When Should a Chatbot Ask a Follow-Up Question?

“Show me the employees.” Does that include former staff? Are names enough, or does the user expect contact details? The bot has to make some choices. Trouble starts when it hides them and presents the result as the only possible answer.

Clarification also has a cost. A conversation that requires confirmation of every detail quickly becomes tiring.

Ask about a difference that matters

When interpretations lead to substantially different results, ask a specific question: “Only employees with a current contract?” That is easier to answer than “Could you clarify?”

For a smaller choice, use an appropriate default and state it: “Showing current employees, sorted by surname.” The user receives an answer and a way to correct the assumption.

Missing evidence is a different problem

Sometimes the question is clear but the necessary data are unavailable. Rephrasing the request will not fix that. Explain which source or check is missing and how it affects the answer.

Adding “approximately” to an unverified number does not make it useful. If the requested figure cannot be established, provide the supported part or explain why the result is unavailable.

When a person needs to take over, include the available evidence and previous attempts. The user should not have to describe the entire problem again.

Route on observable conditions

A model's statement that it is “90% confident” is insufficient by itself. More useful conditions include multiple matching entities, an unspecified time range, an unavailable source, or failed validation.

Those conditions can determine whether to continue, clarify, or stop with an explanation. Include ambiguous and unanswerable questions in the test set. They show whether the bot can handle missing information as well as clear requests.