An AI chatbot for a small business helps when it answers a narrow set of questions quickly or catches enquiries out of hours, and it hurts when it stands between a buyer and a person.
The technology is no longer the weak point. The weak points are what the bot is allowed to say, how it hands over to a human, and whether anyone reads the transcripts afterwards.
This article covers where a chatbot earns its place on a small business website, where it loses you enquiries, a real liability case worth knowing about, and a four-week test you can run before signing a contract.
Do customers actually want to talk to a chatbot?
Most do not, and UK evidence points the same way: in a 2025 survey of 2,000 UK adults commissioned by the communications vendor 8x8, 83% preferred a human agent and 4% preferred a chatbot.
The figures were reported by PPC Land. Treat them with the usual caution, since a vendor paid for the research, but the direction is hard to argue with. People will use a bot when it is faster than waiting, and will leave when it is in the way.
That gives you the design rule for everything below. A bot should be the quickest route to an answer, never a gate.
Vendors will quote conversion lifts of two or three times. Those numbers often come from vendor-funded studies and rarely say what the baseline was. Ask any supplier for the method behind a figure before you believe it.
Where does an AI chatbot earn its place?
It earns its place on questions with a known answer, on out-of-hours enquiries and on the first step of a booking.
Known-answer questions are the tidy category: opening times, parking, what a service includes, how to rebook, what documents to bring. A visitor asking those wants the answer in five seconds, and a bot trained on your own pages can give it. If the same five questions fill your inbox every week, that is your use case.
Out-of-hours capture is the second. A visitor at 9pm on a Sunday can leave their details, the bot asks two qualifying questions, and the lead is waiting with a note when you open on Monday.
For clinics and dental practices, where patients often browse in the evening, a bot that can offer booking slots directly is more useful than a contact form that promises a call back.
The third is the first step of a booking or quote. If it collects the postcode, the job type and a photo, whoever picks it up already has what they need. That saves a call, not a human.
None of these asks the bot to be clever. They ask it to be quick, accurate and polite.
Where does a chatbot lose you enquiries?
It loses them when it blocks a human, answers questions it should not, or pops up on pages where visitors already know what they want.
The first is the loop. The visitor types "I want to speak to someone", and the bot replies with three options that are not a person. Everyone has met this, and it ends with the tab closing.
Every chatbot should offer a visible route to a phone number or a call-back request in the first message, and should hand over automatically when it fails to answer twice.
The second is overreach. A bot that is happy to explain your refund policy, quote a price or comment on a legal question will, sooner or later, say something wrong.
The third is placement. A pop-up on a pricing page or checkout adds friction at the exact moment someone was about to act. If your best pages convert well already, put the bot on the ones that do not, such as help pages and long service pages, and leave the rest alone.
Who is responsible when the chatbot gets it wrong?
You are. In February 2024 a Canadian tribunal ruled that Air Canada was liable for wrong information its website chatbot gave a customer, and rejected the argument that the bot was a separate legal entity.
The facts, as the American Bar Association's summary of the Moffatt case sets out, were simple. The chatbot told a grieving passenger he could claim a bereavement fare after travel, which was not the airline's policy.
The British Columbia Civil Resolution Tribunal found the bot was "just a part" of the airline's website and ordered it to pay about $650 CAD in damages, plus interest and fees.
The tribunal also said it made no difference that the correct policy was elsewhere on the site. The ABA's account of the tribunal's decision is short and worth reading.
It is a Canadian ruling and it does not bind a UK court. But the reasoning is intuitive enough that you should assume a UK customer, or a UK regulator, would take the same view.
For a small business, the practical consequence is a short list. Restrict the bot to answers drawn from pages you control. Stop it from quoting prices or making commitments unless it is reading them from a live source. And for regulated work, such as law firms, limit it to intake and scheduling, with no opinions on anyone's case.
What does a chatbot do with the personal data people type in?
It collects it, stores it and, if it runs on a third-party language model, may send it to that provider, which makes the chatbot a UK GDPR decision as much as a marketing one.
Chat transcripts often hold names, phone numbers and details of health or legal problems. Decide how long you will keep them, who can read them and which provider processes them. Make sure there is a data processing agreement in place, and say clearly in the chat window that the visitor is talking to an AI.
Our guide to UK GDPR and AI tools explains what you can and cannot send to a model, and it applies directly here.
What does a chatbot cost to run?
Expect a monthly subscription, usage charges that rise with conversations, and staff time to keep the answers right.
Suppliers usually price the first. The second depends on the model underneath: some bundle it, others pass the cost of each conversation through. The third is the one that gets forgotten. Someone has to update the bot when your prices, hours or services change, read the transcripts and fix the pages it draws from.
Ask for a monthly estimate at your current traffic and at double it, and ask who pays when the underlying model provider changes its pricing. A supplier who cannot answer has probably not run one at your scale.
Do not forget the cost of a wrong answer, which can be a refund, a complaint or a lost customer. That is why the narrow version of a bot is safer and cheaper to run.
How do you test a chatbot before you commit?
Run it for four weeks on a limited set of pages, read every transcript in the first week and measure what changes on the enquiries that matter.
Week one, collect the twenty questions your team answers most often and check that the bot gets them right. Week two, turn it on for out-of-hours only. Week three, add one or two pages. Week four, compare enquiry numbers and, more importantly, quality: how many of the leads the bot passed on turned into a conversation.
Three numbers tell you most of what you need:
- the share of chats where the bot handed over to a person, and how many of those handovers worked
- the number of wrong or invented answers you find in the transcripts
- the change in completed enquiries, compared with the same weeks before
If the second number is above zero and you cannot fix it by editing the source pages, switch to a simpler, rule-based bot or drop the idea.
Our guide on responding to new leads in under two minutes covers the case where speed, rather than conversation, is the problem. A fast automated email and a text to your phone often beat a chat window.
Is a chatbot the right first automation?
Usually not, and the businesses that save the most hours start somewhere quieter.
Enquiry acknowledgement, quote follow-up and invoice chasing return time without putting anything in front of customers. A chatbot is not among them.
If you do want one, start with the narrowest version: opening hours, bookings, one service line. Widen it only when the transcripts say it is ready. For a wider view of the options, including voice assistants and agents, see what people mean when they search for an AI agency.
We design and build chatbots as part of our AI automation services, including the hand-off to a person and the transcript review. If you would like to talk through whether one fits your website, get in touch.
