Article 9 Risk Management as a Continuous Process
Article 9 of the EU AI Act requires organisations to implement systematic risk management procedures for high risk AI systems. This is not a one-time assessment conducted before deployment. Risk management under Article 9 is a continuous process that evolves as the system operates, data changes and usage expands.
The risk management system must identify and analyse risks associated with the high risk AI system across its entire lifecycle. This includes risks during development, deployment, ongoing operation and eventually retirement. For each identified risk, the organisation must evaluate the probability of occurrence and the severity of potential harm. A system that could make incorrect decisions affecting employment must assess both how likely misclassification is and what consequences workers would face.
Risk evaluation must consider multiple sources. Technical risks include model failures, data distribution shifts, adversarial attacks and edge case performance degradation. Operational risks include human error in system deployment, misconfiguration and inadequate monitoring. Organisational risks include inappropriate use cases assigned to the system or decisions made without adequate human review. External risks include regulatory changes and reputational harm from system failures.
The organisation must then implement risk mitigation measures. These are actions taken to reduce either the probability or severity of identified risks. Mitigation for technical risks might include ensemble methods to improve reliableness, continuous monitoring to detect performance degradation, or retraining protocols when model accuracy falls below thresholds. Mitigation for operational risks includes training for system users, documented procedures for deployment and clear escalation protocols. Mitigation for inappropriate use includes access controls limiting which decisions the system can influence.
The critical phrase in Article 9 is “continuous process.” This means risk management does not stop after initial deployment. The organisation must establish ongoing procedures to monitor for new risks, emerging failure modes and changing circumstances. A system performing well initially might face risks as new data characteristics emerge. A system developed for a narrow use case might face new risks if later deployed in different contexts. Seasonal changes in data, shifts in user demographics or changes in real-world conditions that the system operates on can all introduce new risks requiring re-evaluation.
Documentation is essential. The organisation must maintain written procedures for risk management, including how risks are identified, analysed and mitigated. Risk management documentation must be specific to the high risk system in question. A general policy stating that an organisation conducts risk management does not satisfy Article 9. The documentation must identify actual risks for the actual system, document how those risks were evaluated and explain what mitigation measures were implemented.
Consider a practical scenario. A bank implements a high risk AI system for credit decisions. Article 9 risk management must identify risks including discrimination against protected characteristics, model accuracy degradation over time, data quality problems, inappropriate loan officer overrides and reputational damage from system errors. The bank documents each risk, evaluates probability and severity, and implements mitigation. Mitigation for discrimination includes validation testing against protected characteristics. Mitigation for accuracy degradation includes monitoring key performance metrics and retraining protocols. Mitigation for data quality includes data profiling procedures. Mitigation for overrides includes audit logging and quarterly override analysis. Mitigation for reputational damage includes incident response procedures.
The process continues after deployment. Monthly, the bank reviews performance metrics. They find that model accuracy on loan applications from a particular geographic region has dropped from 92 percent to 87 percent. This suggests a new risk: data distribution shift affecting a subpopulation. The bank re-evaluates this risk, updates their risk assessment and implements new mitigation (retraining the model on regional data or adjusting decision thresholds for that region).
Six months later, the bank expands the system to make decisions on larger loan amounts. This expansion changes risk profiles. The consequences of errors are now higher, so severity of accuracy risks increases. The bank must re-conduct risk evaluation and potentially implement additional mitigation measures.
Article 9 ties directly to other obligations. Risk assessment conducted under Article 9 informs the technical documentation required by Article 11, the logging requirements of Article 12, the instructions for use in Article 13, the human oversight decisions in Article 14, and the quality management system in Article 17. An organisation cannot properly implement these Articles without starting with Article 9 risk management.