Three things people mean by “chatbot”
Establishing which one you want takes two minutes and saves a wasted project, because the three are not versions of each other. They are different products with different costs and different failure modes.
| Kind | Roughly what it costs | What it can answer |
|---|---|---|
| Scripted / button menu | Free to a few pounds a month, usually inside a helpdesk you already pay for | Only what you scripted. Reliable, limited, and honest about it |
| AI trained on your website | A monthly subscription, typically low tens of pounds | Anything on your public pages. Cannot see orders, stock or accounts |
| Custom, wired into your systems | A build, priced from hours; running cost is usually tens of pounds a month in model usage | “Where is my order”, “is this in stock”, “move my appointment” — the questions people actually ask |
Source: Cybrial’s own categorisation, written from builds. Subscription figures are the shape of published self-serve pricing across the category rather than any single vendor’s, and no competitor agency has been used as a source.
The gap between rows two and three is where most disappointment lives. A bot trained on your website will answer “what are your delivery charges” beautifully and then fail, in front of a customer, on “where is my order” — which is the question that generates most of the contacts you were trying to deflect.
The running cost, which surprises people downwards
Model usage for a small business assistant is usually tens of pounds a month, not hundreds. The expensive part is the build — and most of the build is integration rather than anything to do with AI.
That is worth knowing because it changes what a large monthly platform fee is actually buying. If a provider charges several hundred a month for a wrapper around a model that costs a few pounds to run, the fee is for the wrapper. Sometimes that is fair — someone maintains it. Often it is not.
The position taken here, and it is a pricing position rather than a technical one: the build is quoted from hours, and it then runs on your own account with the model provider at their published price. You can see the bill. Nothing is resold at a markup.
When a chatbot makes a website worse
This is the most-requested AI build and the one most likely to damage the thing it was added to. Three specific ways, all common.
- It intercepts people who were about to buy. A bubble that opens itself over the checkout is a conversion problem wearing a helpfulness costume.
- It hides the phone number. If the fastest route to a person is now three clicks inside a widget, the business has made itself harder to contact and called it customer service.
- It answers confidently and wrongly. A bot that invents a delivery date is worse than no bot, because the customer believes it — and then holds you to it.
The version that works is narrow and slightly boring: it answers the twenty questions you genuinely get every week, it says it does not know when it does not, and it hands over to a person quickly and visibly. That is a smaller product than most people set out to buy, and a considerably more useful one.
The bits with rules attached
Chatbots are not illegal in the UK, and the question comes up often enough to be worth answering plainly. What applies to them is the law that already applies to everything else you do with customers.
- Data protection. Conversations frequently contain personal data, so UK GDPR applies: tell people what is collected, keep it no longer than needed, and be able to delete it on request.
- Consumer law. Anything the bot says about price, delivery or refunds is a statement by your business. “The bot said it” is not a defence.
- Accessibility. A widget that cannot be operated by keyboard or read by a screen reader excludes customers, and the public sector has stricter obligations again.
- Transparency. Nothing requires a specific disclosure wording for a chat widget, but making it obvious it is automated is both good practice and the way to avoid the complaint.
None of that is a reason not to have one. It is a reason to know what it is allowed to say, and to write that down before it goes live rather than after the first refund dispute.
How to tell whether it is worth it at all
Count the contacts. Take a fortnight of emails, calls and messages and sort them into the questions that repeat. If a handful of questions account for most of the volume, an assistant that answers those well is worth building — and you now also know exactly what to build.
If instead the contacts are all different, a chatbot will deflect very little and the money is better spent on the pages that would have answered them. Quite often the honest finding from that fortnight is that the website does not state the prices, the lead time or the service area, and the phone is doing the work a page should do.
That is an unglamorous conclusion and it is the right one more often than not — which is why this page says it out loud rather than quoting a build price.
