The day my AI assistant looked like it was working and wasn’t

There is a specific kind of gear failure where nothing looks wrong. The gear spins. It sounds right. It just is not touching anything: no load, no work, no result. The teeth never mesh.

That was the follow-up assistant last week. The job was simple: after each session note gets filed, draft the follow-up email and queue it for review. The dashboard said it ran. The logs said success. The week felt lighter.

Then came the check. Zero drafts in the review folder. Seven days of “success” and not one email existed anywhere. The automation had been writing into a folder that no longer existed, and the tool considered that a job well done.

What actually failed

Not the AI. The system around it. There was no checkpoint where the machine had to show receipts: not a report that says done, but the actual artifact, counted and visible. A report is a claim. A draft sitting in a folder is proof.

Athletes know this one. The training log can say whatever it wants. The scale and the mat do not negotiate. Every system needs its version of the scale.

The fix, in plain terms

Each automation now ends with a verification step that counts real outputs and compares that number against the number of inputs. If the two disagree, the system flags it the same day, not seven days later. Boring to build. Completely worth it.

“Trust is earned by the checkpoint, not the dashboard.”

If there is one thing worth stealing from this note: do not ask your tools how they are doing. Ask them to show you the work.