When an AI chatbot actually earns its keep

Most chatbots added to a small business website are gone within a month. Not because the idea is wrong, but because the thing installed was a search box wearing a friendly face, and it answered questions nobody was asking. The difference between a bot that earns its keep and one that costs you customers is almost entirely about what it can see and when it gets out of the way.

Why do most chatbots get switched off?

The standard failure looks like this. A widget is added, pointed at a page of frequently asked questions written eighteen months ago, and launched. It answers the questions in that document confidently and everything else with a variation of 'I'm not sure, please contact us'.

Customers learn within two messages that it cannot help, and start skipping it. Meanwhile it is intercepting people who would otherwise have used the contact form, so the enquiries it fails to answer do not arrive anywhere else either. It is not neutral — it is negative, and the damage is invisible because nobody reports a conversation they abandoned.

A badly configured assistant is genuinely worse than none. That is the honest starting position, and any assessment that skips it is selling something.

Why does stale knowledge ruin a chatbot?

Almost every chatbot disappointment traces back to one thing: it is answering from a snapshot. A document, written once, that drifts out of date the moment a price changes or a product sells out.

The questions people actually ask are current-state questions. Is this in stock. Does it come in blue. How much is delivery to my postcode. When will my order arrive. None of those can be answered from a static page, and all of them can be answered from the shop's own live data.

So the real dividing line is not which model a tool uses. It is whether the assistant reads your live catalogue and order status, or a copy of your FAQ page.

When should a chatbot hand over to a person?

The second thing that separates a useful assistant is that it recognises the end of its usefulness and hands over cleanly, with the conversation attached.

Two situations always need a person: anything about money that has already been taken, and anything where the customer is annoyed. A bot that keeps trying to resolve either one turns a recoverable problem into a review. A bot that says a person will pick this up, and makes sure one actually does, protects the relationship.

Buying intent deserves the same treatment. Somebody asking whether you can deliver forty units by Friday is not a support query. That should reach a human immediately, with everything they already said.

Should a chatbot admit when it does not know?

An assistant that invents an answer is worse than one that admits it does not know, because a confident wrong answer about delivery times or returns becomes a commitment a customer will reasonably hold you to.

This is a configuration decision, not a technical inevitability. An assistant restricted to what it can actually see, and instructed to say plainly when something is outside that, will say 'I don't have that' more often — and every one of those is a small win rather than a liability.

Some subjects should be refused outright rather than answered carefully. Anything touching medical, legal or financial advice belongs with a person, regardless of how reasonable the question sounds.

What is an AI assistant genuinely good at?

The value is concentrated in hours you are not working. A question at eleven at night either gets answered then or waits until morning, by which point the person has often bought elsewhere. That is the gap an assistant closes, and it is the main reason one pays for itself.

It is also good at the repetitive middle of the funnel — the same six questions about sizing, delivery, materials and returns, asked hundreds of times. Removing those from your day is a real saving even when none of them would have been lost sales.

What it is not good at is persuasion. An assistant will not talk a hesitant buyer into a purchase, and configuring it to try makes it worse at the things it is good at.

How do you judge a chatbot after a month?

Read the transcripts. Not the summary dashboard — the actual conversations, twenty of them, chosen at random. You will know within ten minutes whether the thing is helping.

Look for three specific patterns: questions it answered that would otherwise have waited until morning, questions it got wrong, and conversations where the person gave up. The third is the one that matters most and the one no default report shows you.

If you cannot read the transcripts, that is the answer. An assistant you cannot audit is one you cannot be responsible for, and you are the one whose name is on the answers.

What does it cost to do properly?

The work is not installing a widget. It is connecting the assistant to live data, deciding what it must refuse, and setting the handover rules — which is why the cheap end of this market disappoints so consistently.

The OpenX24 AI Assistant is £249 to set up and runs on a Run plan from £29 a month. It is licensed software rather than something sold outright, which is why it carries a plan: the plan is what runs it and what the reply allowance sits in.

Whether you buy it here or elsewhere, the questions to ask are the same three: does it read live data, what does it refuse, and can I read every conversation it has had.

What to do next

Judge any assistant on three things before you install it: whether it reads your live catalogue rather than a static FAQ, whether it hands over to a person on money and on anger, and whether you can read the full transcripts afterwards. If the answer to any of those is no, it will be switched off within a month — and in the meantime it will be quietly intercepting enquiries that used to reach you.

Common questions