The CFO's first question is usually ambiguous
“Show me revenue for the quarter” leaves several choices open: the quarter, the legal entity and the revenue definition. The person asking may consider all three obvious. The assistant may have several plausible answers. I want the response to make that gap visible when it matters, without turning every routine request into an interview. The difficult design choice is deciding which assumptions are safe in this particular workflow.
Not every vague phrase needs a follow-up. An assistant that interrogates users before every answer makes simple work frustrating, and people stop using it. The useful test is consequence: would the missing information materially change the answer or the decision it supports? Design clarification around that, not around a blanket rule to always ask or never ask.
Have the business owner list common ambiguous requests from real work. For each, decide which assumptions are safe inside the workflow and which require confirmation. That gives the implementation team a practical specification instead of a philosophy.
Make the question easy to answer
A good clarification offers a few meaningful choices in business language. “Do you mean booked orders or recognized revenue?” is far easier than asking the user to define the metric. The choices should come from approved definitions, not terminology the assistant invented. This is where a governed semantic layer earns its keep: the options offered are the certified ones.
Explain why the distinction matters when it is not obvious, in one sentence rather than a tutorial. Then continue the task.
Do not ask for information the system can already obtain reliably and with authorization. Repeatedly requesting a known reporting period or the user's own scope adds friction and can prompt people to paste in sensitive detail they should not share. Separate missing business intent from missing technical context.
Price the cost of a wrong assumption
The lesson I return to most from Claude rollouts is workflows over features. People adopt AI when it maps to work they already do, which means the cost of a wrong assumption is real business cost: a variance memo built on the wrong revenue definition, or a member service response using the wrong eligibility date. Build test pairs where the same wording leads to different valid answers in different contexts, and check whether the assistant uses context correctly and signals remaining uncertainty.
Include a case where the user declines to clarify. The system might give a qualified answer, show the alternatives or say it cannot give a reliable answer. What it must not do is switch silently from careful questioning to an unsupported definitive result.
Track clarification frequency by question type instead of treating every follow-up as a failure. Frequent clarification can signal unclear product language, missing defaults or a genuinely complex decision. Find the cause before optimizing for fewer turns. A short conversation is worthless if it answers the wrong question.
Carry clarified context forward, carefully
Once the user chooses a definition or period, show it in the answer. If the number is pasted into a board pack or an audit committee memo, the reader should not need the whole conversation to know what it means. A concise label does the job.
Decide when context expires. A choice made for one entity or reporting task should not govern every future question. Test a change of topic or entity to confirm the boundary holds.
Ask pilot reviewers to compare two versions of a response: one with a stated assumption and one with a clarifying question. Their reactions help identify where a default saves time and where it hides an important choice. Test with users from different entities or teams. An assumption that feels obvious to the report’s author may be unfamiliar to someone receiving it elsewhere.
Clarification is decision support, not failure
A well-placed question is not an admission that the system failed. It is the assistant doing what a good analyst does before handing a number to an executive: confirming what was actually asked.
The goal is an efficient conversation that lands on a well-defined answer, with enough context that the business owner can understand it, defend it and use it responsibly. In regulated industries, where an examiner or auditor may later ask what a figure meant, that context is part of the answer itself.