Measure completeness by the actions an instruction permits
Suppose you ask an agent to review important invoices first. Importance could mean an approaching due date or a high cost of delay. If one invoice ranks first on both criteria, the agent can review it first without settling which meaning you intended.
An instruction is complete when the intentions consistent with the instruction and shared context imply the same set of actions in every situation. An incomplete instruction can still support a decision when all plausible intentions have an optimal action in common for that case. Completeness is a property of the instruction across situations. The need to clarify is a question about the decision at hand.
Clarification has value when it changes the choice
Now imagine one small invoice is due soon and another is due later but carries a much higher cost if it is paid late. The order depends on what the person meant by important. A focused question about which criterion takes priority can change the action and prevent a costly mismatch.
Asking takes time, too. The person must interrupt their work and the process waits for an answer. For a low-impact decision that can be reversed, the expected loss from proceeding may be smaller than the cost of asking. The same ambiguity can justify a question when the consequences are larger.
Budget human time before and during execution
Compare doing the work directly, writing a detailed instruction upfront, and delegating with questions at selected decision points. Writing a rule can pay off when the same exception will recur. For an unusual case, waiting until it occurs and asking a focused question can consume less of the person's time.
When human time becomes scarcer, consultation during execution can decline without making complete instructions a better choice. Preparing those instructions also consumes the same scarce time. If a clarification applies to later cases under the same conditions, recording the condition alongside the decision can reduce the need to ask again.