm, to show where the line falls.
Visitor: Do you do end of tenancy cleans in Bristol? Chat: Yes, we cover Bristol for end of tenancy cleans. How many bedrooms is the property, and when do the tenants move out? Visitor: 3 bed, they're out on the 14th. Can you do it for £200? Another company quoted that. Chat: I can't confirm prices for individual jobs, but someone from the team can. Can I take your name and the best email or number to reach you? Visitor: Ana, ana@example.com Chat: Thanks, Ana. I've passed on: 3-bed end of tenancy clean in Bristol, tenants out on the 14th, and that you've had a quote of £200. Someone will reply by 10am tomorrow.
What the person picks up: one enquiry record with the service, location, size, date, the competitor price, and Ana's contact details. They don't need to ask anything twice. The bot answered what it knew (coverage), took what follow-up needed, and stopped at the price.
Keep chat on the enquiry path
The chat should write into the same enquiry and follow-up path as the form. A second inbox for bot transcripts splits the lead the same way a stray form does.
In over five years of business analysis, the pattern I see most is a new channel bolted on with its own inbox. If you're adding phone answering too, the same rule applies; see what a voice agent should capture from the call. Before any of it goes live, test the agent before customers see it. For the wider picture, browse our guides.
See Conversational AI and chatbots for chat that follows your offers, hands a stuck or sensitive thread to a person, and records it on the enquiry path you already run.
Questions people ask
What should website chatbots for small businesses automate?
Working out which service someone wants, collecting their details and the facts follow-up needs, and answering routine questions from information you've approved, such as service areas, opening hours and published prices.
When should a chatbot hand over to a person?
For unpublished prices, complaints, anything sensitive or involving a judgement call, and any chat that's going in circles. The person should see the conversation so far.
Am I responsible for what my chatbot says?
Assume so. In the Air Canada case, a Canadian tribunal held the airline responsible for its chatbot's wrong answer. Limit the bot to approved material, and have it hand over when it doesn't know.
Should I tell visitors they're talking to a bot?
Yes. Say it at the start, explain that a person will pick up anything it can't answer, and link to your privacy notice.
Sources
- Air Canada found liable for chatbot's bad advice on plane tickets (CBC News)
- Decision (CBC News)
- Guidance on AI and data protection (Information Commissioner's Office)