What is in an ISO 42001 Annex A checklist?
This checklist turns the control topics of ISO/IEC 42001 into checklist items, and adds one row for each of clauses 4 to 10. Each item says what evidence satisfies it and who usually produces it: your organization, your engineering team, or a system such as Arcjet that records it. Use it before an audit.
ISO/IEC 42001:2023 is the international standard for an AI management system. Its text is copyrighted, and the full list of Annex A controls, with their numbers and titles, is only in the standard. So this checklist doesn't use Annex A numbers. It groups items into topics of our own, in our own words: AI policy, roles, resources, impact assessment, the life cycle, data, information for others, use, and third parties. Map each item to the matching Annex A control in your copy of the standard, which you can buy from iso.org. For what the standard is and how certification works, read the ISO/IEC 42001 guide first.
How do you run an ISO 42001 gap analysis with this checklist?
To run a gap analysis, go through each row and record its status, the evidence you hold, and the owner of any gap. An auditor checks two things: that a process exists, and that records show it ran. A row is done only when you can show both.
Follow these steps:
- Confirm the scope of your AI management system: which AI systems, teams, and locations it covers. An AI system is an application or agent together with the model providers and tools it uses.
- Work through the clauses table first. Clauses 4 to 10 hold the requirements a certification audit checks, and which controls you need follows from your risk assessment.
- Work through the control topics table, and map each row to the Annex A controls in your copy of the standard. Mark each item as done, partial, not started or excluded.
- For each excluded item, write down why.
- For each partial or not-started item, name an owner and a date.
- Collect the evidence for done items in one place, so the auditor sees it with the item it supports.
The Who column uses three labels. You is your organization: leadership, risk, compliance, legal or the AI governance lead. Eng is your engineering team. Arcjet means Arcjet records the evidence automatically from what it observes and enforces. "None" in the Arcjet evidence column means Arcjet supplies nothing for that item.
What do clauses 4 to 10 need before you start on Annex A?
Clauses 4 to 10 define the management system itself, and all of them are your organization's work. No product performs them. Each row in the following table summarizes the general purpose of one clause in our own words; read the AI-specific requirements in your copy.
| Ref | Item | Who | Arcjet evidence |
|---|---|---|---|
| Clause 4 | The context, interested parties and scope are defined | You | None |
| Clause 5 | Top management commits, approves an AI policy and assigns roles | You | None |
| Clause 6 | AI risks are assessed and treated, AI objectives are set, and you record which controls you apply and why | You | None |
| Clause 7 | Resources, competence, awareness, communication and documents are in place | You | None |
| Clause 8 | The planned processes and risk treatments run | You | None |
| Clause 9 | Monitoring, internal audit and management review happen | You | None |
| Clause 10 | Nonconformities are corrected and the system improves | You | None |
Clauses 8 and 9 ask you to show that processes ran and that you monitor results. Records from the control topic rows in the next table can serve as part of that evidence, but the processes, audits, and reviews are yours.
What should you check for each ISO 42001 control topic?
The following table breaks each control topic into checklist items, in our own words, with the evidence Arcjet records where it has any. A provider decision is a team administrator's choice to approve or prohibit an AI provider, or a client of it such as the ChatGPT app versus the OpenAI API.
| Topic | Item (our wording) | Who | Arcjet evidence |
|---|---|---|---|
| AI policy | An AI policy is documented and approved | You | None. Guard policies are runtime rules, not policy documents |
| AI policy | The AI policy aligns with your other policies, such as security and privacy | You | None |
| AI policy | The AI policy is reviewed at planned intervals | You | None |
| Roles | AI roles and responsibilities are assigned | You, Arcjet | Owner per AI system, actor and provider; ownership lapses when the owner leaves the team |
| Roles | People can report concerns about AI systems | You | None. Concern reporting is yours |
| Resources | The resources each AI system depends on are documented | Eng, Arcjet | Providers, Model Context Protocol (MCP) servers and tools per AI system, from observation |
| Resources | Data, compute and human resources are documented | Eng, You | None. Arcjet doesn't record data sets, compute or staffing |
| Impact assessment | An impact assessment process exists | You | None |
| Impact assessment | Impacts on individuals, groups and society are assessed and recorded | You | Risk tier recorded per AI system. The assessment is yours |
| Life cycle | Objectives and requirements for each AI system are documented | Eng | Documented purpose per AI system |
| Life cycle | Design, development, verification and validation are documented | Eng | None |
| Life cycle | Changes to controls are tested before release | Eng, Arcjet | A rule must pass a test before it goes live; dry run results |
| Life cycle | Changes to controls are approved by a second person | Eng, Arcjet | Two-person publish for Guard policies, with policy revision history |
| Life cycle | Systems are monitored in operation, and events are logged | Eng, Arcjet | Guard decision records, captures, opt-in long-term retention |
| Life cycle | Systems are retired in a controlled way | Eng, Arcjet | Retired actors, with keys revoked and the record kept |
| Data | Data for AI systems is managed for quality, provenance and preparation | Eng | None |
| Data | The data classes each system handles are known | Eng, Arcjet | Declared versus observed data classes, from sensitive-information findings |
| Information for others | Users and other interested parties get the documentation and information they need | You | None |
| Information for others | Adverse impacts and incidents are communicated to those affected | You | None |
| Use | Processes for responsible use are defined | You | None |
| Use | Each system is used within its intended purpose | Eng, Arcjet | Purpose linked to the Guard policies that enforce it; purpose drift notices to the owner |
| Use | Consequential actions have human oversight | Eng, Arcjet | Hold decisions routed to the system's owner, with the outcome recorded |
| Third parties | Responsibilities are allocated between you, suppliers and customers | You | None |
| Third parties | Third-party AI providers are approved or prohibited | You, Arcjet | Provider decisions at provider or client level |
| Third parties | Provider decisions are enforced and use after a prohibition is recorded | Arcjet | Enforcement on managed devices; violations attributed to a person |
| Third parties | Suppliers are assessed | You | None. Supplier assessments are yours |
The pattern in the table is consistent. Topics about documents and judgment, such as AI policy, impact assessment and information for others, are yours. Topics about what running systems use and do, such as resources, the life cycle, use and third parties, can be recorded by the controls that enforce them. Roles and data are split: people assign owners and decide what data a system should handle, and a system can keep those records current and compare them with what it observes.
