← Blog

ISO 42001 Annex A Controls (Full Mapping)

Parminder Singh
Parminder Singh··15 min read
Summarize with AI

Full walk-through of ISO 42001 Annex A: 38 controls, 9 areas, deployment layers, and auditor-accepted evidence artifacts.

Compliance & Regulationiso-42001annex-aaimsai-management-systemai-compliancecertification
ISO 42001 Annex A Controls: The 38 AI Management Controls and Where Each One Lands in the Deployment

ISO 42001 Annex A lists 38 controls across nine areas (A.2 through A.10) that an organization implementing an AI Management System (AIMS) has to consider. The Statement of Applicability records implementation status, justified exclusions, and incomplete work with a target date. The auditor samples applicable entries at the certification audit and the surveillance audits that follow.

I want to walk through the nine areas and their control sets. I will also identify the deployment layer and the evidence artifact a certification-body auditor can inspect.

One thing to settle before the list, because it trips up every first-time reader of the standard: every control number begins at the second decimal position. The policy area starts at A.2.2 and the internal-organization area starts at A.3.2. The first decimal position in each area holds the control objective, the statement of what the area is for, rather than an implementable requirement. No standalone entry occupies the first policy-area position. The lifecycle area then splits into two sub-areas, A.6.1 and A.6.2, so its controls carry four-level numbers.

TL;DR

Annex A has 38 controls across nine areas. A Statement of Applicability ties each applicable control to a justification, implementation status, and evidence reference. Gateway records can evidence the routed HTTP AI traffic controls, while data lineage, impact assessments, and design documentation remain with their owning teams.

The full control set

The list below is a deployment map. It does not assert that a technical gateway satisfies every listed control. A control can need policy, people, or lifecycle evidence in addition to runtime records.

A.2 and A.3 deployment map

These policy and organizational controls establish ownership and the review record an auditor expects to see.

  • A.2.2, AI policy: policy layer.
  • A.2.3, alignment with other organizational policies: policy layer.
  • A.2.4, review of the AI policy: policy layer.
  • A.3.2, AI roles and responsibilities: organization.
  • A.3.3, reporting of concerns: organization.

A.4 and A.5 deployment map

These entries cover the people, systems, data, and impact-assessment evidence behind the AI management system.

  • A.4.2, resource documentation: organization.
  • A.4.3, data resources: data layer.
  • A.4.4, tooling resources: application layer.
  • A.4.5, system and computing resources: infrastructure.
  • A.4.6, human resources: organization.
  • A.5.2, AI system impact assessment process: policy layer.
  • A.5.3, documentation of AI system impact assessments: policy layer.
  • A.5.4, assessing AI system impact on individuals or groups: policy layer.
  • A.5.5, assessing societal impacts of AI systems: policy layer.

A.6 deployment map

The lifecycle controls attach to planning, delivery, operation, and the evidence generated as the system runs.

  • A.6.1.2, objectives for responsible development: policy layer.
  • A.6.1.3, responsible design and development processes: application layer.
  • A.6.2.2, AI system requirements and specification: application layer.
  • A.6.2.3, design and development documentation: application layer.
  • A.6.2.4, verification and validation: application layer.
  • A.6.2.5, AI system deployment: application layer.
  • A.6.2.6, AI system operation and monitoring: gateway layer.
  • A.6.2.7, AI system technical documentation: application layer.
  • A.6.2.8, AI system recording of event logs: gateway layer.

A.7 and A.8 deployment map

Data-handling work and information supplied to users, regulators, and other interested parties sit in these areas.

  • A.7.2, data for development and enhancement: data layer.
  • A.7.3, acquisition of data: data layer.
  • A.7.4, quality of data for AI systems: data layer.
  • A.7.5, data provenance: data layer.
  • A.7.6, data preparation: data layer.
  • A.8.2, system documentation and user information: application layer.
  • A.8.3, external reporting: organization.
  • A.8.4, communication of incidents: organization.
  • A.8.5, information for interested parties: organization.

A.9 and A.10 deployment map

The final controls deal with responsible use, intended use, suppliers, and customer relationships.

  • A.9.2, processes for responsible use: gateway layer.
  • A.9.3, objectives for responsible use: policy layer.
  • A.9.4, intended use of the AI system: gateway layer.
  • A.10.2, allocating responsibilities: organization.
  • A.10.3, suppliers: model-provider layer.
  • A.10.4, customers: organization.

