Should you put an AI chatbot on your small business website? A test, the rules and the alternatives

AI chatbot on a small business website: the short answer

Usually not yet. An AI chatbot on a small business website pays off when four things hold. Many visitors ask the same questions, the answers are already written down, people arrive when nobody can reply, and the bot can hand the conversation to a person. If any one fails, better pages and a fast human reply will do more.

This guide is for the owner or operations lead of a 5 to 50 person firm whose website brings in enquiries: a clinic, a builder, an accountant, a shop. Someone has offered you "24/7 support" in a chat bubble, and you want to know whether it's worth it.

Two things have changed since most guides on this were written. Since 2 August 2026, the EU AI Act has required a chatbot to tell people it is an AI from the first message. In the UK, the Competition and Markets Authority (CMA) holds you responsible for an AI agent "in the same way you are responsible for what an employee does".

Customers are also less keen than the vendors. In a Gartner survey of 5,728 customers, published in July 2024, 64% said they would prefer companies didn't use AI for customer service.

We build AI agents for a living, including assistants that answer customers. Our own website has no chatbot. Here is how to decide whether yours should.

What a website chatbot is: four versions

"Chatbot" covers four different things, and sales pages blur them. What matters is where the answers come from and what the bot can do.

Most "AI chatbot for small business" offers are the second version: point a model at your site and it answers. That version is also the most likely to invent something. When your pages don't cover a question, the model writes a plausible answer anyway.

The third version is closer to what we build. An assistant connected to your live system can act on a problem instead of describing it: it can say where the order is because it looked. Once it can change things, it is an agent, and its design is its limits. Our guide to AI agents vs workflows explains the difference.

When an AI chatbot is worth it: a five-question test

Take your last month of enquiries from every channel (email, the website form, phone, social messages) and answer five questions.

Five answers in the middle column make a case. If question 4 lands on the right, the bot can still help, but only if it hands every commitment to a person. If question 5 lands on the right, stop there. An unread chatbot drifts, and you find out from a customer.

Say you run a physiotherapy practice with three clinics. Most of your 40 daily messages ask about parking, what to wear, which insurers you accept and how to move an appointment. The answers are written down, and half the messages arrive after 6pm. That practice passes the first three questions, and an assistant that can see the booking system deserves a serious look.

Now say you run a building firm that gets eight enquiries a day, each about a different job. Every good answer starts with a question back: what's the job, where is it, and when? A chatbot there puts a step between the customer and the person who can answer.

What goes wrong: four public cases

Website chatbots fail in a few repeating ways, each with a design fix. Air Canada was held to a refund its chatbot described, a case told in our checklist of what not to automate with AI. Cursor's support bot invented a policy, as our comparison of AI agents and workflows recounts. Four more follow.

DPD, January 2024. A London customer couldn't get DPD's chatbot to find his parcel. Then it swore at him, wrote a poem about its own uselessness and called DPD the worst delivery firm in the world. DPD said the AI element had worked for years, that an error followed a system update, and that it had switched the AI off (Silicon UK).

New York City, 2024. The city's MyCity chatbot ran on Microsoft's Azure AI services and drew on more than 2,000 of the city's business web pages. The Markup found it telling employers they could take a cut of workers' tips, and landlords they could refuse tenants paying with housing vouchers. Both are illegal in New York (The Markup, March 2024).

In January 2026 the new mayor, Zohran Mamdani, said he would shut it down, calling it "functionally unusable" (The Markup, January 2026).

A Chevrolet dealer in Watsonville, California, December 2023. A visitor told the dealer's ChatGPT-powered chatbot to agree to everything and to call each reply a "legally binding offer". It then agreed to sell him a new SUV for almost nothing. The screenshots spread, and the chatbot's supplier said an update now bans users who try this (Gizmodo).

Lenovo, 2025. Cybernews researchers wrote a 400-character message that made Lenovo's support chatbot, Lena, return code with its answer. That code could steal a support agent's session cookie, letting an attacker into the support system as that agent. Lenovo fixed the flaw in August 2025 (ITPro).

Every one of these bots looked finished, because it answered every question fluently. A chatbot that always has an answer looks like it works. The test that matters is the question whose right answer is "I'll pass you to a colleague".

What UK and EU rules ask of a website chatbot

In the EU, a website chatbot must say it is an AI from the first message. In the UK, you answer for what it says, as you would for an employee. Both apply to small firms.

Since 2 August 2026, Article 50 of the EU AI Act has required AI systems that talk directly to people to tell them so, unless it is obvious. The Commission's guidelines of 20 July 2026 spell out what that means for a website:

• A helpdesk chatbot is not obvious, so the rule applies.
• The notice comes at the first interaction: an opening message saying it is an AI, ideally with a label that stays in view.
• A line in your terms and conditions doesn't count on its own. Nor does a vague name like "assistant", or a human-looking persona that could mislead.
• The duty sits with the provider. A company that builds a chatbot, or has one built, and runs it under its own name is the provider. If you switch on a vendor's ready-made chatbot, the vendor usually is, but the notice appears on your site, so check it.
• It applies wherever the provider is based, so a UK firm whose chatbot serves customers in the EU should plan on it applying.