How does this checklist feed your record of applied controls?
Certification needs a record of which Annex A controls you apply, which you exclude, and why. The checklist gives you its raw material: for each item, the Annex A controls it maps to in your copy of the standard, whether it applies, its status, and the evidence behind it.
Keep the two linked. When an auditor reads that a control is applied, they ask for the evidence. If the evidence for that row is a record from a running system, attach the export rather than a description of it. Arcjet's evidence export is a JSON or CSV file that maps its records to ISO/IEC 42001, which you can attach directly or push to Vanta or Drata. For the steps, see how to export AI governance evidence.
What Arcjet does not cover
Arcjet doesn't cover clauses 4 to 10, which is every row in the clauses table. It isn't a management system, and it doesn't write your AI policy, assess impacts, run internal audits or hold management reviews.
Among the control topics, Arcjet supplies no evidence for the AI policy, for impact assessment beyond a recorded risk tier, or for information for others. It also supplies nothing for concern reporting, data quality and provenance, design and validation records, responsible-use processes, or supplier assessments and contracts.
The inventory covers only what Arcjet observes: managed macOS and Windows devices with Discovery enrolled, actors that report through the SDK, and coding agents with hooks installed. An AI tool used only on an unmanaged device doesn't appear. The evidence page states inventory coverage as the share of healthy enrolled devices, so you can show an auditor how complete the inventory is.
Where Arcjet fits
For the rows marked Arcjet, the evidence comes from the control itself, not from a survey about the control. Discovery records which AI providers people use on managed macOS and Windows devices and enforces provider decisions on managed devices. The SDK and coding-agent hooks report the agents you build and run. Guard decides tool calls, prompts, and model use, records each decision, and requires a rule to pass a test before it goes live. Sensitive-information detection finds personally identifiable information (PII), secrets and protected health information (PHI), which is how Arcjet compares declared data classes with observed ones.
Owner, purpose, risk tier and review date sit on each AI system. In the Console, only team administrators change them, every change is written to the WorkOS audit log, and an owner who leaves the team reads as no owner. The evidence page maps all of it to ISO/IEC 42001, lists every requirement as supported or not covered, and rolls up every standard project in your team.
For the same exercise against the US framework, see the NIST AI RMF checklist. For a framework-neutral list of controls that count toward ISO/IEC 42001, the NIST AI RMF and the EU AI Act at once, see the AI governance checklist.
Frequently asked questions
Why doesn't this checklist use Annex A numbers?
The Annex A control numbers and titles are in the copyrighted standard. This checklist groups items into topics of our own instead: AI policy, roles, resources, impact assessment, the life cycle, data, information for others, use, and third parties. Map each item to the matching Annex A control in your copy.
Can I exclude Annex A controls?
Decide which Annex A controls apply from your risk assessment, and record why you exclude any. Check the exact rules for exclusions in your copy of the standard and with your certification body.
Does Arcjet make you ISO 42001 compliant?
No. Clauses 4 to 10 and most documentation and judgment controls are your organization's work. Arcjet supplies records for some control topics, and its evidence page lists every requirement as supported or not covered.
Which ISO 42001 control topics can be evidenced from running systems?
Resources, the life cycle, use, and third parties produce records that controls can keep, such as the providers and tools each AI system uses, policy tests and approvals, decision records, and provider decisions. Parts of roles and data can be recorded too. The AI policy, impact assessment, and information for others need people.
Does this checklist quote ISO/IEC 42001?
No. Every item and topic is our own wording, and only the clause numbers come from the standard. Check each item against your licensed copy, which you can buy from iso.org or a national standards body.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Get allow, deny, and redact on agent actions before the side effect.