2.2.3. Model Drift and Ongoing Monitoring

💡 First Principle: A model is a snapshot of the world at training time, but the world keeps moving. As customer behavior, products, prices, regulations, and language change, the gap between the snapshot and reality grows, and performance decays without any error message. Monitoring is how you see the gap; retraining or updating is how you close it.

Two kinds of drift explain most decay:

• Data drift: the inputs change. A fraud model trained mostly on card-present purchases now sees mostly mobile-wallet payments. A support classifier meets product names that did not exist at training time.
• Concept drift: the relationship between inputs and the correct answer changes. Fraudsters adopt new tactics, so patterns that once signaled safety now signal risk. After a pricing change, the factors that predict churn are different.

Generative AI systems drift too, in their own ways: the documents a RAG assistant relies on become outdated, the provider updates or retires the underlying model version and behavior shifts, or users start asking questions the system was never designed for.

What should a business leader require?

1. Business-level metrics, not just technical ones. Track outcomes the business cares about (approval accuracy, resolution rate, complaint volume) alongside model accuracy.
2. Thresholds and owners. Define the level of decline that triggers review, and name who is accountable for acting.
3. A retraining or update plan with budget. Decide in advance how often the model is refreshed, or what event triggers it, and fund it as an operating cost.
4. Feedback loops. Capture corrections from human reviewers and customers so they can inform the next version.
5. Re-validation after change. Any retrained model or new model version is re-evaluated against baseline and responsible AI criteria before it replaces the old one.

Monitoring also catches risks covered in Phase 4: bias that emerges as populations shift, rising hallucination rates, and inappropriate use. Treat it as part of the cost of owning AI, the way maintenance is part of the cost of owning a building.

⚠️ Exam Trap: "The model was validated before launch, so no further review is needed" is never the right governance stance. When a scenario describes gradual, unexplained decline in AI performance after a business or market change, the answer involves monitoring for drift and retraining or updating the model, not replacing the whole solution or blaming users.

Reflection Question: A churn model's accuracy fell from 88% to 71% over nine months after the company launched a new subscription tier. Is this data drift, concept drift, or both, and what should have been in place to catch it sooner?