Why most business chatbots get switched off within a few months

The failure is almost never the model

Worth establishing first, because it changes where you look. The models underneath these products are extremely capable of holding a helpful conversation. Very little of what goes wrong is about the quality of the writing.

What goes wrong is that the thing was deployed without the material it needed, without a route to a person, and without anyone owning it afterwards. Those are all decisions made before launch, which is the good news: they are fixable in advance and very expensive to fix later.

1. It has nothing to answer from

The most common failure by a wide margin. The bot is pointed at a website that never answered the questions customers actually ask, so it does what a system with no source does: it generalises, or it invents.

Invention is the version that gets it fired. A bot that confidently tells somebody you are open Sunday, or that you cover a town you do not, produces a complaint that lands on the owner's desk with a screenshot attached.

• What it looks like: answers that are technically about your business and wrong in the details. Hours, coverage, turnaround, guarantees.
• Why it happens: the site answers what you do and never answers what happens if, how long, or do you cover me.
• The fix, and the order: write the answers first, deploy second. Not the other way around, and it is worth delaying a launch by a month over.

2. There is no way out to a person

A bot that cannot hand over traps people. Everyone has met one, everyone remembers it, and that memory is what the next visitor brings to your widget.

The trap is rarely deliberate. It comes from the handover being technically possible but practically useless: it opens a contact form, or it emails an inbox nobody watches, or it works only during hours when a person was available anyway.

• What it looks like: visitors typing "agent", "human", or something less polite, three messages in.
• Why it happens: nobody defined what happens at 9pm on a Saturday, so the answer became nothing.
• The fix: a real route out, on every single turn, that works at every hour. Out of hours it takes a message and promises a time. That is enough, if the promise is kept.

3. Nobody agreed what it must never say

This one does not surface gradually. It surfaces once, badly.

A bot quotes a price it should not have. It gives something that reads as clinical advice. It confirms someone is a client. It accepts a complaint in a way that sounds like admitting fault. In a regulated business, any of those is a serious event rather than an embarrassing one.

4. It was never given the second half of the job

A bot that only answers questions is worth much less than the effort of deploying it, and this is the failure that looks like success for the first two months.

Traffic goes through it. Questions get answered. And nothing happens as a result, because it cannot book, cannot capture an enquiry into anything that gets worked, and cannot hand a qualified person to a salesperson with context attached.

• What it looks like: healthy usage numbers and no attributable revenue. The dashboard looks fine, so nobody investigates until someone asks what it is for.
• Why it happens: it was bought as a support tool and judged as a sales tool, or nobody connected it to the systems where work actually happens.
• The fix: decide before launch what it is supposed to cause, not just what it is supposed to answer. Then wire it to that.

5. Nobody owns it after launch

Businesses change. Hours change, services change, a policy changes, and the bot keeps confidently saying the old thing because its source material was written in March and it is now November.

Meanwhile the conversation logs, which are the most valuable thing the whole exercise produces, are read by nobody.

• What it looks like: accuracy quietly degrading over months, with no single moment where it broke.
• Why it happens: it was a project rather than a responsibility, and the project ended at launch.
• The fix: a named owner, and someone reads the unanswered questions monthly. That review is where the next round of improvements comes from, and it takes half an hour.

6. It was the wrong tool for the actual problem

The quietest failure. Everything works, and the business did not have a website-chat problem.

Plenty of businesses put a bot on a site that gets modest traffic while the phone rings out at lunch and Instagram DMs sit unread. The bot is not failing; it is answering a channel nobody was using while the busy channels stay unanswered.

What the working ones have in common

We have built and inherited enough of these to be reasonably confident about the pattern. The ones still running after a year share five things.

1. Real source material, written before launch. The questions customers actually ask, answered properly, in your words.
2. A person one message away, at any hour. Out of hours it captures and promises a time, and the promise is kept.
3. An explicit never-say list, enforced and tested. Signed by whoever owns the risk.
4. Something it causes. A booking, a captured enquiry, a routed handoff with context. Not just an answer.
5. A named owner and a monthly look at what it could not answer.

Where to start, honestly

The prerequisite is the source material, and it is also the part that gets skipped because it is the least fun. It is worth doing whether or not you ever deploy a bot: the same answers improve your website, your search visibility and how you get described by AI assistants.

The FAQ extractor reads your site and returns the thirty questions buyers ask, flagged where your site does not answer them. Those flags are the work list, and they exist regardless of what you deploy on top.

The wider set of reasons AI output comes back generic is in twelve reasons your AI gives generic answers, and the material a system needs is in the context sheet. If you are in a regulated sector, settle the Law 25 questions before anything is connected rather than after.