How to prepare for an AI governance audit

What an AI governance audit looks at

An AI governance audit is not an audit of the models. It looks at the management system around them. That means the scope of AI in use, who owns each system, how risk is assessed, what controls sit on top and the records that show all of it happened. ISO/IEC 42001 describes this as an AI management system, with requirements for documented information (clause 7.5), internal audit (9.2) and management review (9.3). A certification audit against that standard tests the same thing a board member, a regulator or an enterprise customer’s due-diligence team would ask to see, so preparing for one prepares you for the others.

What an auditor means by evidence

ISO 19011 describes audit evidence as records, statements of fact or other information that are relevant to the audit criteria and verifiable. Relevant means the evidence bears on the control being tested, not adjacent material that looks reassuring. Verifiable means the auditor can confirm it rather than taking it on the organisation’s word.

In practice a record that stands up has four qualities. It is relevant to the specific control. It is sufficient, so there is enough of it to support the conclusion rather than a single artefact standing in for a practice. It is attributed, so it is clear who did the work and who owned the decision. And it is reliable, so the auditor can trust it is the record it claims to be and has not been altered since. Reliability is the quality organisations most often miss, because a file that could have been edited last night carries less weight than one that can be shown unchanged since its date.

Step 1: Confirm which AI systems are in scope

You cannot evidence governance over systems you have not listed, so the first thing an auditor asks for is the population. That means a current register of the AI systems in use, including the models embedded inside vendor software and the tools a team stood up without a formal decision. The systems missing from the register are the ones that undermine the audit, because a finding the auditor makes that you did not is the finding that costs you credibility. Reconcile the register against what is actually running before the audit starts, not during it.

Step 2: Show each system has an owner and a recorded risk decision

ISO/IEC 42001 asks top management to assign responsibility and authority for the AI management system, and the NIST AI RMF GOVERN function makes accountability for AI risk an explicit outcome. Naming a person answerable for each system is how that accountability holds in practice. An auditor reads ownership as the first control, because a system no one owns has no one to answer for its risk. For each material system, show the named owner and the risk decision behind it. Someone assessed the system, understood what could go wrong and decided the use was acceptable on the record. An assessment with no owner is a document, not a control, and an auditor treats it that way.

Step 3: Assemble the records behind each control

For each control you claim, hold the record that shows it operated. For an AI management system that usually means the register of systems in scope, the impact assessment behind each high-risk use, the owner and the accepted risk, the approval line and action log for an autonomous agent, the monitoring history for a deployed model, the incidents with what was done about them and the reviews over each system’s life. The records have to be the ones captured when the work happened and attributed to the people who did it. A record reconstructed the week before the audit answers a weaker question than one that was there all along.

Step 4: Check each record would count as evidence

Go through the set with the four qualities in hand. AI makes reliability harder than it looks, because these systems change without a formal release. An assessment written before a model was retrained or a vendor updated it does not evidence the system as it runs now, so the date on each record matters. A document in a shared drive proves its contents but not its history. Records whose capture date and integrity a third party can confirm move the burden off your word and onto the record, which is what an auditor is looking for.

Evidence that fails usually fails in one of these ways.

• A policy offered as evidence that a control operates. The policy is the intention, not the operation.
• A single artefact standing in for a whole practice. One assessment does not evidence a programme of work.
• A record with no owner. An unattributed document answers a weaker question than a signed one.
• A record that cannot be dated or shown unchanged. The auditor cannot rely on what it cannot verify.
• Evidence assembled the week before the audit. Reconstruction reads as reconstruction.

Step 5: Close the loop on incidents and reviews

ISO/IEC 42001 and the NIST AI RMF both expect AI systems to be monitored once deployed and reviewed over their life, not only at the start. An auditor checks that the reviews actually happened and that when something went wrong you recorded it and acted. Show the review schedule for each material system, the dates reviews were done and the incidents you logged with what you did about them. A framework with no incidents recorded reads as one that is not watching, not one that has never had a problem.

What auditors most often find

• Shadow systems running in the business that never reached the register.
• Owners named on paper but no risk decision recorded behind them.
• Assessments done once at procurement and never revisited as the system changed.
• Records that exist but cannot be dated or shown unchanged since capture.
• Incidents handled in a corridor conversation and never written down.

In Australia

The Guidance for AI Adoption (AI6), published by the National AI Centre in the Department of Industry, Science and Resources in October 2025, replaced the Voluntary AI Safety Standard. Its six essential practices, starting with deciding who is accountable and running through testing and monitoring, line up with the steps above. Regulated entities may also face sector expectations, such as APRA’s prudential standards for operational risk. Read the AI6 framework page for how the practices map.

Where Aicura fits

An audit samples from your list of AI systems, so the list has to be right. Aicura’s discovery reports AI in use from the sources you connect, which brings the undeclared tools onto the register before an auditor finds them. Each system’s record keeps a new version every time it changes, and risk assessments run against a fixed version of that record with the owner on the record. At Pro, impact assessments do the same, and the Cadence view, Aicura’s screen for recurring work, shows how long ago each piece of work on each system was last done, so a stale review shows up in your own preparation first.

At Pro, the Evidence Vault keeps sealed snapshots of governance events such as approvals, publishes, ownership changes and incident steps as they happen, and an evidence pack gathers them for the auditor. A snapshot shows when something was captured and that it has not changed since, not that its content was right. Aicura does not run the audit, judge sufficiency or decide whether you pass. The platform overview sets out what each area does.

Sources

• ISO/IEC 42001:2023, AI management system, International Organization for Standardization, iso.org
• ISO 19011:2018, Guidelines for auditing management systems, International Organization for Standardization, iso.org
• AI Risk Management Framework (AI RMF 1.0), National Institute of Standards and Technology, nist.gov/itl/ai-risk-management-framework
• Guidance for AI Adoption, National AI Centre, Department of Industry, Science and Resources, industry.gov.au

A note on this page

This guide is general information on how to prepare for an AI governance audit against current international guidance, with an Australian note. It is not legal advice and it does not tell you whether particular evidence, a particular system or a particular framework meets a particular obligation. For how the guidance applies to your organisation, read the primary sources above and take your own professional advice.