More drafts, the same number of approvals
An AI pilot can multiply the drafts a team produces and leave the number of approvals where it was. A count you can run to see where your pilot's output is waiting.
Published · Every number is ours, measured on our own operation.
The pilot did what it promised. It writes first drafts of customer quotes or answers to support tickets in minutes, where a person used to take most of an hour. Months after launch, quotes still leave the building at the pace they always did, and the operations review shows the same turnaround it showed before the pilot started. The software is busy all day, and the work is not moving any faster.
Count two numbers over the same month
Pick one AI workflow and one calendar month. Count what the AI produced in that month, in whatever unit it produces, drafts or summaries. Then count how many reached a customer or the books in that month. Next to each item that went out, write the approver's name.
If the first number is far larger than the second, and the same one or two names fill the approval column, the pilot has not sped anything up. It has moved the wait from writing to signing off.
Where the drafts go
Before the pilot, writing and approving moved at roughly the same speed. Someone spent an hour on a quote, and the approver read it later that day. Writing was slow enough that approval was rarely the constraint. The pilot cut the writing to minutes and left the approval exactly where it was, with a sales director or a controller who has the same hours in the day and the same other work.
So drafts arrive faster than they can be read. They collect in an inbox or a shared folder marked for review, and the approver works through them at the old pace. Checking a draft properly takes about as long as it ever did. A draft written by software deserves a slower read, since the approver did not watch it being written. Measured where work reaches a customer, output per week stays flat.
Follow that through and the shortage turns out to be one of capacity. The model is good enough. What ran out is the organization's capacity to check what the model produces, and nobody planned for that capacity when the pilot was scoped.
The items that never leave the pile
Review piles grow a second way, and it is quieter. Most review lists clear an item once it is approved or sent. A request to quote a product that has since been discontinued can never be approved as written, however many times it is redrafted. A supplier invoice with no matching purchase order can never be marked as matched. Under a rule that clears only items reaching those states, such items stay in the pile for good. If the AI redrafts every open item each morning, it redrafts those too, every day, adding to the count of drafts while nothing reaches anyone.
The question to ask of any removal rule is which items can never reach the state it tests for. Those items need an exit of their own, such as going back to a named person with the reason attached.
What a recount found in our own approval list
We keep a running list of the items in our own operation that wait on a single sign-off, and we treat its length as a warning light. When we compared that list line by line against the record of what was still open, it held 39 items against 25 real ones. Of those, fourteen were orphans. They were work already finished or withdrawn that nobody had cleared, and three of them were items an earlier written summary had already reported as deleted. The number meant to show how backed up approvals were overstated it by 56%, and reading the summaries would never have revealed that.
The repair was mechanical. The line-by-line comparison became a standing step, and it prints the orphan count even when that count is zero, so a clean list is something the comparison reports rather than something anyone assumes.
Before adding a second AI workflow
If your count shows a pile in front of one approver, another drafting workflow will add to the pile. The work sits at the approval step: decide which items a person truly needs to see and which can go out under a written rule, and give the items that can never be approved somewhere to go. Adding a second approver will probably help less than it looks, because that person inherits the same full drafts to read. A pilot is ready to scale when the number reaching customers rises with the number drafted. See how we approach moving a pilot from busy drafts to finished work.