Model drift (degradation of an AI model over time)

The term comes from the data mining literature of the 1990s. It gained enterprise weight only in 2023 to 2024, when companies started keeping models in production long enough to notice the degradation. Earlier, projects ended before the problem could appear.

Two types of drift

The Microsoft Research Concept Drift Survey of 2023 distinguishes two categories. Data drift: the distribution of the input data changes over time. Example: a product recommendation model trained on pre-pandemic behaviour meets post-pandemic behaviour, and the query distribution is different. Concept drift: the relationship between the data and the label changes. Example: a spam classification model where spammers evolve their techniques, so the same email features mean something different in 2026 than they did in 2024.

The scale of the problem in production

Google Cloud AI Best Practices of 2024 reports that 73 percent of ML models in enterprise production have no drift detection mechanism at all. Among models that do have monitoring, the average quality degradation is 8 to 12 percent per year. For models without monitoring, estimated degradation is 15 to 25 percent per year, but no one measures it, so the problem becomes noticeable only at a catastrophic drop.

Four mitigation techniques

The AWS SageMaker Model Monitor docs of 2025 standardise four practices. First, baseline monitoring: a snapshot of the data distribution at deployment, compared against current production. Second, performance metric tracking: accuracy, precision and recall measured on an external validation set. Third, automated retraining: a pipeline that automatically retrains the model on fresh data when a drift threshold is crossed. Fourth, A/B testing of model versions in production.

Cost implication

Drift monitoring costs 5 to 15 percent of the original model's deployment cost per year. Retraining in response to drift costs 20 to 40 percent of the original training cost, on average once every year to year and a half for classification models. The total cost of maintaining ML in production is typically 25 to 50 percent of the deployment cost per year, and drift management is a significant part of that.

The Polish context

In Polish enterprise deployments in 2025 to 2026, model drift monitoring is standard in fewer than 18 percent of cases (our own estimate). Most AI projects in Polish companies are still before the point at which drift becomes operationally visible (usually 6 to 12 months in production), so the problem is only now approaching.

A model monitoring plan is a standard part of AI implementation in your processes.