Almost everyone has had a bad experience with a chatbot. You arrive with a specific question, get offered four irrelevant options, type the question anyway, and receive a confidently wrong answer before finally finding the link to speak to a person. The whole category has earned its reputation.
I build these things, and I still think most businesses that ask for one should not have one. Here is how I decide.
The one thing that makes a chatbot useful
Volume of repeated questions. That is it. That is the whole test.
If a meaningful number of people arrive at your website every week wanting to know the same handful of things — your opening hours, whether you cover their postcode, what happens on collection day, whether a particular product fits a particular model — then a chatbot is answering a real question that a person would otherwise have to answer. It saves your time and it saves the customer waiting.
If people arrive with genuinely varied, specific questions, a chatbot has nothing to do except get in the way of the contact form.
Where they work well
Councils and public bodies. Genuinely enormous volumes of repetitive questions, heavily seasonal, and the answers are all published somewhere already. This is close to a perfect fit, and it is why one of the chatbots I run is for a town council.
Businesses with a long FAQ nobody reads. If your FAQ page has thirty entries, you have already admitted people ask a lot of questions. A chatbot is a better interface for that content than a page of accordions.
Service businesses with coverage and eligibility rules. “Do you deliver to Shropshire” and “can you fit this in a flat” are questions that stop people enquiring. Answering them instantly means the enquiries you do receive are better qualified.
Where they do not
Low traffic. If forty people visit your site a week, you do not have a volume problem. Put your phone number in the header instead.
Complex or emotive purchases. Nobody wants to discuss a funeral, a legal problem or a £40,000 project with a bot. It reads as though you could not be bothered.
As a replacement for clear content. If people are asking a question because the page is confusing, fix the page. A chatbot layered over bad content is an expensive workaround.
What separates a good one from a bad one
If you do decide to have one, three things matter far more than which model is behind it.
It answers from your content only. The failure mode everyone hates is invention. A properly built chatbot retrieves the relevant passage from your own approved material and answers from that, and says it does not know when it finds nothing. That behaviour is a design decision, not a happy accident.
It hands over to a human quickly. The moment the question needs a person, or the visitor asks for one, it should get out of the way and pass the conversation on. A bot that traps people is worse than no bot.
Somebody reads the transcripts. This is the one that gets skipped. The conversation log is the most honest customer research you will ever get — it is a list of exactly what people wanted to know and could not find. If nobody reads it, you have bought a gadget rather than a tool.
So should you?
Count the repeated questions. If your team answers the same ten things over and over, it is probably worth it. If they do not, save your money — and be slightly suspicious of anyone who tells you otherwise without asking about your volumes first.