A customer asks whether your AI assistant is governed. A procurement team sends a questionnaire. A board member asks what standard the program follows. ISO/IEC 42001 is often the answer people reach for, because it is the one AI governance standard an outside auditor can certify against.

The standard itself is sold, not published openly. This guide relies on what standards and accreditation bodies publish in the open: the abstract, the contents pages, catalogue entries and accreditation notices. Where we go beyond those sources, we say it is our reading.

Key takeaways
  • ISO/IEC 42001:2023 specifies requirements for an AI management system. The unit is the organization, so a bank that buys an assistant is as much in scope as the vendor that built it.
  • The standard is voluntary unless legislation, a contract or another binding instrument requires it. It is certifiable by third-party certification bodies, so ask which accreditation body stands behind any certificate you are shown.
  • Requirement clauses 4 to 10 cover context, leadership, planning, support, operation, performance evaluation and improvement. Annex A lists reference control objectives and controls.
  • The standard names two assessments, AI risk assessment and AI system impact assessment, and both need a written record.
  • Neither the abstract nor the contents pages names generative AI. Non-determinism, third-party models, change after launch and agents are where that generality leaves work for a GenAI team.

What ISO/IEC 42001 is

ISO/IEC 42001:2023, "Information technology - Artificial intelligence - Management system," was published on December 18, 2023, as edition 1.0 (IEC webstore). It comes from ISO/IEC JTC 1/SC 42, the joint committee on artificial intelligence.

Its abstract says it "specifies the requirements and provides guidance for establishing, implementing, maintaining and continually improving" an AI management system "within the context of an organization" (IEC webstore). The Standards Council of Canada puts the idea in plainer words: "Instead of assessing a single tool, it requires policy, roles, risk controls and continual improvement for the AI an organization builds, buys or deploys" (IAF Outlook).

So the unit is the organization, not the chatbot. The abstract says the standard is for an organization "providing or using products or services that utilize AI systems," and applies "regardless of size, type and nature." A bank that buys an assistant from a vendor is as much in scope as the vendor that built it.

As of November 11, 2026, the IEC webstore lists no amendment or corrigendum to the 2023 edition. Three related documents have appeared since:

  • ISO/IEC 42005:2025, AI system impact assessment. It "provides guidance for organizations performing artificial intelligence (AI) system impact assessments," including "how and when to perform such assessments" (EVS catalogue, dated May 28, 2025). It is a guidance document.
  • ISO/IEC 42006:2025, requirements for bodies that audit and certify AI management systems (EVS catalogue, dated July 7, 2025). It is written for auditors, not for the organizations they audit.
  • EN ISO/IEC 42001:2026, the European adoption. The Estonian standards body lists EVS-EN ISO/IEC 42001:2026 as valid from April 1, 2026, with "(ISO/IEC 42001:2023)" in its title (EVS catalogue).

Voluntary, certifiable, and who certifies

Voluntary. A standard is not a law. CEN and CENELEC state that "the use of standards is generally voluntary: businesses and other organisations may choose whether to apply them unless legislation, a contract or another binding instrument requires otherwise" (CEN-CENELEC FAQ). The Estonian standards body says the same, and adds that a state can make a standard obligatory by a legal act (EVS FAQ). The contract route matters. A customer can require the standard even where no regulator does.

Certifiable. ISO/IEC 42006 describes certification of an AI management system as "a third-party conformity assessment activity" (EVS catalogue). An independent body audits the organization's management system against ISO/IEC 42001 and issues a certificate if it conforms.

Who certifies. Certification bodies do. National accreditation bodies assess those certification bodies. ISO/IEC 42006 "does not replace ISO/IEC 17021-1," the general standard for management system certification bodies. "Instead, it supplements that standard with AI-specific requirements for auditing and certification processes" (EVS news). The Standards Council of Canada, for example, runs an AI Management Systems Accreditation Program and assesses new applicants against the published ISO/IEC 42006 (IAF Outlook).

Accreditation is the part buyers skip. The same article warns about low-cost, non-accredited offers: "Without accreditation there is no independent check of the certifier's competence or impartiality." If a vendor shows you an ISO/IEC 42001 certificate, ask which body issued it and which accreditation body stands behind that body.

The clause structure

The contents pages show seven requirement clauses, numbered 4 to 10, after the scope, references and definitions. The descriptions below are ours, built from the clause and sub-clause titles.

  • 4 Context of the organization. Understand the organization, its context and its interested parties, then set the scope of the AI management system (4.1 to 4.4).
  • 5 Leadership. Leadership and commitment, an AI policy, and assigned roles, responsibilities and authorities (5.1 to 5.3).
  • 6 Planning. Actions to address risks and opportunities, which include AI risk assessment, AI risk treatment and AI system impact assessment (6.1.2 to 6.1.4). Also AI objectives (6.2) and planning of changes (6.3).
  • 7 Support. Resources, competence, awareness, communication and documented information (7.1 to 7.5).
  • 8 Operation. Operational planning and control, then the same three activities again: AI risk assessment, AI risk treatment and AI system impact assessment (8.1 to 8.4).
  • 9 Performance evaluation. Monitoring, measurement, analysis and evaluation, an internal audit program, and management review (9.1 to 9.3).
  • 10 Improvement. Continual improvement, and nonconformity and corrective action (10.1 and 10.2).

