Criteria for deciding if an AI agent should run in production, human accountability limits, and return metrics for businesses.