Post-Market Monitoring

Learning Objectives

By the end of this chapter, you will be able to:

Design a post-market monitoring system meeting Article 72 requirements
Establish data collection mechanisms with deployers and users
Implement analysis procedures to detect compliance issues
Trigger appropriate corrective actions based on monitoring data
Report serious incidents according to Article 73 requirements

Post-Market Monitoring Framework (Article 72)

Post-market monitoring (PMM) ensures high-risk AI systems remain compliant throughout their operational lifetime. Unlike one-time conformity assessment, PMM is continuous and responsive to real-world performance.

Legal Requirements

Article	Requirement	Key Elements
Article 72(1)	Establish PMM system	Proportionate to nature, risk
Article 72(2)	Active and systematic collection	Performance data throughout lifetime
Article 72(3)	Documentation and availability	Plan documented, available to authorities
Article 72(4)	Integration with existing PMM	Providers under Union harmonisation legislation or financial services law may integrate AI Act PMM into existing systems
Article 73	Serious incident reporting	Report to market surveillance authorities within 2, 10, or 15 days depending on incident type

PMM System Elements

Element	Purpose	Implementation
Data collection	Gather performance information	Feedback channels, telemetry, reports
Analysis procedures	Evaluate collected data	Metrics, thresholds, trend analysis
Corrective action triggers	Identify when action needed	Criteria, escalation procedures
Documentation updates	Keep records current	Version control, change logs
Authority reporting	Meet regulatory obligations	Reporting templates, procedures

Data Collection Framework

Data Sources

Source	Data Type	Collection Method
Deployers	Operational performance, incidents, feedback	Regular reports, surveys, incident forms
Users	Complaints, satisfaction, error reports	Feedback channels, support tickets
System telemetry	Technical metrics, logs, performance data	Automated collection (with consent)
Market surveillance	Regulatory feedback, inspection findings	Authority communications
External research	Academic studies, third-party audits	Literature monitoring, commissioned studies
Media/public	Reports, complaints, incidents	Media monitoring, public feedback

Performance Metrics

Metric Category	Specific Metrics	Monitoring Frequency
Accuracy	Precision, recall, F1, error rates	Continuous or periodic
Reliability	Uptime, failure rates, consistency	Continuous
Fairness	Demographic performance gaps, bias indicators	Periodic (quarterly)
Safety	Near misses, incidents, risk indicators	Continuous
User satisfaction	Feedback scores, complaint rates	Ongoing
Performance drift	Changes from baseline metrics	Continuous

Deployer Feedback Mechanisms

Mechanism	Purpose	Frequency
Regular performance reports	Systematic data collection	Monthly/quarterly
Incident reporting forms	Capture adverse events	As incidents occur
Annual surveys	Comprehensive feedback	Annually
Support interactions	Issue identification	Ongoing
Contract review meetings	Strategic feedback	Annually/bi-annually

User Feedback Channels

Channel	Type	Data Captured
In-product feedback	Integrated reporting	Real-time issues, satisfaction
Support tickets	Issue-based	Problems, errors, complaints
Complaint mechanisms	Formal complaints	Serious concerns, rights issues
User research	Structured research	Usability, satisfaction, impact

Analysis and Evaluation

Analysis Framework

Analysis Type	Purpose	Frequency
Trend analysis	Detect performance changes over time	Monthly
Threshold monitoring	Identify breaches of acceptable limits	Continuous
Root cause analysis	Understand causes of issues	Per incident
Comparative analysis	Compare across deployments	Quarterly
Bias monitoring	Detect emerging fairness issues	Quarterly
Risk reassessment	Update risk profile	Annually or on trigger

Performance Thresholds

Metric	Green	Amber	Red
Accuracy	Within specifications	5-10% below spec	>10% below spec
Error rate	Baseline	2x baseline	>3x baseline
Complaint rate	< 0.1% users	0.1-0.5% users	> 0.5% users
Bias indicator	< 5% gap	5-10% gap	> 10% gap
Incident rate	0 serious	Minor incidents	Serious incident

Trigger Points for Action

Trigger	Indicated Action	Timeline
Red threshold breach	Immediate investigation, potential pause	Hours to days
Amber threshold breach	Enhanced monitoring, remediation planning	Days to weeks
Trend toward threshold	Proactive investigation	Weeks
Serious incident	Incident response, authority notification	Immediate
Pattern across deployments	Systematic review	Days

Corrective and Preventive Actions

Corrective Action Categories

Category	Trigger	Actions
Technical fixes	Performance issues, errors	System updates, patches
Documentation updates	Changed understanding, new limitations	Revised documentation
Training updates	User/deployer misunderstanding	Enhanced guidance
Deployment restrictions	Unsafe in certain contexts	Use case limitations
System updates	Model improvements needed	Retraining, enhancement
Withdrawal	Fundamental safety/compliance issues	Market removal

