Model drift
Model drift is when a model's accuracy quietly fades because the world moved on while the model stayed frozen. The model did not change. Its inputs did: new slang, new products, new fraud tricks, a new pricing page. The trap is that nothing errors out. The model keeps answering with full confidence while getting quietly more wrong.
Why it matters
A model is only as fresh as the data it learned from, and the world keeps editing itself. Skip drift monitoring and you ship a system that was 92% accurate at launch and 74% by spring, with no alarm ever firing. Each wrong answer looks normal, so the failures stay invisible. You usually find out from a spike in complaints or a bad quarter, long after the decay started.
How it works
You catch drift by tracking model quality over time instead of only at launch. Watch two things. Are people asking about stuff the model never saw? Is accuracy on a fresh sample of real cases dropping? In practice you re-grade a labeled slice each week, compare the score to last week, and alert when it falls past a threshold. When it drifts, you retrain or refresh the data. Drift is a monitoring problem before it is a model problem.
A support bot answers refund questions well all year. Then the company changes its refund window from 14 days to 30, and the bot keeps confidently quoting 14 because that is what its training data said. Every answer reads perfectly and every answer is now wrong, until someone notices the refund complaints climbing.