Nine controls of the 38 sit in A.6, which makes the life cycle area larger than any other by a wide margin. Runtime records, rather than documents, satisfy operation and monitoring, event logging, and intended-use monitoring. Organizations buying models from an external provider often have the least evidence for that subset.

A.2: Policies related to AI

A.2 contains three controls covering the AI policy documentation.

A.2.2 (AI policy) requires the organization to maintain an AI policy that describes the objectives and principles for the responsible development, deployment, and use of AI systems. The evidence is the policy document, signed by the accountable executive, with a review date within the past 12 months.

A.2.3 (Alignment with other organizational policies) requires the AI policy to align with the organization's other policies (information security, privacy, ethics, compliance). The evidence is a cross-reference matrix showing the alignment.

A.2.4 (Review of the AI policy) requires the policy to be reviewed at planned intervals. The evidence is the review record with the date, the participants, and the outcomes.

The policy-related controls operate at the policy layer. The policy document is the primary artifact, and the enforcement layer (see A.8.3 and A.8.4) produces the operational evidence that connects the policy to the AI system's operation.

A.3: Internal organization

A.3 contains two controls covering roles and responsibilities.

A.3.2 (AI roles and responsibilities) requires the organization to define, allocate, and document responsibilities for AI-related activities. The RACI matrix identifying the accountable policy head, system manager, risk manager, and Human Oversight Assignee is the primary evidence.

A.3.3 (Reporting of concerns) requires the organization to establish a mechanism for reporting concerns about the organization's AI systems. The evidence is the reporting channel (a defined process), the reports received in the period, and the responses provided.

A.4: Resources for AI systems

A.4 contains five controls covering the resources needed to operate AI systems responsibly.

A.4.2 (Resource documentation) requires the organization to document the resources allocated to AI-related activities. The evidence is the resource inventory covering people, technology, and data.

A.4.3 (Data resources) requires the organization to document the data resources used across the AI system lifecycle. The evidence is the data inventory with data source, purpose, and processing agreements.

A.4.4 (Tooling resources) requires the organization to document the tooling used to develop and operate AI systems. The evidence is the tool inventory.

A.4.5 (System and computing resources) requires the organization to document the compute infrastructure. The evidence is the infrastructure inventory.

A.4.6 (Human resources) requires the organization to document the human resources involved in the AI lifecycle. The evidence is the role catalog with competence requirements.

A.4 stops at A.4.6. The standard's resource-commitment requirement sits in its main body, so an auditor reviews it even when an Annex A entry is marked inapplicable.

A.5: Assessing impacts of AI systems

A.5 contains four controls covering the AI impact assessment.

A.5.2 (AI system impact assessment process) requires the organization to establish a process for assessing the potential consequences for individuals, groups, and society of AI systems. The evidence is the process document plus the completed assessments for each AI system in scope.

A.5.3 (Documentation of AI system impact assessments) requires the organization to document the assessments produced. The evidence is the assessment records.

A.5.4 (Assessing AI system impact on individuals or groups of individuals) requires the assessment to include the impact on individuals. The evidence is the impact analysis on affected persons.

A.5.5 (Assessing societal impacts of AI systems) requires the assessment to include the societal impact where relevant.

A.6: AI system life cycle

A.6 contains nine controls covering the AI system lifecycle from planning through decommissioning. Its management-guidance sub-area holds two controls, while the lifecycle-stage sub-area holds seven.

A.6.1.2 (Objectives for responsible development of AI system) requires the objectives for responsible development to be defined and carried into every lifecycle stage. The evidence is the written objectives, dated before the build started.

A.6.1.3 (Processes for responsible design and development of AI systems) requires processes for responsible design and development. The evidence is the AI development lifecycle documentation.

A.6.2.2 (AI system requirements and specification) requires the AI system's requirements to be documented. The evidence is the requirements document.

A.6.2.3 (Documentation of AI system design and development) requires the design and development to be documented. The evidence is the design records.

A.6.2.4 (AI system verification and validation) requires the AI system to be verified and validated. The evidence is the verification and validation records including the test results.

A.6.2.5 (AI system deployment) requires the deployment to be planned and controlled. The evidence is the deployment plan and the deployment records.

