Compliance Guide

Log Retention Obligations for Deployers of High-Risk AI Systems Under the EU AI Act

Deployers of high-risk AI systems must keep the logs those systems generate automatically. The obligation has a minimum retention period with qualifiers, and other Union or national law, including data protection and sector record-keeping rules, can affect it.

What the deployer obligation asks for

Deployer log retention comes down to four questions, summarized below.

Deployer log retention at a glance
AreaWhat to do
ScopeIdentify which automatically generated logs sit under the deployer's control.
PeriodApply the Act's minimum period and document any qualifier that changes it.
Other lawReconcile with GDPR storage limitation, data minimization, and sector rules.
EvidenceProve integrity, access control, and policy-driven retention and deletion.

Separating deployer duties from provider duties

Scope starts with identifying which automatically generated logs sit under the deployer's control. For each high-risk system, mark which log sources you control, which the provider controls, and which belong to a third party.

Scoping logs in agentic and multi-system workflows

Scope becomes harder when a workflow crosses several systems. An agent may call tools, query data stores, and hand work to other agents, with each step logged in a different place. A single decision can leave fragments across the model runtime, an orchestration layer, tool servers, and downstream applications.

Start from the high-risk system and trace outward. For each system, list the log sources that record its inputs, outputs, and operation, then mark which you control, which the provider controls, and which belong to a third party. Correlation identifiers matter here, because logs that cannot be tied to the same event are of limited value as audit evidence even when every fragment is retained.

Scope decisions should be reviewed when workflows change. Adding a tool, a new agent, or a new data connection can create log sources that nobody classified.

Where other law changes the retention answer

Other Union or national law, including data protection and sector record-keeping rules, can affect the retention obligation. Reconcile the Act's minimum period with GDPR storage limitation, data minimization, and any sector rules that apply.

Record the outcome in a written retention schedule that states the period applied, the legal basis, and any qualifier or competing rule considered.

Controls that make retention defensible

  • A log inventory that maps each high-risk system to its log sources and assigns control to deployer, provider, or third party.
  • A written retention schedule stating the period applied, the legal basis, and any qualifier or competing rule considered.
  • Integrity protection, such as write-once storage or tamper-evident records, so logs can be shown to be unaltered.
  • Role-based access control with least privilege, and a record of who accessed or exported logs.
  • Automated enforcement of retention and deletion, with exceptions such as legal holds recorded and reviewed.
  • Periodic testing that logs are complete, retrievable, and correlated across systems.

How runtime governance can support evidence collection

Retention is defensible when you can prove integrity, access control, and policy-driven retention and deletion. If your high-risk AI workflows span several systems, runtime governance and audit logging can support a consistent evidence trail.