An AI audit checklist is the work programme of whoever audits an AI system: for each area in scope, the objectives, the test steps (inquiry, inspection, re-performance, analytical procedures), the evidence to request and the sample. It separates control design from operation over the period and scales depth to the risk of the system. This tool builds the programme from the characteristics of the system audited.
How to use
- Name the system and give its type, sourcing (in-house or vendor) and internal risk level.
- Say whether it is high-risk under the EU AI Act, whether it uses personal data and how much it weighs in decisions about people.
- Choose the engagement type and tick the areas in scope.
- Read the programme, the scope warnings and the evidence list by owner, then export as CSV for the working papers, or as DOCX.
What each field asks for
| Field | What to enter |
|---|---|
| AI system name | The name the system goes by in the organisation. |
| System type | Generative AI adds hallucination and prompt injection tests. |
| Sourcing | A vendor system adds contract and vendor documentation tests. |
| Internal risk level | The risk rating from the AI inventory, if there is one. |
| High-risk under the EU AI Act? | Yes adds technical documentation, logging and serious incident tests. |
| Uses personal data? | Yes adds lawful basis, DPIA and minimisation tests. |
| Weight in decisions about people | Automatic decisions add contestation and human intervention tests. |
| Engagement type | Changes the sample guidance and the estimated effort. |
| Areas in scope | Tick the areas the engagement will cover. Each area produces 2 to 4 tests. |
Example
The internal audit function of a fictional insurer, Vale Azul Insurance, plans to audit its customer service virtual assistant, a generative AI service from a vendor, with a high internal risk level and customer personal data, not high-risk under the EU AI Act. With eight areas in scope, the tool produces 28 tests, 16 of them high priority, and estimates 12 to 20 days of work (about 127 hours). Because it is generative AI, hallucination, prompt injection and AI disclosure tests are added; because it comes from a vendor, vendor validation and contract tests are added. A warning flags that performance and monitoring were left out of scope for a high risk system.
When to use
- When planning an internal audit of an AI system already in production.
- For a review of an AI system before deployment.
- When the second line needs to review first line AI controls.
- To prepare for an external audit or a certification audit of the AI management system.
Frequently asked questions
What is an AI audit checklist?
It is the audit work programme: for each area it lists what to test, how to test it, what evidence to request and what sample to use. It helps you plan the work, send out evidence requests and record results in the working papers.
How does the tool choose the tests?
Each area in scope gets two core tests (control design and operation). Conditional tests follow: generative AI adds hallucination, prompt injection and AI disclosure; a vendor adds vendor validation, contract and documentation; personal data adds minimisation; EU AI Act high-risk adds logging, technical documentation and serious incidents. Risk caps the total at 3 or 4 tests per area.
How is priority set?
Core design and operation tests weigh more, high risk raises priority by one level and low risk lowers it by one, and conditional tests go up a level because they address the risk specific to the system. Automatic decisions about people raise the human oversight, bias and transparency areas.
What is the difference between design and operating effectiveness?
A design test checks whether the control, as conceived, addresses the risk. An operating effectiveness test checks whether it worked consistently over the period, using a sample. A well designed control that does not operate does not reduce risk.
Does the programme follow an auditing standard?
It uses widely known audit practice, such as sufficient and appropriate evidence and separating design from operation. The IIA Global Internal Audit Standards have been in force since 9 January 2025 and there is no topical requirement for AI yet. For AI management systems, clause 9.2 of ISO/IEC 42001 covers internal audit. Adapt the programme to the methodology of your function.
Is the result an audit opinion?
No. It is an educational planning aid: it does not replace professional judgement or auditing standards, it is not legal advice and it does not certify the organisation. Conclusions depend on fieldwork and the evidence obtained.
Which DPCert certification covers auditing AI systems?
AUDAI (Responsible AI Auditor) has a domain on auditing AI systems and another on audit fundamentals for AI and data. The prerequisite is RAIP or DPAI.