Model drift is the slow, invisible shift in data patterns that makes an AI system less accurate over time, and it works like rust on a steel beam—​the structure stays standing while its load‑bearing capacity silently declines. The drift originates from two sources: external market changes that alter input distributions, and internal feedback loops where the model’s own outputs shape future inputs. When a recommendation engine starts nudging customers toward a narrower product set, purchase histories become homogenized, feeding the model data that reinforce its existing bias and gradually eroding relevance.

In a recent internal review, a mid‑size e‑commerce platform discovered that a six‑month‑old click‑through model had slipped from a 12 % conversion lift to just 6 %, costing the business roughly $480 K in lost revenue. The team traced the loss to a subtle rise in “new‑arrival” SKUs that the model never saw in training, combined with a seasonal surge in bulk orders that altered basket composition. By the time the KPI dashboard flashed a red flag, the model had already been influencing pricing and inventory decisions for weeks, compounding the error.

The lesson is that without a systematic drift‑monitoring loop, the very automation meant to boost efficiency becomes a hidden liability.