A.6.2.6 (AI system operation and monitoring) requires the operation to be monitored. The evidence is the monitoring records: the audit log the gateway produces plus the review records showing a named person acted on what the monitoring surfaced.

A.6.2.7 (AI system technical documentation) requires the technical documentation to be maintained. The evidence is the technical documentation.

A.6.2.8 (AI system recording of event logs) requires the AI system to record events automatically across the lifecycle. The evidence is the event log itself: the per-decision audit records the gateway produces. This is the control most often over-claimed. Application-level logging that records "user submitted a prompt" satisfies nobody once the auditor asks which policy was in force and which identity the call carried. The record integrity properties matter as much as the coverage.

A.7: Data for AI systems

A.7 contains five controls covering the data lifecycle for AI systems.

A.7.2 (Data for development and enhancement of AI system) requires the data used for AI development to be documented and appropriate. The evidence is the training data documentation.

A.7.3 (Acquisition of data) requires the acquisition of data to be documented and lawful. The evidence is the data sourcing records and the legal basis analysis.

A.7.4 (Quality of data for AI systems) requires the data quality to be assessed. The evidence is the data quality assessment records.

A.7.5 (Data provenance) requires the provenance of the data to be documented. The evidence is the provenance records including the transformations applied.

A.7.6 (Data preparation) requires the data preparation to be documented. The evidence is the preparation records.

A.8: Information for interested parties of AI systems

A.8 contains four controls covering the information the organization provides to interested parties.

A.8.2 (System documentation and information for users) requires the organization to provide documentation and information to users. The evidence is the user documentation.

A.8.3 (External reporting) requires the organization to have a process for external reporting where required. The evidence is the reporting records for the applicable regulatory obligations (EU AI Act Article 26.4 incident reports, HIPAA breach notifications, SEC 8-K filings).

A.8.4 (Communication of incidents) requires the organization to communicate incidents. The evidence is the incident communication records.

A.8.5 (Information for interested parties) requires the organization to provide information for interested parties. It is the widest control in A.8 and the one auditors most often find under-evidenced, because the other three name a specific audience and this one does not.

Start by writing down who the interested parties actually are. The standard's interested-party requirement already made you identify them, and A.8.5 is where that list has to produce artifacts. For most enterprise AI deployments the list resolves to six groups: the individuals whose data the system processes, the customers whose contracts reference AI processing, the regulators with jurisdiction, the workforce whose work the system affects, the investors or board receiving AI risk reporting, and the model providers sitting upstream.

Each group needs a named artifact and a named owner. Data subjects get the privacy notice section covering AI processing and the mechanism for exercising rights against it. Customers get the AI disclosure in the contract or trust centre. Regulators get whatever the applicable regime specifies. Employees get the internal AI usage policy and the disclosure of any AI involvement in decisions affecting them. The board gets the periodic AI risk report. Suppliers get the requirements you flow down to them under A.10.3.

An auditor looks for current information, a named maintainer, and a real route by which each group can reach it. A privacy notice with one AI paragraph that nobody has revised since 2023 fails the control despite the document's existence. Bring the change log with the document.

A.8.5 also carries a trap worth naming. Where the information you owe an interested party is a statement about what the AI system did in a specific case, the artifact is a record extract rather than a policy document, and it has to be reproducible on demand months later. That is a retention and query problem, not a documentation problem, and organizations discover it at the first data subject request rather than at the audit.

A.9: Use of AI systems

A.9 contains three controls covering the use of AI systems.

A.9.2 (Processes for responsible use of AI systems) requires processes for responsible use to be defined. The AI Usage Policy plus the enforcement mechanism are the evidence.

A.9.3 (Objectives for responsible use of AI system) requires responsible use objectives to be defined.

A.9.4 (Intended use of the AI system) requires a defined and monitored use case for each AI system. The evidence is a use statement plus monitoring records showing that actual use stays within it.

A.10: Third-party and customer relationships

A.10 contains three controls covering third-party and customer relationships.

A.10.2 (Allocating responsibilities) requires responsibilities in third-party relationships to be allocated. The evidence is the third-party agreements including the AI-specific clauses.

A.10.3 (Suppliers) requires supplier relationships to include AI considerations. The evidence is the supplier evaluation records for AI providers (OpenAI, Anthropic, Google, cloud AI providers).

