Should you put a chatbot on your website?

Short answer: a chatbot is worth it when people keep asking you the same questions and the answers already exist somewhere in writing. It isn't worth it when every question is different, when a wrong answer could cost someone money or hurt their health, or when only a handful of people visit your site each week.

What a chatbot is good at

Most small businesses get the same ten or twenty questions over and over. What are your opening hours, how much is delivery, do you ship abroad, can I change my booking, how long will it take. The answers rarely change, and typing them out for the hundredth time is nobody's favourite part of the day.

A chatbot handles these well because it never gets tired of them and it answers at 3 a.m. on a Sunday. Bots built on today's language models also understand questions phrased in odd ways, with typos or half a sentence, and they can reply in the visitor's own language. If your customers come from more than one country, that alone can be worth a lot.

You can also give the bot everything on your site: the pages, the prices, the FAQ, your terms, even guides and documents that aren't public. When someone asks a question, the bot first searches that material and then writes its answer from what it found. The method is called RAG, short for retrieval-augmented generation, and it's what lets a bot answer almost anything about your business with your own information. When you update the site, you update what the bot knows too.

The best-known numbers come from Klarna, the payments company. In February 2024 it announced that its AI assistant had handled 2.3 million conversations in its first month, two thirds of all its customer service chats, and that the average time to resolve an issue had dropped from 11 minutes to under 2. Those are the company's own figures, from a business with millions of customers asking very similar things, which is close to the ideal case for a bot.

Where it goes wrong

The same models that write fluent answers also write fluent wrong answers, and they can't tell when they're guessing. A few of the results have become well known.

In late 2022, after his grandmother died, Jake Moffatt asked Air Canada's chatbot about bereavement fares. The bot told him he could buy a full-price ticket and apply for the discount afterwards. That wasn't the airline's policy, and Air Canada refused the refund. He took the case to British Columbia's Civil Resolution Tribunal, where the airline argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal called that a remarkable submission. In February 2024 it ruled that the chatbot was simply part of Air Canada's website and ordered the airline to pay CA$812.02 in damages, interest and fees. So a company answers for what its chatbot tells customers, the same as for any other page on its site.

In December 2023 a Chevrolet dealership in Watsonville, California, had a ChatGPT-based assistant on its site. Visitors soon found they could give it instructions of their own. One of them got it to agree to sell him a new Chevy Tahoe for one dollar and to call the deal legally binding. Nobody drove away with a car, but the screenshots spread everywhere, and the dealership took the bot down.

In January 2024 the delivery company DPD switched off the AI part of its customer chat after an update went wrong. A customer who couldn't track down his parcel started testing the bot, asked it to swear, and it did. It then wrote him a short poem about DPD being useless.

Two months later, the newsroom The Markup tested the chatbot New York City had launched to help people start and run businesses. It told employers they could take a cut of their workers' tips, and told landlords they didn't have to accept tenants who pay with housing vouchers. Both are illegal in New York City.

Searching your own material helps, but it doesn't make a bot perfect. A team at Stanford tested legal research tools from LexisNexis and Thomson Reuters, which work this way. GPT-4 on its own gave wrong or invented information in 43 percent of the test questions; the tools that searched real legal documents first got that down to between 17 and 33 percent. The researchers point out that law is a hard case, because legal questions often have no single clear answer in the documents. Questions like your opening hours or your delivery price usually do.

What your customers think about it

Gartner surveyed 5,728 customers at the end of 2023. 64 percent said they would prefer companies not to use AI for customer service, and 53 percent said they would consider moving to a competitor if they found out a company was going to. Their biggest worry was that it would get harder to reach a real person.

Klarna seems to have reached the same conclusion. In May 2025 it started hiring human agents again, and its CEO told Bloomberg that cost had been "a too predominant evaluation factor" and that the result was "lower quality". The assistant is still in use, and customers who need a person can now reach one.

Signs it makes sense for you

You probably have a good case for a chatbot if most of the messages you get are variations of the same few questions, and the answers are already written down, on a FAQ page or a price list. It also helps if people write to you outside working hours and then wait until morning for something simple, or if you regularly get messages in languages you don't speak well.

You probably don't need one if you get a few messages a week and answer them quickly yourself. The same goes if most questions need you to look at something specific, like a photo or someone's order, or if a wrong answer could hurt someone, as with medical, legal or financial advice. And if your prices or availability change every day and nobody will keep the bot's information up to date, it will end up confidently quoting last month's prices.

If you add one, set it up like this

Give it your information and nothing else. A bot that answers from your own FAQ, price list and policies makes far fewer things up than one left to answer from general knowledge. Tell it to say it doesn't know when the answer isn't there, then ask it a few questions it can't answer and check that it really does.

Make a human easy to reach. An email address or phone number in plain sight deals with the biggest worry in the Gartner survey, and it catches the questions the bot gets wrong before they turn into complaints.

Say clearly that it's a bot. Since 2 August 2026, the EU AI Act requires chatbots that talk to people in the European Union to make it clear they are AI, unless that's already obvious. The rule covers anyone whose chatbot reaches people in the EU, wherever the business itself is based. One short line at the start of the chat is enough, and it's the honest thing to do anyway.

Don't let it make promises. Prices, refunds, discounts and deadlines should come from your written terms, and the bot should send people there. The Air Canada and Tahoe stories both started with a bot saying yes to something nobody had approved.

Read the conversations, especially in the first few weeks. You'll find gaps in what you gave the bot and catch wrong answers while they're still easy to fix.