Control · C-050 · Monitoring, Assurance & Incidents
Drift and performance monitoring
Input distributions and output quality are measured in production against thresholds that trigger investigation, retraining or withdrawal.
- Control type
- Detective
- Risks addressed
- 6
- Audience
- Heads-of · Practitioner
- Frameworks
- Article 72 · MEASURE 3 · Principle 3
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-10
What it does
It detects decline that would otherwise be silent. Measures of what the model is receiving and how well it is performing are tracked continuously against the levels seen at validation, with defined action when they depart. The model owner is accountable for acting on the results; engineering operates the monitoring for as long as the model is in use.
How to implement it
- Record at validation the baseline distributions of inputs, outputs and performance.
- Measure accuracy against actual outcomes wherever outcomes become known.
- Where they do not, track proxies: input drift, score distribution, override rates, user corrections.
- Break measures down by segment, so that decline in one group is not averaged away.
- Set thresholds with a defined action and owner for each.
- Flag individual inputs that fall outside the range seen in training.
Evidence it operates
Design: a monitoring plan listing measures, baselines, thresholds and actions. Operation: current dashboards or reports; alert history with the action taken; a record of a model retrained, restricted or withdrawn as a result.
Where it fits
Runs throughout production. It is the primary control for AI model drift and degradation, poor generalisation outside training conditions and self-reinforcing feedback loops, and builds on the data collected by AI system monitoring and logging.
Limits
True outcomes often arrive late or never, as with a loan that takes years to default or an applicant who was declined. Proxy measures can stay steady while accuracy falls. Monitoring that generates alerts without an owner who must act on them changes nothing.
Risks this mitigates
Primary mappings
- AI model drift and degradation High
- Measures inputs and output quality against thresholds that trigger action. A deployed model's performance declines as the world it operates in moves away from the data it was trained and validated on, with no change to the model itself.
- AI monitoring and alerting gap High
- Measures quality and inputs against thresholds that trigger action. A system's behaviour in production is not measured against thresholds that trigger action, so degradation, misuse and harm are noticed late or by accident.
- Poor generalisation outside training conditions High
- Detects inputs unlike the training data. The model meets inputs unlike its training data and fails, without signalling that it is outside its competence.
- Self-reinforcing feedback loops Medium
- Monitors outcome distributions for narrowing over time. A model's outputs shape the data it is later trained or evaluated on, entrenching early errors and narrowing outcomes over time.
Supporting mappings
- Post-market monitoring gap Medium
- Supplies the performance evidence. No plan exists to collect and act on evidence of how a system performs and what effects it has after release.
- Rogue agent behaviour High
- Detects behavioural change over time. An agent departs from its intended function or acts deceptively or outside its mandate, through misaligned objectives, emergent behaviour or a compromise that persists, while each action appears legitimate.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| EU AI Act | Article 72 | Post-market monitoring Post-market monitoring. |
| NIST AI Risk Management Framework (AI RMF 1.0) | MEASURE 3 | Mechanisms track identified risks over time Tracking over time. |
| SS1/23: Model Risk Management Principles for Banks | Principle 3 | Model development, implementation and use Performance monitoring. |
Related terms
How to cite
Axiom Verity (2026). Drift and performance monitoring. Reference Library, C-050. https://axiom-verity.com/controls/drift-and-performance-monitoring. Reviewed 2026-11-10.
Found an error or a better source? Report a correction.