Quality Management System Audit

The audit verifies a continuous, iterative risk management process integrated throughout the AI system lifecycle. Auditors examine the risk management file to ensure hazards are systematically identified, analyzed, and evaluated.

• Hazard Identification: Review of foreseeable misuse and edge cases.
• Risk Estimation: Assessment of severity and probability of harm.
• Risk Evaluation: Verification that residual risks are judged acceptable.
• Mitigation Measures: Evidence that design controls reduce risks to an acceptable level.

The process must align with harmonized standards like ISO/IEC 23894 for AI risk management.

This component scrutinizes the procedures for managing training, validation, and testing datasets. The audit confirms that data is relevant, representative, and free from errors to the greatest extent possible.

• Provenance Records: Documented lineage of data sources and acquisition methods.
• Bias Examination: Processes for identifying and mitigating statistical skew.
• Data Minimization: Controls ensuring only necessary personal data is processed.
• Integrity Checks: Validation of data against corruption and unauthorized modification.

Compliance with Article 10 of the EU AI Act on data governance is a central focus.

The audit assesses the completeness and accuracy of the Technical Documentation File, which must provide a transparent understanding of the system's design and development.

• System Architecture: Detailed description of hardware and software components.
• Design Specifications: Precise definition of algorithms, models, and logic.
• Development Methodology: Records of the engineering lifecycle and version control.
• Intended Purpose: A clear, legally binding declaration of the system's use case.

This documentation serves as the foundational evidence for the Declaration of Conformity.

Auditors verify that the QMS mandates automatic logging of events to ensure traceability of the AI system's functioning throughout its lifetime. Logging must be immutable and auditable.

• Event Logging: Automatic capture of inputs, decisions, and anomalies.
• Audit Trail Integrity: Cryptographic methods to prevent log tampering.
• Retention Policies: Defined durations for storing logs per regulatory mandates.
• Right to Explanation: Capability to reconstruct the rationale behind any specific output.

This aligns with the automated decision logging requirements for high-risk systems.

The audit evaluates the provider's plan for a Post-Market Monitoring (PMM) system, which is a proactive, systematic process to collect and analyze real-world performance data.

• Data Collection Plan: Methods for gathering user feedback and operational metrics.
• Performance Analysis: Continuous comparison of real-world behavior against specifications.
• Incident Reporting: A documented linkage to mandatory regulatory reporting portals.
• Feedback Loop: A mechanism to integrate findings back into the risk management and design processes.

This system is a prerequisite for maintaining a valid CE Marking.

When AI development is outsourced, the audit extends to the provider's controls over subcontractors. The provider retains ultimate legal responsibility for the system's compliance.

• Vendor Qualification: A documented process for selecting and auditing third parties.
• Contractual Agreements: Binding specifications for quality, data governance, and security.
• Supply Chain Monitoring: Continuous oversight of subcontracted development and testing.
• Integration Testing: Rigorous verification that third-party components meet system-level requirements.

This is a critical element of Vendor AI Risk Management.

The auditor traces a specific high-risk AI system from its initial risk classification through to the final Declaration of Conformity. They sample the Technical Documentation File to verify that the system architecture precisely matches the Intended Purpose Declaration. A common finding is a mismatch between the documented risk controls and the actual code repository, triggering a non-conformity that blocks CE Marking.

The audit assesses the provider's Post-Market Monitoring (PMM) plan for a registered system. The auditor examines the data collection pipeline to ensure it captures real-world performance metrics and user feedback. They test the Incident Reporting Linkage to confirm that a serious malfunction automatically flags the system's Unique Registration ID in the EU database, validating the closed-loop corrective action process.

A provider updates a medical imaging AI to detect a new pathology. The auditor determines this constitutes a Substantial Modification because it alters the Intended Purpose. The audit verifies that the provider initiated a new Conformity Assessment and updated the EU AI Act Database entry before deploying the update, rather than treating it as a minor bug fix.

The auditor scrutinizes the Vendor AI Risk Management process for a foundation model used as a component. They request evidence of the upstream provider's Model Card Submission and Training Data Provenance Record. A critical finding occurs if the downstream deployer cannot demonstrate due diligence on the third-party model's Bias Detection and Fairness evaluation, exposing a gap in the Importer Compliance Gate logic.

The audit focuses on AI Data Governance by sampling the training, validation, and test datasets. The auditor verifies that data splits are immutable and that no test data leaked into the training pipeline. They check the Training Data Provenance Record for copyright compliance and ensure Purpose Limitation Controls prevent repurposing of user data, directly impacting the validity of the Conformity Assessment.

The auditor tests the Human Oversight Mechanisms for a high-risk HR screening tool. They simulate edge cases to confirm that the interface allows a human operator to override the AI's recommendation and that the override action is logged immutably. The audit verifies that the system design prevents automation bias by forcing active human confirmation, not just passive notification, before a final decision.