When past patterns stop working
A model trained to spot fraud in past payments may struggle when people use new payment methods or scammers change tactics. If the kinds of payments change, that is data drift. If the clues that signal fraud change, that is concept drift.
An AI assistant’s results can also shift when users’ questions, source documents, instructions, or the model it uses change.
Check what it gets wrong
Teams compare the model’s results with real outcomes over time, including for different groups of users. A change in the data does not always mean worse results, and a steady overall score can hide problems for one group. If performance drops, teams may update the data, retrain the model, or change how the app uses it.