Map approvals, sign-offs, and accountability in an AI-assisted process

Review an AI-assisted workflow and show me exactly where humans need to stay in control. The workflow: [STEPS, noting which ones AI does] The final output or decision: [e.g. "supplier selected", "customer refund approved", "report sent to the board"] People and roles involved: [WHO does what today] Stakes: [money, customers, legal, safety, reputation - and any limits, e.g. "refunds over $500 need a manager"] Existing approval rules: [company policy or delegation of authority, if known] Return: 1. A table: Step | Human judgment needed? | Verification needed (what to check) | Approval needed (who) | Escalation trigger | Accountable person. 2. The final decision owner - one named role who is accountable for the outcome, and why it cannot be the AI or the tool vendor. 3. Approval thresholds - the amounts, risk levels, or situations that require sign-off, and at what level. 4. Escalation - which situations go to whom, and how fast. 5. Gaps - steps where nobody currently checks AI output, or where accountability is unclear or shared by "everyone". 6. Any step where a human approval would be a rubber stamp, and how to make the review meaningful (what the reviewer must look at, time allowed). Keep the controls proportionate: heavy review only where errors are costly or hard to reverse. Align thresholds with my company's existing approval rules where I have shared them, and mark anything you assumed as [CONFIRM].