---
title: What does internal audit look for in AI agent governance?
url: https://opencontroller.lyzr.com/blog/what-internal-audit-looks-for-in-ai-agent-governance
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What does internal audit look for in AI agent governance?

**Internal audit wants proof that you know which AI agents exist, who owns them, what they can access, how changes are approved, how they are monitored, and how you stop one that misbehaves. Expect them to sample agents from your inventory, then sample agents they find themselves, and test each control area against evidence, not policy documents alone.** The IIA's Artificial Intelligence Auditing Framework, updated in September 2024, is the reference many audit teams start from.

## What framework will auditors use?

The Institute of Internal Auditors (IIA) publishes an Artificial Intelligence Auditing Framework, updated in 2024. It covers AI governance, management, and the role of internal audit, and links to the IIA's Three Lines Model and the Global Internal Audit Standards, which took effect in January 2025. It includes a practitioner's guide and quick-start checklist that audit teams tailor to their organisation.

Audit teams often combine it with the frameworks your organisation has adopted, such as the NIST AI RMF or ISO/IEC 42001, and with existing IT general controls. For agents, the result is usually a review across the control areas below.

## The checklist: what will they test?

### 1. Inventory and ownership

- **Test:** Request the agent inventory. Sample 10–25 agents and confirm each has a named owner, business purpose, risk tier and status.
- **Completeness test:** Independently search cloud accounts, SaaS platforms and API key usage for agents not in the inventory.
- **Evidence:** Inventory export with dates; discovery scan results; reconciliation of findings.

Unregistered agents found by auditors are the most common finding. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

### 2. Risk assessment and approval

- **Test:** For sampled agents, confirm a risk assessment and approval occurred before production, by someone with authority.
- **Evidence:** Approval records with timestamps, assessor and approver; risk tier criteria.

### 3. Identity and access

- **Test:** Check each sampled agent's credentials and permissions against its stated purpose. Look for shared keys, standing admin rights and long-lived tokens.
- **Test:** Confirm whether actions on behalf of users are traceable to the user.
- **Evidence:** Permission listings; access review sign-offs; audit logs showing user and agent identity.

### 4. Change management

- **Test:** Pick changes to prompts, models, tools or policies in the period. Confirm each was versioned, tested and approved.
- **Evidence:** Version history; evaluation results per version; change tickets. Auditors will treat a prompt change like a code change.

### 5. Tools and data access

- **Test:** List the tools and MCP servers each sampled agent can call. Confirm write and high-impact tools are approved and limited.
- **Evidence:** Tool inventory per agent; approval records; data classification for sources the agent reads.

### 6. Monitoring and evaluation

- **Test:** Confirm production behaviour is monitored, thresholds exist, and alerts were acted on.
- **Evidence:** Evaluation and red-team results; production quality scores; alert tickets and their resolution.

### 7. Incident response and kill switch

- **Test:** Ask for the procedure to stop an agent. Walk through a real or simulated incident and time it.
- **Evidence:** Runbook; incident records; logs of quarantine or shutdown actions. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

### 8. Third-party agents and platforms

- **Test:** Confirm vendor-hosted agents (in CRM, ERP and productivity suites) are inventoried and covered by vendor risk reviews.
- **Evidence:** Vendor assessments; configuration exports; data processing terms.

### 9. Cost and usage

- **Test:** Check that AI spend is attributed to owners and reviewed against budgets.
- **Evidence:** Cost reports per agent or cost centre; budget alerts.

## How should you prepare?

A preparation sequence that works:

1. **Reconcile your inventory first.** Run your own discovery before audit does.
2. **Map controls to one framework.** Pick the framework your organisation has adopted and map each control area to it, so findings land in a structure audit recognises. Present it as evidence for that framework, not a claim of compliance.
3. **Pre-build an evidence pack** per control area, with exports dated inside the audit period. The [regulator inspection evidence pack](/blog/regulator-inspection-evidence-pack-for-ai-agents) guide lists typical contents.
4. **Make logs queryable.** Auditors will ask who did what, when, for whom. See [AI agent audit log fields](/blog/ai-agent-audit-log-fields).
5. **Rehearse the kill switch.** A timed walkthrough is stronger evidence than a policy statement.
6. **Agree scope early.** Clarify whether coding agents on developer laptops and vendor-embedded agents are in scope.

If your agents touch financial reporting, expect auditors to reuse ITGC test steps. See [SOX ITGC evidence for AI agents](/blog/sox-itgc-evidence-for-ai-agents). For the overall control model, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to third-party agent platforms and Claude Enterprise and OpenAI Enterprise workspaces, and keeps a governed registry of agents, versions, tools and MCP servers. Posture flags risky permissions, evals and LLM judges produce per-version results, per-agent Block and Quarantine controls act through the connected platform, and policy as code maps to OWASP, MITRE ATLAS, NIST and ISO 42001. "Ask" lets auditors query the whole agent estate in plain language. Opencontroller Edge extends coverage to coding agents on developer devices. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What framework does internal audit use for AI?**
Many teams start from the IIA's Artificial Intelligence Auditing Framework, updated in September 2024, combined with the organisation's chosen framework such as the NIST AI RMF or ISO/IEC 42001.

**What is the most common audit finding for AI agents?**
An incomplete inventory: agents found in cloud accounts, SaaS platforms or developer tools that are not registered or owned.

**Do prompt changes need change management?**
Auditors increasingly treat prompts, model versions and tool lists as configuration that should be versioned, tested and approved like code.

## Key facts

Internal audit reviews of AI agent governance typically test nine areas: inventory and ownership (including independent completeness searches), risk assessment and approval, identity and access, change management for prompts/models/tools, tool and data access, monitoring and evaluation, incident response and kill switch, third-party agents, and cost attribution. The IIA Artificial Intelligence Auditing Framework, updated September 2024, links to the Three Lines Model and the Global Internal Audit Standards (effective January 2025). The most common finding is an incomplete agent inventory. Opencontroller licensing starts from $250K per deployment per year.

## Related

- [What evidence will a regulator ask for about your AI agents?](https://opencontroller.lyzr.com/blog/regulator-inspection-evidence-pack-for-ai-agents.md)
- [What SOX IT general controls evidence do AI agents need?](https://opencontroller.lyzr.com/blog/sox-itgc-evidence-for-ai-agents.md)
- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