A.10.4 (Customers) requires customer relationships to address AI considerations. The evidence is the customer-facing terms and disclosures for AI-related processing.

The Statement of Applicability

The Statement of Applicability (SoA) is the primary Annex A artifact the certification body reviews. Each entry names the control and title, its applicability status, any exclusion justification, its implementation status, the evidence reference, and a target date for incomplete work.

The auditor uses the SoA to plan the audit sample. Applicable, implemented entries get sampled. An exclusion receives scrutiny for its justification. A planned item receives scrutiny for its completion target.

ISO 42006 and who is allowed to audit you

Annex A tells you what to implement. ISO/IEC 42006:2025, published on 7 July 2025, governs the body doing the auditing and changed the market more than most implementation teams noticed.

Before it existed, certification bodies audited AIMS against ISO 42001 under the generic management-system rules in ISO/IEC 17021-1, with no AI-specific competence floor and no AI-specific audit-duration table. The newer standard adds both, then specifies personnel competence, liability rules, certification-document content, and the body's access to the organization's documentation. Its own Annex A sets how audit time is calculated for an AIMS.

The practical consequence is a longer audit than an equivalent ISO 27001 audit at the same headcount, because the duration table accounts for AI-system count and complexity rather than employee count alone. A certificate issued by a body without ISO 42006 accreditation also carries less weight with a customer's procurement team, so ask a prospective certification body for its accreditation status and the name of the granting accreditation body before signing.

Worth knowing alongside this: the climate change amendment that ISO applied across the Annex SL management system standards in 2024 lands in Clauses 4.1 and 4.2 of ISO 42001, not in Annex A. The organization has to determine whether climate change is a relevant issue and whether interested parties have climate-related requirements. No control in the 38 changes, but the auditor will ask.

Maintaining Annex A controls as the AI system changes

The SoA needs a change trigger, not an annual calendar reminder alone. A new model endpoint, a retrieval-data source, an agent role, a sensitive-data class, or a supplier renewal can alter which Annex A controls apply and which evidence proves them. Put the trigger in the same change record that approves the deployment, then assign an owner to update the affected SoA rows before the route reaches production.

For a model-route change, review A.4.4 and A.4.5 for the tooling and computing inventory, A.6.2.2 through A.6.2.6 for requirements, validation, deployment, and monitoring, A.8.2 for user information, and A.10.3 when a provider or processor changes. The answer may be that an existing control and artifact still apply. Record that assessment with the release identifier, reviewer, date, and evidence location rather than leaving the prior row to imply continuity.

Runtime evidence needs its own review. A quarterly control check can sample one permitted request and one denied request from a production route, confirm that each carries the expected principal, policy version, classification, decision, and timestamp, then link the sample to the A.6.2.6 and A.6.2.8 evidence references. A printed SoA beside an unattended dashboard is a familiar audit-room sight. It is weak evidence when the dashboard route changed three releases ago.

My view is that a spreadsheet becomes a liability once it merely lists controls. Treat the SoA as a living index of decisions and proof, with a change record behind every affected row.

What Annex A does not buy you under the EU AI Act

The assumption I run into most often is that an ISO 42001 certificate discharges the EU AI Act's quality management system obligation under Article 17. It does not, and the gap is structural rather than temporary.

Presumption of conformity under Article 40 attaches only to harmonised standards whose references have been published in the Official Journal of the European Union, and only for the requirements those standards cover. CEN-CENELEC JTC 21 is the committee writing them. It reviewed ISO/IEC 42001 for the Article 17 role and concluded the standard's goals and definitions were not aligned with what the AI Act requires of a quality management system, so it wrote a dedicated European standard instead: EN 18286, "Quality management system for EU AI Act regulatory purposes." As of June 2026 EN 18286 has reached the formal vote stage, the furthest advanced of any JTC 21 deliverable, and no JTC 21 deliverable has been cited in the Official Journal. Zero presumption of conformity is available today from any of them.

An ISO 42001 certificate is useful evidence of a functioning management system. It still leaves structural work, because its document set does not map one to one to the AI Act. If you are building toward both, treat the Annex A control set as the operational base and expect a separate mapping exercise against the AI Act's own articles, starting with the high-risk classification question that determines whether Article 17 applies to you at all.

