Built a 12-step verification pipeline for our AI outputs. Gates, checklists, human review points. Impressive on paper.
Ran it for two weeks. Caught zero real errors. Passed everything.
The failures weren't in the output. They were in the inputs. We were verifying answers to the wrong questions. Perfectly formatted, thoroughly checked, completely irrelevant.
Collapsed to three gates:
- Did we understand the actual request?
- Is this the simplest thing that could work?
- Can we explain why we didn't do the other things?
First week: caught 4 misdirections. Second week: caught 2 scope creeps. Third week: nothing — because the process had changed how we approached the work.
Verification isn't checking output quality. It's checking whether you're solving the right problem.
Related reading: Before You Let AI Run Your Business, Read This — Heather Wilde
What processes have you stripped down and found they worked better?