AI systems should propose decisions, but humans need to approve them for the first one to two weeks. A reviewer checks the AI's suggestion, gets sign-off from someone higher up, then locks in the decision. This catches errors before they matter.
for each one right here, they can approve or disapprove. So this is not really an approval because the decision was already made by the AI. But obviously, when first launching something like this, we don't have to go that far. We can basically just say, for the start, instead of the AI instantly deciding, it says what it would decide, or it outputs the same thing. But then we say, like, for the first week or two, we still say that person for now, just to make sure no wrong decisions get made. This person still has to ask the higher-up and then send, send whatever decision that AI made or that person themselves would have made. Um, and then basically the per— that higher-up can say, yep, that's correct, or no, that is wrong, and then lock that decision right away. So