Corrective Action Process

Stage	Activities	Timeline
Detection	Issue identified through monitoring	Day 0
Assessment	Severity evaluation, scope determination	Days 1-3
Planning	Corrective action plan developed	Days 3-7
Implementation	Changes made, tested, deployed	Days 7-30
Verification	Confirm effectiveness	Days 30-60
Documentation	Update records, close issue	Post-verification

Documentation Update Requirements

Update Type	Trigger	Records to Update
Performance update	Metrics change significantly	Technical documentation
Limitation discovery	New limitation identified	Instructions, model card
Risk update	Risk profile changes	Risk management file
Incident record	Incident occurs	Incident log
Version change	System updated	Version history, changelog

Serious Incident Reporting (Article 73)

Definition of Serious Incident

Category	Examples	Severity
Death	AI decision/action contributed to death	Highest
Serious damage to health	Physical injury, significant psychological harm	Highest
Serious property damage	Significant financial loss, destruction	High
Serious/irreversible environmental damage	Environmental harm	High
Critical infrastructure disruption	Essential service interruption	High
Fundamental rights breach	Discrimination, rights violation at scale	High

Reporting Timeline

Event	Timeline	Action
Incident occurs	T+0	Become aware
Initial assessment	Within 24 hours	Determine if serious incident
Authority notification	15 days for general incidents (Art. 73(2)); 2 days for widespread infringements (Art. 73(3)); 10 days for deaths (Art. 73(4))	Initial report
Full report	As soon as possible	Detailed incident report
Follow-up	As required	Additional information, updates

Report Content Requirements

Element	Content
System identification	AI system name, version, unique ID
Incident description	What occurred, when, where
Impact assessment	Who affected, nature of harm
Root cause (if known)	Preliminary or confirmed cause
Immediate actions	Steps taken to address
Planned actions	Further remediation planned
Contact information	Point of contact for follow-up

Reporting Process

Step	Action	Responsibility
1. Detection	Identify potential serious incident	Operations/Support
2. Escalation	Notify incident response team	Detector
3. Classification	Determine if serious incident	Compliance/Legal
4. Notification	Report to relevant authority	Compliance
5. Investigation	Full root cause analysis	Technical team
6. Response	Implement corrective measures	Cross-functional
7. Follow-up	Provide updates to authority	Compliance
8. Closure	Document resolution	Compliance

PMM System Design

Organisational Structure

Role	Responsibilities
PMM Owner	Overall system responsibility, authority liaison
Data Analyst	Collect and analyse monitoring data
Technical Lead	Assess technical issues, implement fixes
Compliance Lead	Regulatory reporting, documentation
Deployer Relations	Manage deployer feedback channels
Quality Assurance	Verify corrective actions effective

System Architecture

Component	Function	Implementation
Data collection	Gather inputs from all sources	APIs, forms, integrations
Data storage	Store monitoring data securely	Database with retention
Analysis engine	Process and analyse data	Analytics platform
Alert system	Notify of threshold breaches	Automated alerts
Reporting	Generate reports for stakeholders	Dashboards, reports
Documentation	Maintain records	Document management

Integration with QMS

QMS Element	PMM Integration
Document control	PMM procedures documented, controlled
Corrective actions	PMM-identified issues in CAPA system
Management review	PMM data in management reviews
Internal audits	PMM system audited
Training	PMM responsibilities in training

PMM Documentation Requirements

PMM Plan

Section	Content
Scope	AI systems covered, monitoring boundaries
Data sources	All sources, collection methods
Metrics and thresholds	What measured, acceptable limits
Analysis procedures	How data analysed, frequency
Action triggers	When action required
Responsibilities	Who does what
Reporting procedures	Internal and regulatory reporting
Review cycle	When PMM system reviewed

Records to Maintain

Record	Content	Retention
Performance data	Raw and analysed monitoring data	Lifetime + 10 years
Incident reports	All incidents, serious and minor	Lifetime + 10 years
Corrective actions	Actions taken, effectiveness	Lifetime + 10 years
Authority reports	Submitted reports, correspondence	Lifetime + 10 years
Analysis reports	Periodic analysis outputs	Lifetime + 10 years
System changes	Updates made based on monitoring	Lifetime + 10 years

PMM Implementation Checklist

System Setup

PMM Owner designated
PMM Plan documented
Data collection mechanisms established
Analysis procedures defined
Thresholds and triggers specified
Reporting templates prepared
System integrated with QMS

Deployer Relationships

Feedback mechanisms communicated
Reporting requirements in contracts
Regular review meetings scheduled
Incident escalation path clear
Training provided on feedback

Ongoing Operations

Data collection active
Analysis performed per schedule
Thresholds monitored continuously
Corrective actions tracked
Documentation updated
Management reviews conducted

Authority Readiness

Reporting procedures in place
Contact points identified
Report templates available
Escalation path to legal/compliance clear
Training on serious incident definition