If your team spends part of every day answering the same questions about opening hours, prices, delivery times, what a treatment involves or whether you cover a postcode, a chatbot can take most of that load. I build AI chatbots for websites and WhatsApp that answer from the content you have approved, rather than from whatever a language model happens to know about the world.
Done well, a customer gets a correct answer at 10pm on a Sunday, your staff stop retyping the same reply, and anything the bot cannot handle reaches a person with the conversation attached.
Who this suits, and who it does not
It suits businesses with a steady flow of repeat questions and answers that are written down somewhere, or could be: clinics and salons, trades and home services, online shops, letting agents, training providers, venues and professional firms with a clear service list. It also suits businesses running click-to-WhatsApp ads that now receive more messages than anyone can answer quickly.
It is the wrong fit if most of your enquiries need judgement on a case-by-case basis, such as legal or financial advice, or if your prices and policies change daily and nobody owns keeping them current. A bot only answers as well as the material behind it. If you get a handful of messages a week, a well-organised FAQ page and a fast human reply will serve customers better.
Two ways to build a chatbot
Most chatbots fall into one of two kinds, and choosing the right one matters more than choosing the software.
A guided flow
The customer taps through buttons and menus that you have scripted in advance. It cannot go off script, which makes it predictable and cheap to run. It suits narrow jobs: booking a slot, checking an order, choosing between three services. It frustrates people who type a question in their own words.
An AI bot that answers from your content
The customer types naturally, and the bot searches your approved material and writes a reply from it. This approach is called retrieval-augmented generation: the model is handed the relevant passages before it answers, so it does not rely on memory. It copes with the long tail of questions a script never anticipates.
Where both kinds of question come in, I combine the two. Free-text questions go to the AI, and anything with a fixed path, such as a booking or a callback request, switches to a guided step so the details are captured correctly.
What the chatbot does
Answers from your own material
I build the knowledge base from your website pages, FAQs, price lists, policies and any internal notes you are happy to make public. Each answer is grounded in those sources, and where it helps, the bot links the customer to the page it drew on.
Admits what it does not know
A language model will produce a confident wrong answer if nothing stops it. This is known as hallucination, and the defence is in the build: strict instructions to answer only from the supplied content, a set reply for questions outside it, and testing against the awkward questions your staff actually receive.
Hands over to a person
When a customer asks for a human, gets frustrated, or raises something the bot should not handle, the conversation passes to your team with the full transcript. Keeping a person in the loop at these moments is what stops a chatbot damaging trust.
Captures the enquiry
If the customer wants a quote, a callback or a booking, the bot collects name, contact details and the reason for getting in touch, then sends them to your inbox or CRM so nothing is lost overnight.
Works on your website and on WhatsApp
On a website, the bot sits in a chat window on the pages you choose. On WhatsApp, it runs through Meta’s WhatsApp Business Platform on a number registered to your business. Both channels can share one knowledge base, so a price change is made once. If the messages are coming from click-to-WhatsApp and Messenger ads, the bot can greet each new conversation within seconds.
What I will not build it to do
- Pretend to be human. The bot says it is an automated assistant in its first message and whenever someone asks. Clear AI disclosure is honest, and it means nobody feels misled when the conversation passes to a person.
- Give medical, legal or financial advice. It can explain what a service involves and how to book it. It does not tell a customer whether a treatment is right for them or whether a contract term applies to their situation.
- Make commitments your staff have not approved. No discounts, refunds or delivery promises that are not already written in your policies.
- Collect sensitive details it does not need. Health conditions, card numbers and similar information are kept out of the chat, and the bot steers customers to the proper channel.
Chatbots and UK data protection
A chatbot that takes names, phone numbers and free-text messages is processing personal data, so UK GDPR applies. Before launch I check four things with you: that your privacy notice explains what the bot collects, why, where it is sent and how long it is kept; that conversation logs are kept no longer than you need them; that the AI provider and chat platform are covered by a data processing agreement; and whether any data leaves the UK, and on what basis.
Clinics, therapists and anyone else whose customers might mention their health need extra care, because that is special category data. For those sites I set the bot to discourage health details in chat and to move clinical questions to a person straight away. I am not a lawyer, and where your situation is unusual I will say so and suggest you take advice.
How a build runs
- Question audit. We collect a few weeks of real enquiries from email, phone notes, WhatsApp and contact forms, and sort them into what the bot should answer, what it should hand over and what it should refuse.
- Content clean-up. Gaps and contradictions in your existing pages are fixed first. This is the stage where a second, out-of-date price list or a conflicting returns policy tends to surface.
- Build. I set up the knowledge base, write the bot’s instructions, tone and set replies, and connect handover and enquiry capture.
- Testing. I run the bot against your real questions plus deliberately difficult ones, and you test it too. Wrong or vague answers are traced back to the content or the instructions and fixed.
- Soft launch. The bot goes live on a few pages or to a share of WhatsApp traffic first, with your team watching transcripts closely.
- Review. After launch we read the transcripts together, add missing answers and adjust the handover rules.
Problems I see most often
- Bots switched on with a single instruction to “be helpful”, which then answer from the open internet and quote prices the business has never charged.
- No route to a person, so customers type “human” five times and leave.
- Knowledge bases built from an old site, so the bot cheerfully describes services that ended last year.
- Chat windows that pop up over the content on mobile and block the page the customer came to read.
- Conversation logs kept forever in a third-party dashboard nobody checks, with no mention in the privacy notice.
What you receive
- A working chatbot on your website, WhatsApp or both, set up in accounts your business owns.
- The knowledge base and the bot’s written instructions, in a form your team can edit.
- Handover and enquiry capture connected to your inbox, CRM or helpdesk.
- A test log showing the questions checked and how each was resolved.
- A short guide to updating answers, reading transcripts and switching the bot off if you ever need to.
- Suggested wording for the chatbot section of your privacy notice, for you or your adviser to approve.
What it costs to build and to run
The build cost depends mainly on how many channels you need, how tidy your existing content is and what the bot must connect to. I write each quote in GBP after a free first call, and it is agreed in writing before the build starts.
Running costs are separate and paid by you directly to the providers: a chat platform subscription if one is used, AI model usage that rises with the number of messages, and Meta’s charges for some WhatsApp Business Platform messages. I estimate these from your current enquiry volume before you commit, so you know the likely monthly cost in advance.
Next step
Send me a list of the ten questions your team answers most, and tell me where those answers currently live. On a free 30-minute call I will tell you whether a chatbot is worth building for you, which kind it should be, and what I would fix in your content first.