There is a single word that decides whether human oversight of an AI is real or decorative, and it is a word about timing: before, or after. A review that happens before the AI acts is a control — it can stop a mistake reaching the outside world. A review that happens after the AI has acted is a report — it can only tell you, neatly and too late, what already went wrong.
Picture the invoices an AI has drafted, sitting in a holding area. They are complete, addressed, and ready — and, crucially, they physically cannot leave. There is no send button wired up in that room. Not a polite promise that we don't send yet, but a genuine inability to send. A person opens each one, sees exactly what the client would see, and only their approval opens the door.
The difference sounds subtle and is not. A holding area that can send, paired with a note that says we usually check first, is not a safeguard — it is a good intention with a live wire behind it. On a busy morning the intention loses.
Three things make the check real. The room genuinely cannot send on its own. The approval records what the reviewer actually saw — who, when, which version — so that a tick against a document that has since changed counts for nothing. And approvals are made one at a time. Approve all forty with a single click and you have signed your name to nothing you read.
The trap comes months later, when everything has run smoothly and the reviewer starts waving batches through unread. That feels like success right up until the day it very much doesn't. Real oversight is a permanent fixture, not a phase you grow out of.
“A review before the act is a control. A review after it is a report. The only difference is one word about timing.