Implementation Guide

Applying the IIA AI Auditing Framework to LLM Applications

To apply the IIA AI Auditing Framework to LLM applications, treat the framework as the top-level structure and build an LLM-specific layer beneath it. Inventory every application, model, retrieval pipeline, tool, and agent with a named owner. Map each framework domain to specific LLM risks and control objectives. Define tests that examine the controls around non-deterministic outputs rather than the outputs alone. Confirm that runtime logs, policy enforcement records, and change records exist to support those tests.

From framework to audit evidence

The approach moves through four stages, each building on the one before it.

StageDescription
1. ScopeInventory LLM applications, models, pipelines, tools, agents, and owners.
2. MapLink each framework domain to LLM risks and control objectives.
3. TestDefine design and operating effectiveness tests for each control.
4. EvidenceIdentify the runtime records that show controls operated in production.

Scope the LLM system inventory first

Inventory every application, model, retrieval pipeline, tool, and agent, and assign each a named owner. Build the LLM inventory with owners, data access, and tool or agent capabilities.

Decide the sampling approach and population source before fieldwork begins, so that testing draws on a defined and complete inventory.

Map domains to LLM risks and control objectives

Treat the IIA framework as the top-level structure and build an LLM-specific layer beneath it. Map each framework domain to specific LLM risks and control objectives, and write testable control objectives for each framework domain you adopt.

Example risk-to-control mapping

A risk-to-control mapping connects each adopted framework domain to its LLM risks and control objectives, then to the tests that show whether each control is designed and operating effectively, and finally to the evidence that supports those tests.

Identify the evidence sources and where runtime governance helps

Confirm that the following records exist to support your tests:

  • Runtime logs
  • Policy enforcement records
  • Change records

Confirm log completeness, retention, tamper evidence, and auditor access. Record gaps where evidence cannot be produced and report them as control findings.

Audit preparation checklist

  1. Obtain the current IIA framework text and note any updates issued since your last review.
  2. Build the LLM inventory with owners, data access, and tool or agent capabilities.
  3. Write testable control objectives for each framework domain you adopt.
  4. Decide the sampling approach and population source before fieldwork begins.
  5. Confirm log completeness, retention, tamper evidence, and auditor access.
  6. Record gaps where evidence cannot be produced and report them as control findings.

Review your audit evidence for agent actions

If your team needs runtime policy enforcement records, agent identity, and audit logging for AI agents, speak with a Trussed AI expert about how these map to your audit objectives.

Talk to an Expert