What I built, what it cost, and what broke the first time. Written from inside a working business rather than about one.
My AI system gave me a confident recommendation, built entirely on someone else's video. Asking it to argue against itself is what caught it, and it's the same build-measure-learn loop that decides everything else I keep.
An AI built to catch compliance mistakes gets nothing, not worse, on a genuinely new case. What changed once that gap got a permanent, searchable record instead of a retrain.
Killing a bad AI suggestion was the easy version of judgment. The harder version showed up weeks later, when the same assistant started asking permission for things that were never its call to escalate.
Same software, same access, same instructions. One drafted an email faster, the other rebuilt half a process. Compute is electricity, and almost nobody was handed the appliance. Includes the rule that cost me the most to learn.
Four sweeps of the AI voices in this region. Government, vendors, executives, all crowded. The practitioner lane, the person who runs the experiment and reports what broke, is empty. Including two of my own failures from this month.