Four annexes follow. Annex A, "Reference control objectives and controls," and Annex B, "Implementation guidance for AI controls," are marked normative. Annex C, on potential AI-related organizational objectives and risk sources, and Annex D, on use across domains or sectors, are informative (contents pages).

The Annex A control areas

The contents pages give the annex titles but not the controls. A crosswalk hosted by NIST lists them by name. It was contributed by Microsoft, and it is keyed to the final draft of the standard, so numbering can differ from the published text (NIST crosswalk list). It cites each control by its Annex B number. Grouped by number, the controls it lists fall into nine areas (crosswalk). The area labels are ours.

  • Policy (B.2). "AI policy," "Alignment with other organizational policies," "Review of the AI policy."
  • Internal organization (B.3). "AI roles and responsibilities," "Reporting of concerns."
  • Resources (B.4). "Resource documentation," plus data, tooling, system and computing, and human resources.
  • Impact assessment (B.5). "AI system impact assessment process," "Documentation of AI system impact assessments," and impacts on individuals, groups and society.
  • AI system life cycle (B.6). Objectives and processes for responsible development, then "AI system verification and validation," "AI system deployment," "AI system operation and monitoring," "AI system technical documentation" and "AI system recording of event logs."
  • Data (B.7). "Acquisition of data," "Quality of data for AI systems," "Data provenance," "Data preparation."
  • Information for interested parties (B.8). "System documentation and information for users," "External reporting," "Communication of incidents."
  • Use of AI systems (B.9). "Processes for responsible use of AI systems," "Objectives for responsible use of AI system," "Intended use of the AI system."
  • Third parties and customers (B.10). "Allocating responsibilities," "Suppliers," "Customers."

Treat this list as a map, not as the control text. An audit runs against the published standard, so buy it before you write a control statement.

AI risk assessment and AI system impact assessment

The standard names two assessments, and each appears twice. Clause 6 plans them (6.1.2 and 6.1.4). Clause 8 carries them out (8.2 and 8.4) (contents pages). Risk treatment sits between them in both clauses.

The two are not the same exercise. The impact assessment looks outward. ISO/IEC 42005 describes it as an assessment "for individuals and societies that can be affected by an AI system and its foreseeable applications," and says the process "can be integrated into an organization's AI risk management and AI management system" (EVS catalogue). The crosswalk lists matching controls for impact on "individuals and groups of individuals" and for "societal impacts" (crosswalk).

For a customer-facing assistant, our reading is practical. The risk assessment asks what could go wrong with this system. The impact assessment asks who is on the other end of the conversation, and what a wrong answer costs them. Both need a written record, because "Documentation of AI system impact assessments" is a listed control and documented information is its own sub-clause (7.5).

What generative AI makes harder

The abstract speaks of "AI systems" in general, and neither it nor the contents pages names generative AI. The Standards Council of Canada reports the same lesson from its accreditation pilot: "ISO/IEC 42001 is intentionally general" (IAF Outlook). The rest of this section is our inference about where that generality leaves work for a GenAI team.

Non-determinism. "AI system verification and validation" is a listed control. A scoring model can be validated against a holdout set. An assistant gives a slightly different answer each time, so one passing run proves little. The evidence has to be a set of scenarios, run repeatedly, with thresholds agreed in advance and results kept by version.

Third-party models. Most assistants sit on a foundation model the organization did not build. The crosswalk lists controls for "Suppliers" and "Allocating responsibilities." The Canadian pilot "underscored the need for AI inventories and data lineage, management of third-party and model-as-a-service risk" (IAF Outlook). In practice that means knowing which model and version each application calls, and what the vendor will and will not tell you.

Change after launch. Clause 6.3 is "Planning of changes." Clause 9.1 is monitoring and measurement. "AI system operation and monitoring" and "AI system recording of event logs" are listed controls. The same pilot stressed "disciplined change management." A vendor model update, a new prompt or a new retrieval source can change behavior with no release on your side. Approval evidence gathered before launch goes stale unless production is monitored and the system is re-evaluated when something changes.

Agents. An agent takes actions, not only writes text. "Intended use of the AI system" is a listed control. For an agent, intended use has to cover which tools it may call and when it must hand over to a person. The open material says nothing specific about agents, so this is ours.

How it relates to other regimes

ISO/IEC 42001 is a management system standard. It does not replace sector rules, and sector rules do not require it. The useful question is where the evidence overlaps.

