The numbers feel off, but I can't prove they're wrong.
Out-of-date data or models
A model or dataset was accurate when it was built. Conditions changed, and nobody updated it. The output still looks current, but it reflects how things were when the model was built, not how they are now.
- Nobody knows when the model or data was last checked against real results.
- The business has changed in ways the model can't know about.
- The team that built it has moved on.
Who: whoever relies on the model's output. Ask: "Who owns this, and when was it last compared with what actually happened?" Look at: its last five predictions next to the actual results. You're probably looking at this pattern if nobody owns it or its misses are growing. Probably not if someone checks it on a schedule and the misses are steady.
Nobody knows where the AI can be trusted. That pattern is about which tasks the AI is good at. This pattern is about a model that was good and stopped being good.
Set aside a handful of known-answer cases and run them now and then, against the model and against the team.
For the next month, write down each time you overrule the AI and why. At the end of the month, check which of those decisions turned out to be right.
Our reasoning. No study we found measures how often models get rechecked after launch. We recommend retesting where an AI is reliable with each major new generation of models.