ISO 42001 Explained: Certifying an AI Management System (with a free template)

ISO 42001 (formally ISO/IEC 42001) is the first international standard for artificial intelligence management systems (AIMS). ISO 42001 provides requirements for establishing, implementing, maintaining, and continually improving an AI management system. Organizations can have that management system independently certified.  In this article, you will learn what ISO 42001 requires, how to prepare for certification, and get a free template with every column ready to use at the end.

What ISO 42001 requires you to prove

To pass an ISO/IEC 42001 audit, you need to show a working AI management system (AIMS). This system includes specific policies, dedicated owners, assessed risks, selected controls, and records demonstrating that your AI management system is operational.

You prove that you govern AI usage through a repeatable cycle: Plan–Do–Check–Act . In ISO 42001, Clauses 1 to 3 cover scope, references, and definitions, so we’re not covering them for you. Instead, the auditable requirements are specified in clauses 4 to 10, and Annex A lists the controls you choose from.

Tip: Here, controls are the specific measures your company puts in place to prevent, detect or manage AI-related risks. Basically, the things you do to keep AI usage under control.

If you already hold ISO 27001, the structure will be familiar. Both standards follow ISO's Harmonized Structure, so you can reuse your document control, internal audit and management review processes and add the AI-specific parts on top. 

The eight-tab ISO/IEC 42001 readiness workbook

To prepare for ISO 42001, these eight tabs provide a practical core workbook. Not 25 tabs because a consultant enjoys spreadsheets. Eight simple tabs with information that describes what AI you use, what can go wrong, what you do about it, and how you prove your AI management system works.

Therefore, the workbook is your index of evidence. Policies, procedures, test results and approvals can be posted in your document management system. Basically you can link and reference to them instead of copying raw policy content into Excel.

1. AI inventory

The AI inventory defines what you're governing: every AI system your organization develops, buys, integrates, or uses, including the tool one team signed up for last spring.

Example: Compliance uses a vendor model to score transactions for AML alerts. The owner is the MLRO. Customer transaction data goes in, and an analyst reviews every alert before a report is filed.

2. AI risk register

The risk register lists the specific ways each AI system could fail and what you do about each one. Use one row per risk, not one row per system. A single chatbot can carry five separate risks.

Example: an internal assistant answers staff questions on MiCA and invents an obligation the regulation does not contain. A staff member passes it to a client. The treatment is answered based on source documents, mandatory human review for client-facing advice, and a quarterly test set of known questions.

3. AI impact assessments

The impact assessment covers the effects of an AI system on people, groups and society, separately from the business risk to your company. Clause 6.1.4 requires a defined process for it, and clause 8.4 requires you to carry it out and record the results.

Columns: system ID, intended use, affected parties, potential benefits, potential harms, severity, likelihood, safeguards, consultation needed, assessment owner, approval, review date, evidence link.

Example: an automated KYC check rejects applicants whose identity documents it reads poorly. Record how you test rejection rates by document type and country, who reviews the results, and how a rejected applicant reaches a human.

4. Statement of Applicability (SoA)

The Statement of Applicability (SoA) is the list of all 38 Annex A controls, each marked as applied or excluded, with a justification and evidence. Clause 6.1.3 requires it, and auditors usually read it first.

The controls cover nine areas, from AI policy (A.2) to third-party relationships (A.10). The life cycle area is the largest, with 9 controls.

Columns: control ID and title, applicable (yes/no), justification, implementation status, owner, evidence link, related risk IDs, gap and remediation action, target date, approval.

You choose which controls apply, but every exclusion must be justified by your scope and risk assessment. If your firm only uses vendor models and builds none, you can justify excluding parts of A.6 that way. Writing "N/A, not relevant" seventeen times can't be justified, and it's one of the most common findings at Stage 1.

5. Action tracker

The action tracker assigns an owner to every gap and sets a due date. A gap without either one won't get closed.

Policy gaps, missing assessments, untrained staff, vendor issues, overdue reviews and audit findings all go here. Use one tracker instead of five competing lists.

