Compliance Guide

ETSI EN 304 223 audit evidence: runtime logging and monitoring requirements

Treat every ETSI EN 304 223 provision that touches logging, monitoring or record keeping as an evidence obligation. For each one, define the record that proves the control operated at runtime: the interaction, the policy decision and the rule that governed it, any tool call, monitoring outputs, and the retention settings. Then test whether your runtime produces those records without reconstruction.

  • Interaction records: time stamp, user or process identity, the model or agent involved, and the outcome of each interaction.
  • Policy decision records: the decision taken, the rule invoked and the policy version in force at that moment.
  • Tool call records: which tool was called, under which identity, whether the call was permitted, and the result.
  • Access events: who or what accessed the AI system and its records, and whether access succeeded.
  • Monitoring outputs: alerts, review results and incident reports, retained as operational evidence.
  • Retention and deletion settings: period, storage location and deletion process, with a history of changes.
  • Pick one interaction and retrieve its record with time stamp, identity, outcome and the rule applied.
  • Show the policy version in force at the time of that interaction without rebuilding it from change history.
  • Retrieve a tool call and show whether it was permitted, and by which rule.
  • Export alerts and review results and confirm they are retained, not only generated.
  • Show that the retention period and deletion process follow written policy, and that changes to those settings are themselves recorded.
  • Simulate a logging failure or full storage and show that it is detected and handled.

Read the logging and monitoring provisions as evidence obligations

Teams that adopt ETSI EN 304 223 as a baseline for securing AI systems often start by asking what must be logged. An auditor starts from a different question: what record shows that a required control operated, at runtime, for a specific interaction? That shift changes the work. Each provision that touches logging, monitoring or record keeping should be translated into a named control, the record that proves it operated, an owner, and a retention setting.

The distinction matters because application logs usually show what happened, not what the system was permitted to do. An auditor can verify an action from a log. Verifying that the action was authorized requires a record of the policy decision and of the policy that governed it.

NIST guidance gives a useful analogue for what auditors treat as a record. Control AU-2 in SP 800-53 Rev. 5 asks organizations to identify the event types a system can log in support of audit, and to choose them according to the monitoring and auditing needs of each implemented control. That is a NIST control, not an ETSI provision. The principle still carries over: records are defined event types with defined content, selected because a control depends on them.

Runtime evidence versus reconstructed evidence

The practical difference is when the governing policy is captured. If the decision and the record are produced by the same runtime event, the link between them is part of the record. If a team reassembles the story later from application logs, tickets and configuration history, the link is an inference the auditor has to accept.

  • Governing rule: Runtime: recorded with each decision. Reconstructed: inferred from configuration history, which may have changed since.
  • Completeness: Runtime: every enforced interaction produces a record. Reconstructed: depends on which systems happened to log.
  • Timing: Runtime: evidence accrues continuously. Reconstructed: assembled when an audit is announced.
  • Auditor effort: Runtime: records can be sampled and exported. Reconstructed: each sample needs an explanation of how it was derived.

Who produces the evidence when several parties are involved

AI systems are often built by one party and operated or hosted by another, with models, tools and data coming from several places. Evidence then spans organizational boundaries. Whatever the roles in the standard assign, the compliance lead should be able to show, for each record type, who generates it, where it is stored, who can read it and who owns its retention.

NIST AU-2 asks organizations to coordinate event logging with other entities that need audit-related information, and the Rev. 5 baselines list cross-organizational auditing as its own control topic. The NIST AI RMF names system operators, evaluators and auditors, compliance experts and organizational management among the actors in the operate and monitor phase. Bringing auditors and compliance experts into the definition of evidence early avoids discovering a gap during the audit itself.

Monitoring practice for deployed AI is still maturing. NIST's March 2026 report on monitoring deployed AI systems organizes open questions into who, what, when, why and how to monitor. Those are useful prompts for internal decisions. Record the answers, along with retention periods, deletion rules and review cadence, in policy so they can be shown to an auditor.

Where runtime enforcement fits

Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Each individual interaction or tool call can be allowed, redacted, blocked, rate limited or escalated for human approval, and that enforcement is applied at runtime, not reviewed after the fact. The decision generates continuous audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST.

Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, so storage location and retention periods can be set to your policy. In production, the platform generates 30+ GB of audit evidence per day. Trussed does not claim ETSI certification, and its evidence mapping is limited to the frameworks listed here.

Evidence chain for one AI interaction

Interaction record

Time stamp, identity, model or agent, outcome.

Policy decision

Allow, redact, block, rate limit or escalate, with the rule and policy version.

Tool call record

Which tool, under which identity, permitted or not.

Monitoring output

Alerts and review results retained as evidence.

Retention setting

Period, location and deletion process, with change history.

Check your runtime evidence against what an auditor can verify

Walk through your interaction records, policy decisions, tool call records and retention settings with a specialist.

Talk to an Expert