For an ISO/IEC 42001 audit, prepare evidence that your AI management system exists and runs: a defined scope, an AI policy, risk and impact assessments, a Statement of Applicability, operating records, an internal audit, a management review and tracked corrective actions. Auditors look for records, not intentions.
ISO/IEC 42001:2023 is described by secondary sources as the first certifiable standard for an AI management system. The ISO catalogue pages returned HTTP 403 when we fetched them, and the standard is a paid document, so the structure below comes from secondary summaries. Check wording and numbering against your licensed copy.
What is the structure auditors work from?
Secondary summaries say the mandatory requirements sit in clauses 4 to 10: context, leadership, planning, support, operation, performance evaluation and improvement. Annex A is a reference catalogue of 38 AI-specific controls in nine objectives, numbered A.2 to A.10, and Annex B gives implementation guidance. You choose applicable controls from your risk and impact assessments and record exclusions with reasons in a Statement of Applicability. Objective A.5 deals with impact assessment.
Which evidence should you collect?
Group it by what the auditor will ask.
Scope and context. The AI systems you build, provide or use, the interested parties, and the boundaries of the management system. An AI inventory with owner, purpose, model source and data sources is the core artifact.
Leadership. A signed AI policy, named roles, and proof management approved resources.
Risk and impact. A documented method, completed assessments per system, and treatment decisions with owners. Include impact on people affected by the AI system's output, not only on your business.
Statement of Applicability. Every control marked applicable or not, with implementation status and an evidence link.
Operation. Records from the AI lifecycle: data sourcing notes, evaluation results, release approvals, monitoring of live behavior, incident handling, supplier checks for third-party models and services.
Performance evaluation. An internal audit report and minutes of a management review.
Improvement. A log of nonconformities and corrective actions with closure dates.
How do audits usually run?
ISO management system certification audits commonly happen in stages, a documentation review followed by an assessment of operation over time. We could not confirm the exact rules for 42001, so ask your certification body how many stages it runs, how long, and how much operating history it wants to see.
How do you survive the audit itself?
- Nominate one person to coordinate requests and keep a shared log of what you provided.
- Give auditors the real thing. Producing a document written the night before is worse than showing a gap with a corrective action open.
- Brief interviewees. Engineers should be able to explain how an AI system gets approved and monitored in their own words.
- Fix findings with root cause, not with a rewritten policy.
What to do
- Build the AI system inventory and mark in or out of scope.
- Write the AI policy and assign roles.
- Run risk and impact assessments for each in-scope system.
- Draft the Statement of Applicability from your own risk results.
- Operate the controls for long enough that records exist.
- Hold the internal audit and the management review before the certification audit.
- Pick a certification body and ask for its stage plan in writing.
Quick check
Is ISO 42001 only for companies that build AI models?
We could not confirm the full applicability text from the paywalled standard. Secondary summaries describe it as applying to organizations that develop, provide or use AI systems, so check the scope clause in your copy.
Do I need all 38 Annex A controls?
Per secondary summaries, you select applicable controls based on your assessments and justify exclusions in the Statement of Applicability.
How Safeguard helps automate this
Much of the operating evidence for AI systems lives in code and dependencies. Safeguard's compliance suite covers frameworks, controls, evidence, continuous monitoring, automated tests and questionnaires, so evidence is collected from connected systems and code instead of by screenshot, and monitoring runs continuously between audits. Technical evidence comes from your real repositories through scans, SBOMs and reachability. Autonomous remediation can open tested fix pull requests for vulnerable dependencies, but it is off by default, every category starts on Manual, and you set which run automatically. The certification body, not Safeguard, decides the outcome. Book a call to talk through your evidence.
Sources
- ISO: ISO/IEC 42001 (HTTP 403 on fetch; standard is paid)
- Secondary web search results on clause and Annex A structure, 18 October 2026; verify against the standard
Last reviewed: 18 October 2026.
