Post-Market Monitoring

A documented, proactive strategy for gathering data from real-world use, not passive complaint waiting. This plan must define the specific metrics, collection methods, and feedback channels used to capture performance data.

• User Feedback Mechanisms: Embedded interfaces for deployers and end-users to report issues or unexpected outputs.
• Automated Telemetry: Direct collection of system logs, input/output pairs, and performance metrics from the live environment.
• Relevant Documentation: The plan itself is a critical part of the Technical Documentation required under the EU AI Act (Art. 11).

The core objective of PMM is to detect novel risks that were not reasonably foreseeable during the initial Risk Management System analysis. This involves analyzing collected data for anomalies, drift, and failure modes.

• Data Drift Analysis: Statistically comparing live input data distributions against the original training and validation datasets to detect shifts.
• Model Decay Detection: Monitoring for degradation in key performance indicators (e.g., accuracy, precision) over time.
• Systemic Risk Triggers: For General Purpose AI (GPAI) models, identifying capabilities that could lead to large-scale harm.

A non-negotiable legal obligation triggered when a malfunction or failure of the AI system directly or indirectly leads to a specific harm threshold. This is a reactive but critical component of the PMM framework.

• Reporting Deadline: Providers must notify the relevant Market Surveillance Authority immediately, typically within 15 days of becoming aware of the incident.
• Reportable Harm: Includes death, serious damage to health, or serious and irreversible damage to property.
• Investigation Protocol: The report must include a root cause analysis and a description of the immediate corrective actions taken.

PMM is not just an observation exercise; it mandates a closed-loop system that feeds findings back into the AI lifecycle to trigger mandatory updates. This ensures the system remains compliant over time.

• Mandatory Provider Action: If a risk is identified, the provider must immediately take corrective measures, which may include withdrawing or recalling the system.
• Substantial Modification Trigger: A significant performance change identified through PMM may constitute a Substantial Modification, requiring a new Conformity Assessment.
• Risk Management Update: Findings must be systematically fed back to update the original risk assessment and mitigation strategies.

The responsibility for PMM is shared. Deployers of high-risk AI systems have a legal duty to monitor the system's operation and report any suspected risks to the provider, creating a collaborative surveillance network.

• Operational Monitoring: Deployers must ensure human oversight and monitor for signs of malfunction or unexpected behavior in their specific use context.
• Data Provision: Deployers are obligated to provide relevant data back to the provider to facilitate the provider's own PMM activities.
• Fundamental Rights Impact: Deployers must monitor for impacts on fundamental rights that were identified in their Fundamental Rights Impact Assessment.

PMM is a formal, auditable component of the provider's mandatory Quality Management System (QMS) . It must be a documented, controlled, and continuously improved process, not an ad-hoc activity.

• Audit Trail: All monitoring data, risk analyses, and corrective actions must be logged immutably for audit purposes.
• QMS Documentation: The PMM plan, procedures, and reports are key artifacts reviewed during a Conformity Assessment by a Notified Body.
• Continuous Improvement: The QMS must ensure that lessons learned from PMM are used to improve the design and development of future AI systems.