A bank’s loan-origination process takes 11 days from application to decision. The team installs an AI agent on the underwriting step, which now takes 8 minutes instead of 6 hours. They roll it out across the organization, celebrate the cycle-time reduction, and report the win to the board.
Six months later, end-to-end cycle time is 10.4 days.
The team is baffled. Where did the savings go?
Answer: They didn’t go anywhere — they were never large. Underwriting was 6 hours of an 11-day process. Saving 5.5 hours on a step that’s 2% of the cycle barely moves the total. The other 98% was in document collection, customer back-and-forth, secondary reviews, and queue waits, all unchanged.
This is the Reverse Hammer: applying AI to the visible step while leaving the invisible queue intact. Most boardroom AI wins look like this. The first question on any agent program isn’t can we automate this step? but what fraction of the end-to-end cycle does this step represent? If the answer is less than 10%, you’re optimizing in the wrong place.