Where the human belongs in your automation

Help me decide where humans belong in this automation. You are a systems designer who has seen automations cause expensive mistakes.

The automation: <steps, in order>
Actions it takes in the world: <sends messages / changes records / moves money / books things / publishes content>
Cost of a wrong action, per type: <e.g. wrong email = embarrassment; wrong refund = money; wrong booking = angry customer>
Volume per day: <number>
Who could review, and how much time they realistically have: <role, minutes per day>

Produce:
1. Each action classified: Run alone (low cost of error, easily reversible), Review after (run, then a human spot-checks), Approve before (human confirms each one), Never automate (human does it with AI assistance). One line of reasoning per action, based on my cost-of-error answers.
2. For "Approve before": the approval message design: what the reviewer sees (the proposed action, the evidence, one-click approve/edit/reject), where it arrives, and what happens after 4 hours with no answer.
3. For "Review after": sample rate, what the reviewer checks, and the threshold that moves the action to "Approve before".
4. Graduation rules: the evidence needed (e.g. 200 runs, error rate under X%) to move an action from Approve to Review to Alone.
5. The kill switch: how anyone can pause the whole thing in under a minute, and what gets logged.
6. The reviewer's daily time budget, computed from my volume, and whether it fits what they have.