Drift has two common causes in business AI. The inputs change, for example a supplier redesigns its invoice layout or customers start asking about a new product. Or the model changes, because a provider updates or retires the version the workflow was tested on.
Neither shows up as an error message: the system keeps answering, just less accurately. That is why live AI workflows need monitoring of quality signals, such as how often people override the AI, plus a scheduled rerun of the original test cases. Treat drift as expected maintenance rather than a failure, and agree in advance who reviews the signals and who approves a fix.