When to Transition AI Workflows in Sales from Human Review to Full Automation

I keep seeing teams build a useful AI workflow, then get stuck on when to let it run without someone checking every step. My current bias is to keep human review until the team can explain what the workflow gets wrong, not just what it gets right. For anyone running AI in prospecting, RevOps, or sales enablement: what evidence made you comfortable moving a workflow from suggested actions to real automation? Was it a clean run count, a measurable error rate, or something else?