Give an AI agent its own business surface, then keep every risky step approval-gated until the data says it is safe.