  • NIST AI RMF. The crosswalk maps the framework's Govern, Map, Measure and Manage subcategories to clauses and controls of the final draft of ISO/IEC 42001. NIST notes that listing a crosswalk "does not imply NIST endorsement" of the mapped resource, "nor does it imply that either resource comprehensively covers the contents of the other" (NIST crosswalk list). See our NIST AI RMF guide.
  • EU AI Act. The standard now exists as a European standard, EN ISO/IEC 42001:2026 (EVS catalogue). That alone does not make it a route to legal conformity. A harmonized standard may provide a presumption of conformity only once its reference "is published in the Official Journal of the European Union" (CEN-CENELEC FAQ). Check the Official Journal before relying on it. See our EU AI Act guide.
  • US banking. SR 26-2 leaves generative and agentic AI out of scope and points banks back to their own risk management and governance practices. A management system is one way to structure those practices. It is not a substitute for supervisory expectations. See our SR 26-2 guide.
  • US insurance. The NAIC AI Model Bulletin expects a written AIS Program covering governance, risk management controls and internal audit. Those themes overlap with clauses 5, 6 and 9. An overlap is not equivalence, and a state regulator will ask its own questions. See our NAIC AI Model Bulletin guide.

How GenGuardX helps

GenGuardX is where business teams test a GenAI application, risk teams approve it against evidence, and both monitor it after launch. The table maps clauses and controls of ISO/IEC 42001 to evidence a GenAI program can produce, and to the GenGuardX capability that records it. The evidence is organized against the standard. The mapping does not certify conformity, and it does not make any organization certified or compliant. A certification body decides that, against the full management system. GenGuardX itself is SOC 2 Type 2 certified. It does not hold an ISO/IEC 42001 certificate.

Clause or control, as titledGenAI evidenceGenGuardX capability
Determining the scope of the AI management system (4.3); "Resource documentation" Every prompt, retrieval source, model and agent, with an owner and a use case. Inventory management registers datasets, models, prompts, RAGs and end-to-end agents.
AI risk assessment and AI risk treatment (6.1.2, 6.1.3, 8.2, 8.3) The risks that apply to each application, how each was measured, and the threshold it had to pass. Standardized evaluations and Risk dashboards & curated reports.
AI system impact assessment (6.1.4, 8.4) Realistic customer scenarios run by the people who know the customer, with the issues they raised. Human Integrated Testing lets business users run scenarios, flag issues and rate responses. Annotation queues route outputs that need a human.
"AI system verification and validation" Versioned test results for groundedness, policy adherence, PII leakage and safe escalation. Standardized evaluations that are versioned and re-runnable, and the Pipeline Builder to compare configurations.
Roles, responsibilities and authorities (5.3); documented information (7.5) Who approved the application, on what evidence, and when. Approval workflows and role-based access in the Governed Workbench, with lineage tracking and approval history.
Monitoring, measurement, analysis and evaluation (9.1); "AI system operation and monitoring" Evidence the application still behaves as approved after vendor updates and prompt changes. Monitoring for drift from the approved version, with alerts on live traffic.
Planning of changes (6.3); nonconformity and corrective action (10.2) Each issue tracked from raised to resolved, and a retest after the fix. A findings database with tracking, and re-evaluation fed by production issues.

GenGuardX does not write your AI policy, run your internal audit or hold your management review. Those belong to the organization. What it gives an auditor is a record that can be followed without a guided tour: what the application is, how it was tested, who approved it, and how you know it still works.

FAQ

Is ISO/IEC 42001 mandatory?

No. CEN and CENELEC state that "the use of standards is generally voluntary" unless "legislation, a contract or another binding instrument requires otherwise" (CEN-CENELEC FAQ). A customer contract or a procurement questionnaire can still ask for it, so check your own obligations.

Can an organization be certified to ISO/IEC 42001?

Yes. ISO/IEC 42006:2025 describes certification of an AI management system as "a third-party conformity assessment activity" (EVS catalogue). Certification bodies carry out the audit, and national accreditation bodies assess those certification bodies. The organization is certified, not an individual AI application.

Does ISO/IEC 42001 cover generative AI?

It covers AI systems in general. The abstract says it applies to any organization "that provides or uses products or services that utilize AI systems" (IEC webstore). The abstract and the contents pages do not name generative AI. Our reading is that a customer-facing assistant or agent is in scope like any other AI system, and that the organization has to work out the tests itself.

Has ISO/IEC 42001 been amended since 2023?

As of November 11, 2026, the IEC webstore lists edition 1.0, published December 18, 2023, and no amendment or corrigendum. Two companion standards followed in 2025: ISO/IEC 42005 on AI system impact assessment and ISO/IEC 42006 on certification bodies. A European adoption, EN ISO/IEC 42001:2026, is listed by the Estonian standards body as valid from April 1, 2026.

Is GenGuardX certified to ISO/IEC 42001?

No. GenGuardX is SOC 2 Type 2 certified and does not hold an ISO/IEC 42001 certificate. Using GenGuardX does not make an organization certified or compliant. It records testing, approval and monitoring evidence that an organization can organize against the standard.

Last reviewed: November 11, 2026. This article summarizes openly published material about a paywalled standard. It is not legal advice and not a substitute for the standard's text.

See how GenGuardX organizes GenAI testing, approval and monitoring evidence against ISO/IEC 42001.