What an AI chatbot can and cannot do with customers

Chatbots have a bad reputation, and they earned it. The old generation answered from fixed menus and trapped you in a loop until you gave up. The current generation understands the question, but that creates a new problem: it can answer convincingly and wrongly.

What it does well

Questions that repeat

Opening hours, address, order status, how to return a product, what documents an application needs. These are the majority of support messages and they need no judgement.

Searching documentation

This is where it is genuinely useful. A person would spend ten minutes searching a manual; a chatbot with access to it finds the relevant passage immediately.

Gathering information before handing over

Even when it cannot resolve something, it can collect what is needed: order number, what happened, when. The person picks up a problem already described, not one from scratch.

Being available at impossible hours

Saturday at 11pm, a correct answer to a simple question beats a contact form.

What it does not do well, however hard anyone pushes

• Situations where the person is angry. Nobody wants to be calmed down by a program. Anger is the clearest signal to hand over to a human.
• Decisions with consequences. Exceptions to the returns policy, discounts, commitments. These have to be made by somebody who answers for them.
• Cases it has never seen. Exactly where a chatbot gives its most confident and most wrong answer.
• Anything it was not told. A chatbot knows what you gave it. If the documentation is old, the answers are old.

The rule that makes the difference

Handing over to a person must be easier than persisting with the bot.

A visible, permanent button that reaches a person. Without asking for the reason three times. Most bad experiences come from here: not from the bot not knowing, but from it not letting you leave.

How to measure whether it is worth it

Not by “how many conversations it had”. By:

1. How many ended without reaching a person, where the customer did not come back with the same question the next day.
2. How fast somebody who does reach a person gets an answer — it should be faster, because the team has fewer trivial messages.
3. How many wrong answers were flagged. If nobody tracks that, you do not know what is happening.

What has to be prepared first

A chatbot does not invent information about your company. It needs the answers written down somewhere: the returns policy, the deadlines, the procedures. If they are not written, the first stage is not technology, it is writing them — and that helps anyway, including your own team.

The honest conclusion

A well-built chatbot takes over the boring, repetitive part of support and leaves people the cases that genuinely need a person. It does not replace the support team. It changes what lands on their desk.

If you want to see whether it makes sense for you, tell us which questions you receive most often. That list makes it clear within minutes.

← Back to the blog