rethink your priors
Should a human approve every step an AI agent takes?
The short answer
No. "Human in the loop" usually turns into "human in the way". If a person has to approve every step, they stop reading after the first few dozen approvals and start clicking yes. You get the delay of oversight without the safety. Put people where it matters: at the steps you can't undo. Sending money, emailing customers, deleting data, signing something, publishing something. Let the reversible steps run, log what happened, and review a sample afterward. Then make the irreversible steps easy to check, with the key facts on one screen, so the approval takes seconds and actually gets read. Oversight works when it's rare enough to be taken seriously. A rule of thumb: if undoing it would take an apology, a person approves it. Decide in advance which steps those are.
Outcome first
- Put a human at the steps
- you can't undo. Only those.
Want this applied to your actual workflow? That part isn't a cartoon. Work with Eduardo. Longer writing: The Compounding Founder.