Luca BertonArticle 72 of the AI Act requires providers to establish post-market monitoring systems "proportionate to the nature of the AI technologies and the risks of the high-risk AI system." Translation: you need continuous monitoring infrastructure that catches problems before they become incidents.
What to Monitor
Performance Degradation
- Accuracy metrics tracked against deployment benchmarks
- Prediction distribution shifts (concept drift, data drift)
- Latency and throughput degradation
- Error rate trends by category, geography, and time period
Fairness Drift
- Continuous fairness metric computation across protected groups
- Disparate impact ratio monitoring
- Equalized odds tracking over time
- Subgroup performance comparison against deployment benchmarks
Safety Signals
- Edge case frequency — inputs outside training distribution
- User complaint correlation with model predictions
- Human override rates — increasing overrides suggest model degradation
- Adversarial input detection
Alerting and Escalation
Define alerting thresholds tied to AI Act requirements:
- Warning — Metric approaching threshold, investigate within 24 hours
- Critical — Metric breached threshold, initiate remediation
- Incident — Serious risk to health, safety, or fundamental rights — trigger incident response and consider reporting to market surveillance authority (Article 62)
Kubernetes Recipes
A practical guide for container orchestration and deployment by Grzegorz Stencel & Luca Berton (Apress).
Watch on Skillshare →Serious Incident Reporting
Article 62 requires reporting serious incidents to market surveillance authorities. Your infrastructure must: detect incidents automatically, classify severity against AI Act criteria, generate structured incident reports, and submit within regulatory timelines. This mirrors CRA and DORA reporting — build a unified reporting pipeline.