The scene repeats itself in every company. The system runs for three weeks, gets most of it right, and one error reaches a director. The next meeting decides to go back to the manual process. Nobody asks how often the manual process was wrong — because nobody ever measured it.
You are not comparing the AI with perfection. You are comparing it with what your company already does — and that number exists.
Measure the baseline first
Before any contract, take a hundred already-resolved cases from the process you want to automate. Review them by hand. Count how many came out wrong, how many needed a correction afterwards, and how long each correction took. That is your current number. It almost always surprises people, and almost always upwards.
From there the conversation gets concrete: the system has to be wrong less often than that, cost less than that, or both. With no baseline, any error becomes an argument to cancel and any success looks like luck.
Three numbers you have to agree on
The cheap error and the expensive one. Filing a document in the wrong category costs two minutes. Extending credit to someone who should not have it costs the whole loan. Those are different regimes and cannot share a limit.
The time it takes to find out. An error that surfaces the same day is an incident. The same error found at quarter close is a loss. Agree the maximum time to detection, not only the rate.
What happens next. An error that only becomes a complaint teaches nothing. An error that becomes a written example and enters the rules improves the system. That is a routine with an owner and a date, not a good intention.
The warning sign
If a supplier promises it never gets anything wrong, they are selling, not building. The good answer is a different one: "here is how it fails, here is how long it takes you to find out, and here is the path to fix it". You are buying the ability to know, not a promise of perfection — and that is what decides whether the project survives its first visible error.
What is your process's error rate today?
If the answer is "we never measured it", that is the first thing the diagnostic delivers.