Implementation Guide

ISO/IEC 24970 Log Event Fields for AI Agents

ISO/IEC 24970 is the AI system logging standard. It describes common capabilities, requirements and a supporting information model for logging events in AI systems, and as of October 2026 it is listed at FDIS stage, under development. For AI agents, define one event schema covering event type, time, source, identity, target, decision and outcome. Capture it at the point of policy enforcement so each record holds the decision and its result, not only the activity. Reconcile the schema against the version of the standard you adopt, and treat the records as evidence of individual interactions and tool calls, not of whole workflows.

Where in the execution path each event can be captured

  1. 1

    Inbound message

    Time, calling user or agent identity, originating application, and the policy decision applied to the content.

  2. 2

    Tool request

    Target tool, requested action, requesting agent, and the decision made before execution.

  3. 3

    Tool result

    Outcome of execution, success or failure, and any redaction applied to returned content.

  4. 4

    Memory and knowledge operations

    Which store was read or written, by which agent, and under which decision.

  5. 5

    MCP and agent-to-agent messages

    Server or peer identified as the target, protocol message type, decision and outcome.

  6. 6

    Escalation and approval

    Hold event, approver identity and final disposition, recorded as separate events.

What the standard covers and how to cite it

ISO/IEC 24970, Artificial intelligence: AI system logging, states its abstract as describing common capabilities, requirements and a supporting information model for logging of events in AI systems. The SC 42 catalogue lists it as ISO/IEC FDIS 24970, under development, at stage 50.20. Earlier iso.org pages still carry AWI and CD labels, so any internal document that references the standard should name the stage it was reviewed against.

That has a practical consequence. Describe your work as alignment with a draft-stage information model, not as conformity or certification against a published standard. Build the schema so fields can be renamed or added when the final text is published, and keep a version note on the mapping.

Two general references are useful as cross-checks, though neither is a 24970 field list. NIST SP 800-171r3 describes audit record content as what type of event occurred, when, where, the source, the outcome, and the identity of the individuals, subjects, objects or entities involved. The OWASP AI Agent Security Cheat Sheet recommends logging all agent decisions, tool calls and outcomes. Use them to test whether your schema is complete, and keep that distinction visible in your documentation.

Define a minimum field schema

Start with a small, strict schema that every capture point populates the same way. The fields below mirror the audit record elements above and the AOS decision events. Treat them as a working design, then map each one to the corresponding element in the standard's information model and record the mapping.

Two design rules matter more than the exact field names. First, log tool requests and tool results as distinct events. This is how blocked calls and executed calls are both evidenced: a blocked request has no result event, and the absence is itself explained by a decision record. Second, keep decision and outcome in separate fields. A call can be allowed and still fail, or be redacted and still succeed.

Where agents, tools and MCP servers emit different formats, normalise at the enforcement point rather than downstream. OCSF is one documented option. AOS extends the OCSF API Activity class (6003) for agent events, so records can reach existing SIEM and security monitoring tools.

  • Event type and identifier: Category such as message, tool_call, tool_result or MCP message, with a unique event ID and an interaction reference.
  • Time: A consistent timestamp format and clock source across all capture points.
  • Source and identity: The agent, the user or entity on whose behalf it acted, and the originating system.
  • Target: The tool, MCP server, data store or peer the event addressed.
  • Decision: Allow, redact, block or escalate, with the policy identifier and version that produced it.
  • Outcome: Success or failure, event-specific result, and the state of the interaction after the event.

Recording policy decisions and mapping vocabularies

The decision field is what separates audit evidence from activity records. It shows what the agent was permitted to do, which is the question an auditor asks first. Record the decision at the moment it is enforced, with the policy version, so the record remains interpretable after policies change.

Vocabularies differ across schemas. AOS decision events use Allow, Deny and Modify. The set used here is allow, redact, block and escalate. A reasonable mapping pairs allow with Allow, block with Deny and redact with Modify, but that is an implementer's design choice, not a standard-defined equivalence. Escalate has no obvious counterpart, so keep your native value as the system of record and store any exported mapping in a documented field.

For redact, record that redaction occurred and the category of content affected, not the removed content itself. For escalate, record the hold, the human approver and the final disposition as linked events, so the approval is evidenced rather than assumed. Security-relevant events, such as repeated blocks, should also feed alerting, as the OWASP guidance suggests.

Making records durable enough to rely on

A well-designed schema does not make a record trustworthy by itself. Audit reliance depends on how records are stored and who can change them. The controls below are internal design decisions to agree with audit and legal teams. They should not be described as requirements of the standard.

For integrity, write to append-only storage and consider hash chaining or signed batches, so that alteration or deletion is detectable. Restrict write access to the enforcement component and separate it from the administrators of the systems being observed. For retention, set periods per record class and make them configurable, since evidence needs may differ from operational needs. For residency, decide where records are stored before deployment, because moving them later is difficult to justify to a reviewer.

Redaction needs the most care. Prompts and tool results can contain sensitive data, so decide whether to store content, a hash, or metadata only, and apply the same rule at every capture point. Storing full content creates a second sensitive data set that needs its own protection. Storing too little may leave the record unable to support a later question.

Limits of per-interaction records, and where Trussed fits

Records captured at the enforcement point describe individual interactions and tool calls. They show who or what acted, what was requested, what was decided and what happened. They do not, on their own, reconstruct a multi-step workflow, establish transaction-level governance across tasks and agents, or document model or dataset provenance. A reviewer who needs those answers will need other evidence, and the page that presents your records should say so plainly.

Trussed is one approach to producing this evidence. It enforces policy in the execution path across models, agents, tools and MCP servers, with decisions to allow, redact, block, rate limit or escalate for human approval, so audit-ready evidence is generated by enforcement at runtime, not reviewed after the fact. Deployment is self-managed in the customer cloud, with configurable retention and data residency. Its scope is individual interactions and tool calls.

A useful test for any approach is a short list of questions. Which standard version was the mapping built against? Which fields come from the enforcement point and which are derived later from application logs? How are blocked and modified calls recorded? What evidence exists for integrity, retention, residency and redaction, and what do the records not show?

From agent event to audit evidence

Capture point

Events recorded where policy is enforced in the execution path.

Field schema

Type, time, source, identity, target, decision, outcome.

Decision

Allow, redact, block or escalate, stored apart from the outcome.

Durable record

Integrity, retention, residency and redaction handled as explicit controls.

Review your agent event schema against enforcement-point evidence

Walk through how decisions and outcomes for agent interactions and tool calls are captured and retained in a regulated enterprise.

Explore Agent and Tool-Call Enforcement