Post-Market Monitoring

The foundational mechanism of PMM involves the automated ingestion of operational logs directly from the deployed AI system. This is not a periodic audit; it is a persistent data pipeline.

• Telemetry Streams: Captures input data distributions, model inference outputs, and confidence scores.
• Drift Detection: Compares live production data against the baseline training data provenance record to identify data drift.
• User Feedback Loops: Systematically logs reported errors or overrides from the human-in-the-loop interface to flag potential safety issues.

PMM mandates that the initial residual risk disclosure is a living document. Providers must actively search for unknown failure modes that were not apparent during the pre-market conformity assessment.

• Safety Signal Detection: Uses statistical anomaly detection to identify clusters of unexpected outputs that may constitute a new hazard.
• Substantial Modification Trigger: If the system's real-world performance deviates significantly from the intended purpose declaration, it legally triggers a re-registration requirement.
• Incident Reporting Linkage: Automatically correlates system behavior with the mandatory reporting portal to ensure a serious incident is filed within the regulatory deadline.

PMM is not an isolated activity; it is a mandatory component of the provider's audited Quality Management System. The QMS must define explicit procedures for feedback loops.

• Corrective Action Protocols: Documented steps for how the organization reacts to PMM findings, including model rollback or emergency patching.
• Audit Trail Immutability: All monitoring data and subsequent human decisions must be logged immutably to satisfy AI audit trail requirements.
• Notified Body Access: The data collected during PMM must be structured and preserved for inspection by the Notified Body during periodic QMS audits.

PMM extends beyond safety to ensure the model maintains its ethical and performance standards over time. A model that was fair at launch can become biased due to environmental shifts.

• Fairness Metric Decay: Continuously calculates disparate impact ratios and other fairness metrics against live data to detect emergent bias.
• Model Drift Evaluation: Compares the current model accuracy against the baseline established in the model card submission to detect performance degradation.
• Segmented Analysis: Breaks down performance by demographic or operational segments to ensure no specific user group is suffering from silent model failure.

The technical mechanism connecting a registered AI system's Unique Registration ID to a mandatory incident reporting portal. When a serious incident or malfunction is detected during post-market monitoring, this linkage triggers an immediate notification to the National Competent Authority. The process requires providers to document the root cause, implement corrective actions, and update the EU AI Act Database within 15 days of becoming aware of the incident.

The automated, real-time verification of AI systems against evolving regulatory standards through policy-as-code enforcement. This process ensures that post-market performance data is continuously checked against the original Declaration of Conformity parameters. Key components include:

• Automated drift detection from registered intended purpose
• Real-time bias metric tracking against fairness thresholds
• Integration with Quality Management System audit trails

A change to an AI system's intended purpose or performance characteristics that triggers a new Conformity Assessment and re-registration obligation. Post-market monitoring must detect when model updates, retraining, or context shifts constitute a substantial modification. Examples include:

• Expanding from adult to pediatric diagnostic use
• Changing from advisory to fully autonomous decision-making
• Integrating new data modalities not covered in the original Technical Documentation File

The mandatory declaration of any remaining risks that could not be mitigated, which must be transparently communicated to end-users and recorded in the registration database. Post-market monitoring validates whether these disclosed residual risks manifest in practice and whether their severity or frequency exceeds initial projections. Providers must update disclosures if real-world data reveals unanticipated failure modes not captured during pre-market testing.

The formal obligation to inform the Market Surveillance Authority and update the EU database when a registered AI system is recalled or withdrawn. Post-market monitoring serves as the early warning system that triggers this process when:

• Safety performance degrades below acceptable thresholds
• Systemic biases emerge that cannot be remediated
• The system no longer meets the essential requirements of the EU AI Act

The assessment of a provider's internal processes for design, testing, and post-market monitoring. Certification of the QMS is a prerequisite for self-assessment and registration. Auditors evaluate whether the monitoring plan includes:

• Defined KPIs for safety and performance
• Procedures for collecting and analyzing user feedback
• Documented escalation paths for adverse events
• Integration with the Incident Reporting Linkage mechanism