6 to 8. Suppliers, training and competence, audit and review

The last three tabs record your vendor due diligence, your people's competence, and proof that the system checks itself.

• Suppliers: vendor, contract terms on data use and model changes, security and privacy checks, known model limitations, reassessment date. Most crypto and fintech firms buy their AI from vendors, so for them this is often the busiest tab.
• Training and competence: role, required skills, training completed, date, evidence. Clause 7.2 requires you to keep these records.
• Audit and review: audit plan, findings, management review dates and decisions, follow-up actions linked to the tracker.

Records the auditor will ask for outside the workbook

The workbook points to your evidence. Most of the evidence itself goes in a document system. Expect to show:

• AI policy and governance. Who can approve a new AI system, who owns AI risk, and who can stop a deployment. Top management signs this, and a thumbs-up in Slack won't count as approval.
• Operational procedures. How you test AI before release, monitor it in production, handle incidents, manage changes and retire systems, plus records showing each step was carried out.
• Objectives and KPIs. Measurable targets with owners and a review frequency. "We take responsible AI seriously" isn't an objective. "100% of live AI systems have an impact assessment less than 12 months old" is one.
• Internal audit and management review records. These show the system is running, which the auditor needs to see in addition to the documents.

7 steps to ISO 42001 certification, in order

Build the workbook in this order, because each step feeds the next:

1. Define the scope. Name the legal entities, locations, teams, processes and AI activities included. Everything after this step depends on that boundary.
2. Build the AI inventory. Interview each department; IT's software list will miss tools teams bought on a card and models built into vendor products.
3. Assess risks and impacts. Use one documented method with consistent rating scales, assign owners and record who accepts residual risk.
4. Build the SoA and treatment plan. Map each risk to controls, mark the gaps, put each gap in the action tracker.
5. Implement the controls. Publish policies, change workflows and train staff. If nobody follows a policy, the auditor will find out from interviews.
6. Run the system. Collect records, monitor results, investigate incidents and review changes for a full cycle. Most certification bodies want to see a few months of records before Stage 2.
7. Audit yourself, then book the certification audit. Run an internal audit, fix the findings, hold a management review. Then go to Stage 1 (document review) and Stage 2 (implementation audit).

An independent gap assessment before Stage 1 shows which SoA exclusions and missing records an auditor would challenge while you still have time to fix them. That's the purpose of Hacken's ISO 42001 readiness assessment, which reviews your workbook and evidence against the standard before a certification body does.

What a finished workbook does not give you

A completed workbook isn't an ISO 42001 certificate. Certification is issued by an independent certification body after Stage 1 and Stage 2 audits. ISO/IEC 42006:2025 sets the requirements those bodies must meet. Certificates generally run for three years with annual surveillance audits, so the workbook has to stay current after certification day too.

It also doesn't give you EU AI Act compliance. ISO 42001 covers a large part of the governance the Act expects, including risk management, documentation and human oversight. But it's not a harmonised standard under the Act, so certification doesn't create a presumption of conformity. Use the workbook as the base and map the Act's high-risk obligations onto it separately. For the security side of the same systems, see LLM security frameworks, LLM security risks and AI red teaming.

How to make sure the tab is done?

Add three columns to almost every tab: Owner, Evidence link, Last reviewed. If a row has no owner, nobody answers for it. If it has no evidence, you can't prove it. If nobody knows when it was last reviewed, the auditor will assume it's out of date. Open the template, fill the inventory first, and check which rows already have all three.

FAQ

How many controls are in ISO 42001?

Annex A of ISO/IEC 42001:2023 lists 38 controls in nine areas, from AI policy (A.2) to third-party relationships (A.10). You assess all 38 in your Statement of Applicability and justify every exclusion against your scope and risks.

How long does ISO 42001 certification take?

For a mid-sized organisation starting from zero, six to twelve months is a common range. Organisations that already hold ISO 27001 often move faster, because document control, internal audit and management review are already in place. Certification bodies usually expect a few months of operating records before Stage 2.

Is ISO 42001 mandatory?

ISO/IEC 42001 certification is generally voluntary, although contracts, tenders or sector-specific requirements may make it commercially necessary.