DeepInspect

The DeepInspect gateway produces evidence artifacts for several Annex A controls in a single record series. A.6.2.6 (operation and monitoring) and A.6.2.8 (recording of event logs) use the gateway's per-decision audit records. A.8.3 (external reporting) can use the record excerpts a CISO prepares for regulatory reporting. The incident-communication control can use the runbook's incident records. A.9.2 uses policy-enforcement events. Classifier verdicts and policy events support A.9.4 intended-use monitoring.

The gateway's identity binding satisfies the identification evidence A.3.2 needs. Its tamper-evident storage satisfies the record integrity properties the auditor tests for.

The boundary is worth stating plainly, because an SoA that overstates coverage is worse than one that admits a gap. DeepInspect sits on HTTP traffic between authenticated users or agents and the LLM. It evidences what happened on that path: which principal called which model, under which policy version, with what classification on the payload, and what the gateway decided. Training-data lineage, impact assessments, and design records stay with the teams that produce them. Treating data-provenance evidence as satisfied by a runtime log is the kind of claim that turns a surveillance audit into a nonconformity.

If you are building out the Annex A control set for an ISO 42001 certification or a Statement of Applicability, a request-level evidence record can support the controls that apply to routed AI traffic.

Let's talk today.

Frequently asked questions

How many of the 38 Annex A controls typically apply to an enterprise AI deployment?

Many deployments mark 30 to 36 controls as applicable. The development-data and acquisition entries may be marked inapplicable when the organization does not train or fine-tune its own models and relies entirely on external providers. The societal-impact entry may be partially applicable when the AI system's reach is limited. Lifecycle-stage exclusions are harder to justify, because deploying and operating a purchased model is still a lifecycle the standard expects you to control.

Do I need all 38 controls implemented at the initial certification?

An initial certification can include partially implemented entries with target dates. The auditor reviews those dates and progress at surveillance audits. Several entries often remain in partial-implementation status at the certification audit.

How does A.10.3 (suppliers) apply when I use OpenAI or Anthropic?

The organization documents the supplier evaluation for each AI provider. The evaluation covers the provider's security posture (SOC 2 report), the provider's AI-specific documentation (model card, system card), the data processing agreement, and the incident notification obligations. The supplier evaluation record is the primary artifact.

Why is there no control A.2.1, A.3.1 or A.4.1?

The first decimal position in each Annex A area holds the group's objective and statement of purpose, rather than an implementation or evidence requirement. Numbering therefore begins at the second decimal position. A.6 adds a level because it splits into A.6.1 (management guidance for the lifecycle, two controls) and A.6.2 (the lifecycle stages themselves, seven controls). An SoA that lists an A.4.1 row has been built from a template rather than the standard.

What is the difference between Annex A and Annex B?

The reference control set is in Annex A, with 38 titles and short descriptions. Teams use Annex B for longer implementation guidance. Annex C lists potential AI-related organizational objectives and risk sources that inform the risk assessment. Annex D covers applying the AIMS across domains and sectors. The Statement of Applicability references Annex A only. The implementation team works from Annex B, and the risk team works from Annex C.

Do I have to implement all 38 Annex A controls?

No. Annex A is a reference control set, not a mandatory checklist. The standard's risk-treatment requirement tells the organization to determine which controls are necessary and justify each exclusion in the Statement of Applicability. The justification is what gets audited. An exclusion for a development-data control can be accepted when the organization does not train models. An entry with no reasoning is a nonconformity.

Who is allowed to issue an ISO 42001 certificate?

An accredited certification body. ISO/IEC 42006:2025, published 7 July 2025, sets the requirements those bodies must meet, covering auditor competence, audit time calculation, liability, and the content of the certification document. Ask a prospective certification body for its accreditation status against ISO 42006 and the name of the accreditation body that granted it before you sign. A certificate from an unaccredited body is not worthless, and it will be questioned in a customer security review.

How does ISO 42001 Annex A compare to ISO 27001 Annex A?

ISO 27001 Annex A (2022 edition) has 93 information security controls across four themes. ISO 42001 Annex A has 38 AI-specific controls across nine areas. The two Annexes overlap where AI systems are information assets (the ISO 42001 vs ISO 27001 mapping covers this in detail). Most organizations that hold both certifications maintain a single integrated Statement of Applicability that references both Annexes, and schedule the surveillance audits together to reduce the sampling burden on the same evidence set.