Watch for the Model Drifting Away From What You Validated
Is the system still as good as it was the day we approved it?
Enhanced: 22 controls
Who it applies to
Enter for every AI system whose letter is E or A (this area has no Baseline controls). Go as deep as your system letter, and implement in your mode.
Why it exists
Monitor for the quiet decay that afflicts every deployed model: the slow slide of accuracy as the world the model was built for drifts away from the world it now operates in. A fraud model trained on last year’s fraud, a demand model trained on pre-pandemic demand, a language model trained on an audience whose behaviour has moved — each degrades not with a crash but with a gradual, invisible erosion that no one notices until the cost has accumulated. Measure performance continuously against reality, so the erosion becomes a signal you act on rather than a surprise you absorb.
What to actually do
Track the model’s accuracy against real outcomes over time. Compare what the model predicted or decided against what actually happened, continuously, and watch the trend. A single good validation at launch tells you the model worked once; only ongoing comparison tells you it still works.
Set thresholds that trigger action, not just alarm. Decide in advance how much degradation is tolerable and what happens when it is crossed — retraining, review, or disengagement. A drift alert that triggers no defined response is decoration; connect the threshold to the retraining decision and, at the extreme, to the kill switch and fallback.
Controls
Enhanced2 controls
Control 201
Monitor model performance against real outcomes on a continuous basis.
EnhancedDetectiveall rolesper system
What it is
Trend-tracked, not validated once at launch.
How to implement it · Steps
Choose the metrics validated at approval, e.g. accuracy, error rate, false positives and negatives
Collect real outcomes and match each to the prediction or decision, allowing for known delays
Compute the metrics at a fixed cadence (weekly or monthly) and per relevant segment
Plot the trend against the validation baseline
Review the trend with the system owner at each cadence
How to start
Pick the one or two metrics used at approval and set up a monthly comparison against actual outcomes on a sample of cases.
Evidence
The performance records and trend charts over time.
Mode: light / standard / full
light
A monthly sample of cases checked by hand against outcomes, logged in a spreadsheet with a simple chart.
standard
Automated weekly metric calculation reviewed by the system owner.
full
Continuous monitoring by a model operations team with dashboards and per-segment analysis.
Control 202
Define degradation thresholds and the action each triggers.
EnhancedCorrectiveall rolesper system
What it is
Retraining, review, or disengagement — connected to a defined response.
How to implement it · Minimum rules
Reference: each metric's value at validation
Warning threshold (e.g. 5% relative drop): system owner reviews within a set time
Action threshold: retrain, recalibrate or tighten human review
Disengagement threshold: halt the system and switch to the fallback
Each threshold names an owner and a response time
How to start
Agree with the system owner, for each monitored metric, the drop that triggers review and the one that triggers disengagement, and write both down.
Evidence
The defined thresholds and records of the actions they triggered.
Mode: light / standard / full
light
Two thresholds per metric in the monitoring spreadsheet, with the owner-manager deciding the response.
standard
Three-level thresholds owned by the system owner, with alerts sent automatically.
full
Thresholds wired to automated alerts, retraining requests and the halt mechanism, with response tracked.
Evidence
The performance-monitoring records and trends; the defined thresholds and triggered actions; evidence of responses taken on drift.