Each step works on what the step before it produced, so the chances multiply. Five steps that are each 80% accurate come out right together 33% of the time, which is 0.8 multiplied by itself five times.
An email draft is a single step, and a person reads it before it goes out, so a weak draft costs a minute of editing. A month-end close is a chain. Picture an AI that pulls the bank lines, matches them to invoices, posts the accruals, reconciles the accounts between group companies and writes the commentary. A wrong match in the second step does not stop anything. It looks reasonable, the next step builds on it, and by the commentary it has become a variance that someone must explain to the board.
So the question to ask a vendor is how many steps its AI runs before a person looks. A controller who spends 15 minutes on the unmatched and doubtful items after each step catches an error while it is still one wrong line in a list. Five such reviews take just over an hour, and the AI still does the days of matching and posting.
A rise from 80% to 98% per step saves a few minutes of editing on a draft, but it lifts a five-step job from 33% to 90%. So multi-step finance work will move to AI in jumps, one model release at a time, and each jump takes a review out of the close.