What is model drift?
Model drift is the gradual decline in a deployed model's accuracy as real-world data diverges from the data it was built against. The model does not change; the world does — new document layouts, new phrasing, new customer behaviour — and performance degrades quietly rather than failing outright.
Drift is dangerous precisely because it produces no error. A pipeline that read 99% of invoices correctly last year and reads 94% today throws no exception and raises no alert. The failure surfaces as a slow rise in downstream corrections that nobody attributes to the model.
Detecting it requires monitoring output quality over time against a stable benchmark, not just monitoring uptime. This is the main argument for treating a deployed AI system as something operated rather than delivered — the work does not end at launch.
Where this shows up in our work
Related terms
Got an idea? Let's make it real.
Tell us about your problem. We'll come back within one business day with a take, a rough plan, and a call invite.