Chatbot

What a chatbot does

A chatbot answers customer enquiries automatically, in the chat window, in a messenger or in an email. The range runs from simple rules with stored replies to systems that use a language model and phrase answers from a knowledge base. Both kinds have their place, but they differ strongly in effort and reliability.

Which tasks pay off

• Status enquiries: where is my order, when will it arrive. These questions come often and can be answered from the system.
• Standard flows: address change, return registration, invoice reprint. Clear rules, clear answer.
• Pre qualification: recognise the matter, capture data and route the case to the right place.
• Advice: this is where it gets difficult. Recommendations need product knowledge and responsibility that a rule set cannot carry.

Where it fails

Three causes explain most failures. The first is a knowledge base that is not maintained, so the bot hands out outdated answers. The second is a missing way out: if the customer cannot leave the loop, frustration appears instead of relief. The third is an expectation the bot cannot meet, such as negotiations about goodwill or complaints that require a decision.

Perception adds to it. A bot pretending to be a person is felt as deception the moment it becomes obvious. Transparency is also mandatory: users have to be able to tell that they are writing to a system.

How the handover succeeds

The handover point decides the benefit. It needs three properties. It must be reachable at any time, so not only after three failed attempts. It must carry the conversation across, so the customer does not have to repeat anything. And it must lead to a person who has time, not to another queue. These three points matter more than the quality of the automatic answers, because the case never disappears entirely.

What comes with it

A bot processes personal data. It therefore needs a legal basis, a data processing agreement with the provider and a statement in the privacy policy. On top of that comes the training question: whether input is used for learning has to be settled before the rollout, not after. Cost does not rise with the number of answers but with the amount of context each answer carries along.

Frequent questions

• Rules first or a language model? Start with rules. They are predictable, and the work on the knowledge base appears in both cases.
• Does a bot replace staff? It shifts work. The frequent cases fall away, the difficult ones remain.
• What is the key figure? Not the number of conversations but the share resolved without a handover.