Conversational interfaces are good at making uncertainty feel smooth. That is useful when the job is understanding a messy request. It is dangerous when the same confidence is allowed to calculate, authorise, and write the result.
Interpretation is not calculation
Natural language is an excellent input layer because people do not think in database schemas. We remember half a recipe, mix languages, omit units, and describe time as “last week.” The interface should meet that reality.
But the system should cross a visible boundary after interpretation. Structured values can be checked against known data, calculations can be repeated, and permissions can be applied by services designed to own those decisions.
Confirmation is a feature
A confirmation step is sometimes dismissed as friction. In a private or consequential system, it is also the moment a user can see what the software believes they asked for.
The best confirmation does not ask a vague “Are you sure?” It exposes the interpreted record: quantities, dates, assumptions, and the action about to happen.
Useful uncertainty
A trustworthy system does not need to pretend every input is complete. It can calculate what is known, mark what is estimated, and ask one focused question when the missing information would materially change the answer.
That is not a failure of intelligence. It is intelligence with an honest boundary.
Let the model make language easier. Let deterministic services make the consequential claim.