Short answer. Model drift is the gradual decline in an AI model’s accuracy after deployment, caused by changes in the data it sees or in the real-world relationships it learned. A credit-scoring model trained on 2023 data can lose predictive power within months as borrower behavior, interest rates, and fraud patterns shift. Left unmonitored, drift turns a validated model into a silent source of wrong decisions.
Model drift is the expected result of putting a static model into a moving world, not a defect in the code. A trained model is a snapshot of the patterns that held on the day its training data was collected. Those patterns keep changing. Customer behavior changes, prices change, regulation changes, and adversaries change their tactics on purpose. The model does not update itself, so the gap between what it learned and what is now true widens week by week.
The damage is quiet. A drifting model still returns confident predictions and clean dashboards. Accuracy erodes underneath a green status light, and the first visible signal is often a business one: rising false declines, more missed fraud, a churn model that stops predicting churn. Common places drift shows up in financial services and insurance:
Drift speed depends on the domain. Fraud and market models can drift in days. A document-classification model may hold for a year. The point is that no deployed model is safe from it, so the question is never whether a model will drift but how fast, and whether anyone is watching.
In banking and insurance a drifting model is a compliance exposure, not only a performance one. The European Central Bank published its revised Guide to Internal Models in July 2025 and extended model risk management expectations from regulatory capital models to all models a bank uses, including pricing, stress-testing, and machine-learning models. Paragraph 63 of that guide sets supervisory expectations for monitoring model drift directly. Supervisors now expect banks to detect degradation and act on it, not to discover it during an audit.
The EU AI Act reinforces the same duty. Article 72 requires providers of high-risk AI systems to run a documented post-market monitoring system that collects field performance data and detects drift throughout the system’s life. Most credit-scoring and insurance-underwriting models fall in the high-risk category, and the high-risk obligations become fully applicable on 2 August 2026. Ongoing monitoring, once a good-practice option, is becoming a legal requirement with named accountability attached to the provider.
Model drift is the decline in an AI model’s accuracy after deployment as the world moves away from its training data.
Data drift is a change in the inputs. The distribution of the features the model receives moves away from the distribution it was trained on, for example a lending model that suddenly sees applicants from a new region or age band. The relationship the model learned may still be valid, but it is now being asked about data it has not seen enough of.
Concept drift is a change in the rules. The relationship between inputs and the outcome you are predicting changes, so the same input now leads to a different result. A spending pattern that signaled fraud in 2023 becomes normal behavior in 2026. Data drift is a change in the questions the model gets; concept drift is a change in the correct answers. They call for different fixes, so detecting which one you have matters before you retrain.
Detection runs on three layers. The first watches the inputs with statistical tests, such as population stability index or Kolmogorov-Smirnov tests, that flag when a feature distribution has moved beyond a set threshold. This catches data drift early, before outcomes are even known. The second watches performance by comparing predictions against ground truth as real outcomes arrive, which is the only way to catch concept drift but which lags, because a loan default or a fraud confirmation can take weeks to land. The third watches outputs, tracking the model’s prediction distribution and confidence for sudden shifts.
None of these works as a one-time check. Drift detection is a feedback loop: field data comes in, tests run on a schedule, thresholds decide when a human reviews, and a documented decision either clears the model or triggers retraining. The EU AI Act frames post-market monitoring in exactly this way, and it expects the loop to be written down, not improvised.
Article 72 requires the provider of a high-risk AI system to establish a post-market monitoring system that is proportionate to the risk, to collect and analyze performance data including input from deployers, and to keep the monitoring plan as part of the Annex IV technical documentation. Where monitoring reveals a risk to health, safety, or fundamental rights, the serious-incident reporting duties of Article 73 can follow. The obligation sits with the provider that placed the system on the market, and it runs for the life of the system rather than ending at launch. For a bank buying a model from a vendor, this makes the split of monitoring duties a contract question worth settling before deployment.
Start with an inventory. You cannot monitor models you have not listed, and most institutions underestimate how many models, including spreadsheet and machine-learning models, are already in production. Then set a drift threshold and a named owner for each model, so a breach has a defined response and a person accountable for it. Automate the input and output checks rather than relying on quarterly manual reviews, connect them to an alert, and define in advance what triggers retraining versus a full revalidation. The goal is that drift is caught by a system on a Tuesday, not by a regulator during an audit.
A model is only as reliable as the monitoring around it, which is why we treat drift detection as part of shipping a model, not a later add-on. Ableneo shipped 34 production AI projects in 2025, 94% of them using large language models, and roughly four in five reached production and stayed there. That track record in regulated Central European financial services, with institutions of the profile of ČSOB, Erste, and UNIQA, comes from building the governance and monitoring that keep models accurate after launch. For the framework that sits above drift monitoring, see our explainer on model risk management for AI.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.