The EU AI Act becomes enforceable for high-risk AI systems on 2 December 2027 (Annex III standalone systems) and 2 August 2028 (high-risk AI embedded in regulated products), under the Digital Omnibus deferral pending final EU adoption. If your organization deploys AI agents in finance, healthcare, employment, critical infrastructure, or law enforcement, those systems likely fall under Annex III high-risk categories. Fines reach 35 million EUR or 7% of worldwide annual turnover.
What the regulation requires
Article 12: Record-Keeping
High-risk AI systems must include automatic recording of events throughout the system's operational lifetime. Logs must capture: period of each use, reference database, input data, and records that allow output verification. Logging must "conform to recognised standards or common specifications." Ad-hoc text files will not suffice. The NIST AI Risk Management Framework provides complementary guidance on structured logging practices.
Article 19: Log Retention
Providers must retain automatically generated logs for at least six months. Logs must be available to national competent authorities upon request in a retrievable, exportable format.
Article 26: Deployer Obligations
Deployers (not just providers) must monitor AI system operation, retain logs independently, ensure input data relevance, and inform individuals subject to AI decisions. Audit trail capabilities must be designed in from the start.
Supporting Articles
Article 9 requires lifecycle risk management, with audit trails as the primary data input. Article 13 requires transparency through reviewable operational records. Article 14 requires human oversight, which depends on accessible audit trails. ISO 42001 provides a management system framework aligned with these requirements.
Audit trail checklist
- Every agent action logged automatically. No optional logging that developers can toggle off.
- Structured, complete records. Timestamp, agent identity, action type, input data, output, context.
- Tamper-evident. Cryptographic measures, not just access controls. If logs can be silently modified, they cannot serve as evidence.
- Retained for at least six months. Enforced by infrastructure, not documentation.
- Exportable for regulator review. Structured, readable format within a reasonable timeframe.
- Independently verifiable. Auditors can verify records have not been altered without relying on your assertion.
- Accessible for human oversight. Interpretable by people responsible for overseeing the AI system.
Most organizations satisfy the first two and struggle with the rest.
Why traditional logging falls short
Application logs are mutable. Anyone with server access can edit them. Database records can be updated or deleted. Timestamps can be fabricated without independent anchoring. Without cryptographic proof, "we logged it" is a claim, not evidence.
How to prepare
Step 1: Identify which AI systems fall under Annex III high-risk categories.
Step 2: Evaluate logging against the seven-point checklist above. Document the gaps.
Step 3: Implement cryptographic audit trails. The single largest gap for most organizations is tamper-evidence.
Asqav provides tamper-evident audit trails for AI agents. Every action is signed with ML-DSA (NIST FIPS 204), creating tamper-evident records that support Articles 12, 19, and 26. See our EU AI Act compliance page for a detailed mapping.