The starting point
The measured approval time was ten days. Requests had to make it through operating and capital expense checks, budget warnings, approver lookup, messages, and approvals.
What I changed
I rebuilt the process with AI checkpoints that caught expense issues earlier and surfaced budget warnings before they reached the end of the queue. The build used n8n and Claude to connect the checks, lookup, messages, and approvals into one repeatable run.
Approval time
10 days → 2 days
The result came from changing when the routine checks happened, not asking approvers to work faster.
What carries over
The system itself belongs to Meta and is not a product for sale. The useful lesson is the shape of the work: find the repeated checks that hold a request up, handle those checks earlier, and keep the people who own the decision in control.
That is the same way I approach a smaller inbox, report, or approval queue. I learn how it runs now, agree what should change, build it around the software already there, and come back to check it is still working.