Show auditors the controls working, not slides about them. Brief internal audit by giving them the agent inventory, the owner and limits for each agent, and one real decision traced from brief to filed verdict. Then give them read-only access to the record so they can choose their own samples. Each step leaves a document.
What is this briefing for, and who owns it?
Internal audit gives the board independent assurance that controls work. Agents are new to most audit plans, and auditors need to understand what an agent is allowed to do, who decided that, and how they can test it without relying on management's word. The briefing equips them to plan that work.
The person accountable for agent governance owns the briefing, usually the head of agent operations or a risk lead. The chief audit executive decides what audit does with it. Auditors must stay independent, so the briefing explains and gives access; it does not tell them what to conclude. As internal audit needs its own view of agent decisions argues, a management dashboard is not enough.
What are the steps, and who acts at each one?
- Person, governance lead, with the chief audit executive: agrees scope. Which agents, which processes, and what audit wants to understand: design, operation, or both.
- Agent, depth 2: drafts the briefing pack from the agent inventory, the policy, each agent's brief and capability matrix, certification results, the change log and the incident log.
- Person, governance lead: checks the pack. Every figure and claim is traced to its source record, because audit will test it.
- Person, governance lead: runs the walkthrough. Pick one real action. Show the agent's brief, the capability and depth that applied, the sources it used and their provenance, the verdict, the approver if it was held, and the filed record.
- Person, platform owner: grants audit read-only access to the inventory and the append-only record, under auditors' own identities.
- Agent, depth 1: retrieves samples auditors choose. Auditors pick the dates, agents or verdict types; the agent pulls the records. Management does not select the sample.
- Person, governance lead: logs questions and commitments made in the session, with owners and dates.
- Person, chief audit executive: decides whether and how agent governance enters the audit plan.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Scope | Nothing | none | Agreeing what audit wants |
| Pack | Drafts from source records | 2 | Checking accuracy |
| Check | Nothing | none | Tracing every claim |
| Walkthrough | Nothing | none | Choosing and presenting the case |
| Access | Nothing | none | Granting read-only rights |
| Samples | Retrieves records auditors select | 1 | Selecting the sample |
| Commitments | Nothing | none | Owning follow-ups |
| Audit plan | Nothing | none | Audit's independent decision |
What evidence does each step leave?
- Scope: a short memo signed by both sides.
- Pack: a versioned document with source references beside each claim.
- Check: the governance lead's sign-off, noting any figure that could not be traced.
- Walkthrough: the record identifier of the action shown and auditors' notes.
- Access: the access request and approval in your identity system, with an expiry.
- Samples: a retrieval log showing what audit asked for and what was returned.
- Commitments: an action log with owners and due dates.
- Audit plan: audit's own planning record.
What should the pack contain?
Keep it short and factual. Auditors want to understand the design, then test whether it operates.
- Inventory: every agent, its owner, its status and the processes it serves.
- Policy: what agents may and may not do across the firm.
- Per agent: the brief, the role scope, the capability matrix with depths, and certification dates. Explain what a brief is, using agent brief: what it means for AI agents, because auditors will want to see that the brief and the deployed agent match.
- Bounds: how every action is checked against role scope, certificate, sources, your limits and platform guardrails, and how the verdict names the bound that decided.
- Change control: how briefs, sources and depths are changed and approved.
- Incidents: what went wrong, what was done, and what changed.
Which controls should audit be able to test themselves?
Completeness of the inventory. Audit compares identities in your access management system with agents on the inventory. Any identity without an entry is an unregistered agent.
Enforcement. Audit picks actions above an agent's depth and checks they were held or refused.
Approvals. For held actions, audit checks a named person approved them and that the person was on the approver roster.
Certification currency. Audit checks no agent acted on a lapsed certificate.
Record integrity. Audit checks the record cannot be edited after the fact.
If you are working towards ISO/IEC 42001, internal audit of the AI management system is expected; ISO/IEC 42001 internal audit of AI agents covers how that audit runs.
What should you measure after the briefing?
Track the questions audit raised that you could not answer from records, the time taken to return each sample, and the commitments closed on time. A sample that takes days to produce is itself a finding.
Where does it usually go wrong?
The briefing is a presentation. Auditors leave with slides and no access. Without their own view they can only test what management chooses to show.
Management picks the walkthrough case. Choose a typical one, or better, let audit pick it from the record.
The pack describes intended design. If depth limits are planned but not yet enforced, say so. Overstating controls in front of audit is worse than a known gap.
Agents outside the inventory are not mentioned. Disclose known unsanctioned use and the plan to deal with it.
See how to prepare AI agents for an audit for the work before fieldwork starts, and agent governance is a bound, not a model for the controls being audited.
Questions people ask
Should internal audit use an agent to test agents?
Audit can use an agent to retrieve and compare records at depth 1, which speeds up sampling. Conclusions and findings stay with auditors, and the retrieving agent should have its own identity and read-only access.
How technical does the briefing need to be?
Not very. Auditors need the control design, where the evidence lives and how to get it. Model internals matter less than what the agent was permitted to do and whether that permission was enforced.
How often should internal audit be briefed?
Once in full before agents enter the audit plan, then whenever an agent gains a new capability, moves to a higher depth or has a significant incident. A short written update between audits keeps the picture current.