Should you put an AI chatbot on your website?

We get asked for these most weeks now, and we build them. We also talk about half the people who ask out of it, so this is the conversation we would have had anyway, written down.

The question that decides it

Do you get the same questions over and over, and are the answers written down somewhere?

Both halves matter.

If people ask you twenty variations of “do you deliver to Hyderabad”, “what are your hours in Ramzan” and “do you take card”, and those answers exist in a document, a chatbot has something real to do. It answers instantly at 2am, and you stop retyping the same four sentences.

If every enquiry is different, or the answer depends on judgement — a quote, a diagnosis, whether your product suits their situation — a chatbot mostly gets in the way of a conversation that was going to convert.

And if the answers are not written down anywhere, that is the actual project. You cannot skip it. A chatbot is a way of serving a knowledge base; it is not a substitute for having one. Half the value people get from building one comes from finally writing down what they know, and they would have got that value from a good FAQ page for a fraction of the cost.

What they are genuinely good at

The same question, at any hour. The one unambiguous win. Businesses in Pakistan selling to the Gulf, the UK or the US get enquiries in the middle of the night, and an instant correct answer beats a reply at 10am.

Getting to the right page. On a site with forty pages, “where do I find the return policy” is a real problem and a chatbot solves it better than a menu.

Qualifying before a human arrives. Collecting what you need — budget, timeline, city — so the first human reply is useful rather than a request for information.

A language you do not staff for. A model answering in Urdu, English and Arabic from the same knowledge base is genuinely cheaper than three people.

The four ways they go wrong

1. It answers confidently and wrongly. The dangerous one. A model asked something outside what it was given will often produce a plausible answer rather than admit it does not know. On a website that means quoting a price you do not charge or promising a delivery time you cannot meet — to a customer, in writing, in your name. Businesses have been held to what their chatbot said. Anything the bot cannot support from your own documents, it should refuse, and you should test that it does by asking it things you never told it.

2. It becomes a wall. The pattern everyone has suffered: a bot that will not let you reach a person, on a site where the bot cannot help. If yours cannot answer, it should hand over immediately and visibly — a WhatsApp link, a form, a number. A chatbot that traps people costs more customers than it serves.

3. It goes stale. Prices change and the knowledge base does not. Now a confident bot is confidently quoting last year. Whoever owns updating it needs to be a named person, not “we’ll keep it updated”.

4. It slows the site down. Most hosted widgets ship a substantial JavaScript bundle from a third-party domain, and it loads for every visitor including the 90% who never open it. On a phone on mobile data that is a real cost paid by everyone to serve a few. If you add one, load it on interaction rather than on page load, and measure the difference rather than assuming.

What it actually costs

The widget is the cheap part. The real costs, roughly in order:

• Writing the knowledge base. The largest, and the one that gets left out of quotes. Somebody has to assemble the answers and keep them right.
• Testing the refusals. Deliberately asking things it should not answer, and confirming it declines rather than invents. This is the step that gets skipped and it is the step that prevents the expensive failure.
• Per-message API cost. Small per message and real at volume, and it varies by which model you use — and the models keep changing, so build so you can switch.
• Maintenance. Ongoing, forever, and it is where these quietly die.

A chatbot that nobody maintains is worse than no chatbot, because it is a confident source of stale information wearing your logo.

The cheaper thing to try first

Before building anything:

1. Write the FAQ page properly. Ten real questions with complete answers. This is the knowledge base you would have needed anyway, it costs almost nothing, it helps in search, and it is what an AI assistant can quote when someone asks about you elsewhere.
2. Put WhatsApp on the site. In this market it converts better than most chat widgets, it is free, and the customer is already in the app. A great many “we need a chatbot” conversations end here, correctly.
3. Read your enquiries for a month. Tag them. If eight of ten are the same four questions, you have a real case and you know exactly what to build. If they are all different, you have your answer too.

If you get through those three and still want one, you will build a much better one, because you will know what it needs to say.

If you do build one

• Ground it in your own documents, and make refusal the default when it cannot find support. Do not let it answer from general knowledge about your business.
• Show the handover. A visible, one-tap route to a human on every screen.
• Log every conversation and read them weekly. This is the highest-value part and nobody does it — it is a direct record of what customers do not understand about your offer.
• Name an owner for keeping it current.
• Load it lazily, so visitors who never open it do not pay for it.
• Say it is a bot. People work it out in one message anyway, and pretending costs you the trust you were trying to build.

---

We build these, and we will tell you if we think the FAQ page and a WhatsApp button would serve you better — that answer costs you nothing and it is right more often than it is wrong. Tell us what you are getting asked and we will say which one it is.