Start with a one-week pilot on a single workflow and log total model spend, retry count, reviewer minutes, and whether the change actually passed acceptance checks. Put the spend and retry limits in the execution layer, not just in the prompt, and require a human decision before either limit is raised. For finance reporting, separate cost per attempted task from cost per accepted outcome; cheap generations can still create an expensive review queue. Compare those numbers with the manual baseline before expanding the pilot, and keep a rollback owner named in the execution plan.