.svg){kind=icon}
.svg){kind=icon}
What is model drift?
Model drift (also known as concept drift or data drift) is the phenomenon in which the predictive quality of an AI model deteriorates during operation, even though the model itself has not been modified. This is usually caused by changes in the input data or the underlying relationships.
Four types of drift: Data drift (input distribution changes—e.g., new products, different customer groups), concept drift (the relationship between input and output changes—e.g., new purchasing patterns), Label Drift (target distribution changes—e.g., fraud becomes more frequent), Prediction Drift (model predictions change—a symptom, not a cause).
Typical causes: Changes over time (trends, seasonality), behavioral changes (customers behaving differently), system changes (new data collection systems, different metrics), external events (COVID-19, crises), business changes (new products, modified processes).
For small and medium-sized businesses, drift is a practical problem: Anyone who deploys a model today needs to know tomorrow that it still works. Without drift monitoring, AI operations become a black box—and trust erodes gradually.
Best Practices for Preventing Model Drift
These six principles have proven effective:
- Think ahead from the start: Drift monitoring is part of the model—not an add-on.
- Appropriate Metrics: Don’t use one metric for everything—choose metrics that fit the data type and task.
- Reasonable Thresholds: Thresholds that are too tight lead to alarm fatigue; thresholds that are too wide obscure problems.
- Set up auto-retraining: For stable processes, enable automated retraining when drift occurs.
- Ensure ground truth: Where possible, collect real-world results—objectively measure model quality.
- Collaboration with the business: Link business KPIs to drift metrics—reveal real-world impacts.
Common Mistakes in Drift Management
We often see these pitfalls:
- Focusing only on model quality, not drift: Action is taken only when predictions start to deteriorate—drift goes unnoticed.
- Thresholds set too low: Constant alerts—the team soon ignores them all.
- No action: Drift is measured, but nothing is done—monitoring becomes useless.
- Incorrect reference data: The baseline itself is already skewed—drift detection produces false signals.
- Only global metrics: Feature-specific drift goes undetected—problems in subgroups remain invisible.
Data Drift vs. Concept Drift vs. Model Decay
Three similar but distinct concepts:
- Data Drift: Input data changes — e.g., new customer groups. Most common case.
- Concept Drift: The relationship between input and output changes—e.g., new buyer behavior.
- Model Decay: An umbrella term for all forms of quality degradation during operation.
Getting Model Drift Under Control
In a free initial consultation, we’ll review your production AI models and outline a drift monitoring strategy—one that’s pragmatic and effective.
As an AI partner for small and medium-sized businesses, we build drift monitoring tailored to your needs—with meaningful metrics, clear responses, and automated retraining wherever possible.
What We Offer
- AI Consulting — Drift Strategy and Implementation.
- AI Monitoring as a Service —Drift as an integral component.
- MLOps in the Glossary — Retraining as part of operations.
- AI Audit in the Glossary — Drift is a key audit point.