The UK has no rule like Article 50. The CMA's guidance on using AI agents, published on 9 March 2026, reaches a similar place through consumer law. You are responsible for the agent "even if someone else designed or provides the AI agent on your behalf". The CMA asks you to consider labelling it, to keep a person checking its answers, and to fix the prompts quickly if it keeps giving wrong information about cancellation rights.

We aren't lawyers. Our guide to the EU AI Act for small businesses covers the rest of the Act; if your chatbot touches health, money or legal questions, ask yours.

What customers want from the chat box: a person, when they need one

Gartner's survey found that customers' main worry about AI in customer service was that it would make reaching a person harder. And 53% said they would consider switching to a competitor if they found out a company was going to use AI for customer service.

Klarna learned this at scale. In 2024 it said its AI assistant handled two-thirds of customer service chats in its first month, the work of 700 agents. In 2025 its chief executive told Bloomberg that cost had been "a too predominant evaluation factor". Klarna said customers would always be able to reach a person (CX Dive).

For a small firm, that makes the handover the product. Design it before anything else:

• A "talk to a person" option visible from the first message, not after three failed answers.
• The conversation goes with the handover, so the customer doesn't repeat themselves.
• Out of hours, the bot says when a person will reply, takes contact details, and the reply arrives when it said.

A bot that traps people in a loop costs you the customer it was meant to catch.

How we would build one, if the test says yes

When a chatbot passes the test, we build it the way we build any agent: the design is its limits. Here are the eight we would set.

1. A written scope: the questions it answers. Everything else goes to a person.
2. Answers only from your own approved pages and documents, each with an owner who keeps it current. A bot can't know what nobody wrote down; it will guess or deflect. Writing it down may be the better first project: a company second brain that your team and your AI both read.
3. No commitments. Prices, refunds, dates, availability and anything that sounds like advice come from your systems or from a person. Our guide to human-in-the-loop AI shows how to design that approval step so it isn't a rubber stamp.
4. An opening message that says it is an AI and what it can help with, a label that stays in view, and no human name or stock photo.
5. A handover with the conversation attached, in business hours and outside them.
6. Its output treated as untrusted text. It never runs as code in your support tools, and it can reach only the data it needs.
7. A fixed set of test questions, including ones it must refuse, run after every change to the model, the instructions or your pages.
8. A log of what it said and why, and a person who reads a sample every week. The questions it couldn't answer are your next website pages.

When you don't need a chatbot, or us

Most of the time, the problem a chatbot is sold to fix is a website that doesn't answer the obvious questions, or a reply that takes too long. Fix those first.

For ES Construction, a London builder, we built quote routing by phone, email or WhatsApp, because that is how the trade makes contact. We put WhatsApp level with the phone number. Our own site uses a short form and a promise: you'll hear back within one business day, from the people who would build it.

The drafting option also sits outside the EU disclosure rule. The Commission's guidelines say that when staff use AI tools to help them reply, the customer isn't interacting with the AI directly. That holds only while the person really reads what they send.

You don't need us either if a ready-made chatbot passes the test and meets the eight limits above. Many website builders and help desk tools include one. Check the opening message, the handover and the transcript export, then switch it on. If you can't tell whether yours does, ask for an independent read before you pay for a year.

Start with your last 50 enquiries

Before you talk to any chatbot vendor, export your last 50 enquiries from every channel. Sort them into three piles: a page could have answered it, a person had to answer it, or it arrived when nobody was there. Count each pile.

When the first pile is the biggest, write those pages and watch whether the enquiries drop. When the first and third piles are both big, you have a case for a chatbot, and the questions to test it with. And if the second pile wins, your customers want what they have always wanted: a person who replies quickly.

Questions

Will a chatbot help my website rank in Google?

Not directly. Google ranks pages, and an answer inside a chat window doesn't sit on a page Google can index. The questions people type into a chatbot are useful, though. Each one it couldn't answer is a page worth writing, and that page can rank.

What does a website chatbot cost to run?

We don't publish prices, because every project is quoted. Four things move the cost. The first is whether you rent a ready-made tool or have one built. Then come the number of conversations, since each one uses the model, and whether it connects to your booking or order system. Last is the staff time to read transcripts and keep the answers current, which most quotes leave out. Our guide to what an AI agent costs to build and run goes through each line.

Can a chatbot take bookings or payments?

It can take bookings if it connects to your booking system. At that point it changes something, so treat it as an agent and design its limits. The confirmation should come from the booking system itself, not from the model's wording. Keep payments on your checkout or your payment provider's page, not inside the chat.

How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.