The rule. Use dialog to resolve uncertainty the system cannot resolve on its own, and only when the cost of guessing wrong outweighs the cost of interrupting.
Why. The reference point is Horvitz's mixed-initiative principles: a system uncertain about a user's intentions "should be able to engage in an efficient dialog with the user, considering the costs of potentially bothering a user needlessly" (Horvitz, 1999, p. 160). The same paper turns the instinct into arithmetic: weighing the expected utility of action, inaction, and dialog yields two probability thresholds, so the system stays quiet below one, asks in the band between, and simply acts above the other (Horvitz, 1999, pp. 164-165). Microsoft's HAX pattern G10-A carries the rule into practice: elicit clarification before acting, define the uncertainty threshold that triggers asking, and skip the interruption when confidence is sufficiently high (Microsoft, n.d.).
Seen in the wild. ChatGPT's deep research inserts a dedicated clarification step, asking about goals, preferences, and constraints before committing to a long research run (OpenAI, n.d.).