A chatbot people do not immediately close

Most people already have an opinion about chatbots, and it is bad. It was earned honestly: for years companies installed a window that was good at asking questions and bad at answering them.

Meanwhile a bot that answers a question at midnight sells better than a salesperson who calls back on Monday. The difference between those two bots is not the model and not the budget — it is five decisions made before development starts.

Rule one: the bot answers before it asks

The classic bad exchange: a customer writes “how much is delivery to Manchester” and the bot replies “Hello! May I have your name?”

That looks like courtesy. What it means is that the bot collects leads rather than helps. The person came with a question and wants an answer. They will leave their details themselves if the answer suits them.

Do this instead: answer the question asked, then clarify. “Delivery to Manchester is £6, two days. Shall I set it up?” Contact details are requested at the step where they are actually needed.

Rule two: the bot knows what it does not know

An unconstrained model always answers. Asked about the lead time for a product outside its data, it will invent a plausible one. The customer believes it, because it is written confidently. You handle the aftermath.

This is not a technology problem, it is a configuration decision. The bot must answer only from your data: price list, stock, policies. For everything else there is one correct response: “I would rather not guess — let me check with a colleague and come back to you.”

How to test at handover: ask the bot about something you do not have. A product you do not sell, a branch in a city where you have none. A confident, detailed answer means the bot is not ready for customers.

Rule three: the exit to a human is always visible

Someone rephrasing their question for the third time is already annoyed. If they cannot find a way to reach a person at that moment, they leave — and not towards you.

The “talk to a human” control should be visible from the first message, not appear after three failures. This feels backwards: was the point not to take load off people? In practice a visible escape hatch reduces handovers rather than increasing them. The customer stays calm because they know the exit exists, and does not panic at the first misunderstanding.

Non-negotiable companion rule: on handover, the bot passes the whole conversation across. “Could you explain from the beginning?” after ten minutes of typing is exactly how bots earned their reputation.

Rule four: the bot remembers the line above

Customer: “Do you have the black one in 42?” Bot: “Yes, we do.” Customer: “What about 43?” Bot: “Could you specify which product you mean?”

The conversation ends here. Context is not a luxury feature, it is the minimum requirement for anything calling itself a conversational interface.

Technically this problem was solved years ago. If a supplier’s demo loses the thread on the second question, that is not a limit of the technology, it is a saving on the build.

Rule five: the bot has an owner

A bot ships and then lives its own life. A month later your prices changed, two months later a new service appeared, three months later the promotion ended and the bot is still advertising it.

So a bot needs two things alongside it: a place where you edit the information yourself without a developer, and a conversation log that somebody reads. Not all of it — reviewing the ones where the bot said “let me check” once a week is enough. That is a ready-made list of what it does not know and the best source of improvements.

What this means for the brief

If you are about to commission a bot, put five lines into the brief:

1. Answers the question before asking for anything in return
2. Answers only from our data, honestly says “let me check” about the rest
3. The human handover control is always visible and carries the history with it
4. Holds context within a conversation
5. We edit the copy and data ourselves; conversations are stored and readable

None of these adds meaningfully to the cost. Together they decide whether your bot produces enquiries or collects resentment.

What not to expect

A bot will not replace a salesperson in a complex sale. It will not negotiate on price, will not sense hesitation, and will not read from tone that it is time to back off.

Its job is to absorb the stream of identical questions that eats your team’s day, and to stop enquiries from dying overnight. That is enough to pay back in a couple of months, and expecting more is the fastest route to disappointment.