# ISO/IEC 42001: event logs for AI agents

ISO/IEC 42001:2023 is the management system standard for AI. One control in its Annex A is about logs: A.6.2.8, AI system recording of event logs. This guide says what the standard asks for in plain words, what a signed record of an agent's tool calls gives towards it, and what it does not.

Nothing here is legal or certification advice. The standard is paraphrased, not quoted: read the wording in your own copy.

## What the standard asks for

- **A.6.2.8: event logs.** Decide at which stages of an AI system's life event logs are kept, and keep them at least while the system is in use. Annex B gives the guidance for this control under B.6.2.8.
- **Clause 9.1: monitoring, measurement, analysis and evaluation.** Decide what is monitored and how, and keep evidence of the results.
- **Clause 7.5: documented information.** The information the management system relies on is controlled: protected, available when needed and kept for as long as it is needed.

Annex A is a reference set of controls. An organisation states which of them apply to it and why, and an auditor then looks for evidence that each one operates.

## What this means for an agent

An agent acts through its tools, so its event log is the list of its tool calls: which tool, with what input, what came back, and when. An auditor of the management system asks to see that log for a period, and asks how you know it is complete and unchanged.

A log you write and keep yourself cannot answer the second question: [Why your own logs are not evidence](/docs/guides/why-your-own-logs-are-not-evidence).

## What a record gives

- **An event log kept while the system is in use.** Each tool call is written as a signed `decision` and `outcome` record as the agent runs: by wrapping the tools, through the gateway, or from the traces the agent already has.
- **Protected against change.** An edited record breaks its signature and a removed one leaves a gap in the agent's chain. A witness co-signs the records, so neither you nor the witness can rewrite them alone.
- **Evidence for a period.** An export covers any period, and the auditor verifies it record by record on their own machine with the open verifier: [Verify](/docs/verify).
- **A stated scope.** The signed tool list shows which of the agent's tools are recorded and which are not: [Tool coverage](/docs/coverage).

## What it does not give

- **Not certification.** A certificate comes from a certification body's audit of the whole management system. Records are evidence for one control and for the monitoring that relies on it.
- **Only tool calls that go through it.** A call made around a wrapped tool or the gateway leaves no record, and other event logs the organisation decides to keep need their own logging.
- **Not whether the decision was right.** A record proves who signed it, that it has not changed and when it existed.
- **From traces, the decision can come after the tool ran.** Records made from OpenTelemetry spans say whether the decision was written before or after, and a sampled-out span has no record: [Audit log from OpenTelemetry traces](/docs/frameworks/opentelemetry).

Audit packs that map the records to A.6.2.8 and clause 9.1 are part of Enterprise: [Audit packs](/docs/audit-packs).

Next: start free, with the SDK alone and no account. [Quickstart](/docs/quickstart)
