When the bot answers on its own — and when it hands the conversation to a human

Why handover matters more than answer coverage

A customer forgives a bot for not knowing. They do not forgive a confident wrong answer, or going round in circles. So the first thing to configure in a bot is not the number of topics — it is the exit condition.

A bot that answers seven questions out of ten and honestly hands over the other three works better than a bot that tries to answer all ten.

Four cases where the bot must call a human

• No answer in the materials. The question is outside the approved base. Making things up is not allowed.
• The stakes. Money, contracts, delivery dates, returns. Anything where a mistake costs money or reputation.
• Emotion. The customer is irritated or repeating the question for the third time. From here a person leads the conversation.
• A direct request. The customer asks for a human. This is honoured immediately, with no coaxing to “try once more”.

What a correct handover looks like

What the bot says

Briefly: I can’t answer this precisely, I’m passing you to a colleague. No three-line apologies, and no promises a person will have to answer for.

What the employee receives

The whole conversation, not the last message. The customer should not have to explain their question a second time — that is the main source of irritation during handover.

What happens outside working hours

The bot honestly states the response time and takes contact details. The worst option is a bot that stays silent about nobody being available right now while the customer waits in vain.

What the team prepares

• A list of handover topics. Drawn up before launch and agreed with the people who will answer.
• A duty rota. It is clear who takes handed-over conversations during working hours and what happens outside them.
• A reaction time. The promise in the bot’s message matches what the team can actually keep.
• One place for replies. The employee sees handed-over conversations in one window, not in four messaging apps.

How to test the rule before launch

1. Take real enquiries. Thirty recent conversations are usually enough.
2. Label them. Mark where the bot should answer on its own and where it should hand over.
3. Run them through the bot. Look not at the elegance of the wording but at the match with your labels.
4. Count handover errors. Answering alone where it should have handed over — a defect. Handing over unnecessarily — acceptable.

A check like this takes a day and removes most of the first week’s surprises.

Frequently asked questions