Internal audit is both an assurance provider and a governed AI user
Internal audit provides assurance over other functions, and when it uses generative AI in its own work it becomes a governed user of that technology. The model below applies the five parts described above to audit work, so that each part can be tested and evidenced.
| Stage | What it establishes |
|---|---|
| Classify | Assign each use case to a risk tier before use. |
| Bound | Define which evidence, models, and tools are in scope. |
| Review | Record named reviewer accountability per work product. |
| Trace | Log prompts, outputs, and tool calls for reproduction. |
Classify use cases before approving tools
Assign each use case to a risk tier before it is used, so that tool approval follows from classification. An approved use-case register with risk tiers and named owners is the starting point, and it gives the audit function something concrete to test.
Set data boundaries and least-privilege access
Define which evidence, models, and tools are in scope for each use case. Document the data classes permitted for each model or tool, and enforce those limits rather than relying on guidance alone.
Grant model, tool, and data access by role and engagement, and revoke it when the engagement closes.
Keep judgment, skepticism, and accountability with the auditor
AI-assisted output does not replace auditor judgment or professional skepticism. Each AI-assisted work product should record a named reviewer and the checks that reviewer performed, so accountability stays with a person.
Make AI-assisted work reproducible with logs and runtime enforcement
Log prompts, outputs, model versions, and tool calls, and retain them per policy, so a reviewer can reproduce any AI-assisted conclusion. Pair logging with runtime enforcement of the policy, so that the controls above are applied when the work is performed.
Controls internal audit can test on itself first
- An approved use-case register exists with risk tiers and named owners.
- Data classes permitted for each model or tool are documented and enforced.
- Model, tool, and data access is granted by role and engagement, and revoked at closure.
- Each AI-assisted work product records a named reviewer and the checks performed.
- Prompts, outputs, model versions, and tool calls are logged and retained per policy.
- Policy, approved tools, and framework mappings are reviewed on a defined schedule.