Model Drift

Model drift is the slow decay of an AI system's accuracy as the real world moves away from the data it was trained on. Nothing breaks and no code changes. The model keeps doing exactly what it did at launch, but the world it is predicting has shifted underneath it, so behavior that was right then is wrong now.

In Depth

A model encodes the patterns of a particular moment, the data it saw during training and evaluation. That moment does not hold still. Customer behavior changes, vocabulary changes, products and policies and regulations change, and the mix of inputs the agent sees in production drifts away from the mix it was tested against. The model's understanding goes stale in increments too small to notice day to day, until the cumulative gap is large enough to produce real errors. This is drift, and it comes in two related flavors: the input distribution shifting (data drift) and the relationship the model learned no longer holding (concept drift).

The governance problem is that drift defeats point-in-time assurance. An agent can pass a rigorous risk assessment, clear an enterprise security review, and earn a strong Agent Trust Score (ATS) by Ollive at launch, then degrade for months without tripping a single alert. There is no outage, no exception, no failing test in CI. The artifacts that said "this agent is safe" describe a version of the world that has since moved on. NIST's AI Risk Management Framework (AI RMF 1.0) treats this as a lifecycle issue for exactly this reason: managing AI risk is not a launch gate, it is an ongoing measurement obligation.

Drift also quietly reopens failure modes a vendor thought they had closed. An agent that tested clean for bias at launch can develop disparate outcomes as its population shifts. One that rarely hallucinated on the original input mix can start fabricating as it meets inputs it was never evaluated on. The original test results expire, even though no one changed the model.

What It Looks Like

A vendor ships an AI agent that scores insurance claims for fast-tracking. At launch it is accurate and well-calibrated, validated against a year of historical claims, and it sails through the customer's review. Over the following year the claim mix changes, with new product lines, new fraud patterns, and a shifted customer base. The model, untouched, keeps applying the relationships it learned from the old distribution, and its accuracy erodes month by month. By the time the customer notices a spike in wrongly fast-tracked or wrongly denied claims, the agent has been quietly making bad decisions at scale for two quarters, and the launch-day validation everyone relied on is long stale.

The pattern generalizes. A triage agent grows less reliable as symptoms and presentation shift; a content classifier decays as language evolves. Whatever the domain, the liability accrues silently between the launch that was reviewed and the present that was not.

Why It Matters For AI Vendors

Drift turns a one-time approval into a standing exposure. The certificate, the model card, and the red-team report all describe launch day, but the agent's behavior is whatever it is today. When drift produces a harmful decision, the harmed party and your enterprise customer do not care that the agent was validated a year ago. They care what it did this week, and the indemnity clause routes that back to you.

It also undercuts an assumption baked into most insurance: that a risk assessed once stays assessed. A policy or a review that treats the agent as a fixed artifact misses the way AI risk accumulates after launch. Underwriting AI well means accounting for the version of the agent that exists now, not the one that was tested.

Common Questions

No. The code is doing what it was written to do; the world it models has changed. You manage drift with monitoring, periodic re-validation, and retraining, not with a one-time fix, because the gap reopens as the world keeps moving.
Because liability tracks the harm the agent causes now, not the date it was validated. A stale model making a harmful decision today is a present exposure, and your customer's indemnity clause does not expire just because your last review did not.
It protects the version that was reviewed. Drift is the risk that the reviewed version and the running version stop being the same thing, which is why NIST's AI RMF frames risk management as continuous across the lifecycle.
← PreviousModel Card Next →NIST AI Risk Management Framework (AI RMF)

See where your AI agents stand.

Get an Agent Trust Score, map your liability exposure, and find out what it takes to make your AI agents insurable.