Skip to main content

ISO/IEC 42001: standing up an AI management system without drowning in it

The short answer

ISO/IEC 42001 is the certifiable shell

It is the first international AI management-system standard (published December 2023), voluntary like NIST but, unlike NIST, auditable by an accredited third party who issues a certificate a buyer can trust. It holds the governance you already built and makes it verifiable.

What you will be able to do

  • Explain what an AI management system (AMS) is and how ISO/IEC 42001:2023 structures it, in plain language, without reciting clause numbers.
  • Distinguish a certifiable management-system standard (ISO/IEC 42001) from a voluntary risk method (the NIST AI Risk Management Framework) and from binding law (the EU AI Act), and state the job each one does that the others cannot.
  • Locate your organization inside the standard's Plan-Do-Check-Act structure and its seven management clauses, naming what already exists and what is missing.
  • Scope a right-sized AMS: draw an honest boundary around the systems, sites, and processes the management system will cover, and defend why that boundary is neither too wide to maintain nor too narrow to be meaningful.
  • Apply the thirty-eight Annex A controls against your own AI systems using a risk-based test, deciding for each whether it applies and why.
  • Produce a draft Statement of Applicability that an auditor could read, and connect it to the artifacts you already built (your AI systems inventory, your conformity file, your NIST four-function map).
  • Defend a scoping and control-selection decision against the challenge that certification is a paperwork exercise with no live governance behind it.
  • Distinguish the AI risk assessment (inward, drives control selection) from the AI system impact assessment (outward, on people), and run each for its own purpose.
  • Sequence the certification journey from gap analysis through the operating records an auditor needs (including internal audit and management review) to the two-stage audit and surveillance.
  • Read another organization's ISO/IEC 42001 certificate critically, checking the issuing body's accreditation, the scope, and the currency rather than trusting the logo.

The ideas, one by one

A management system is people and process, not software

Policies, roles, accountability, and records through which you direct and control AI across its life. No tool stands it up for you, and no tool passes the audit for you.

The seven clauses are the same skeleton as ISO 27001 and 9001

Context, Leadership, Planning, Support, Operation, Performance evaluation, Improvement, mapping onto Plan-Do-Check-Act. This is why an AI management system bolts onto management systems many organizations already run.

You do not implement all thirty-eight controls

Annex A is a risk-based menu of 38 controls across 9 objectives. You select what your risk assessment demands and record every include-or-exclude decision, with justifications, in the Statement of Applicability. Implementing everything is the drowning failure.

The Statement of Applicability is the spine

It is the first document an auditor reads. An honest "not applicable" with a one-line reason is stronger than an unevidenced "applicable," and a table of all-applicable-no-justification is a red flag that no risk thinking happened.

The impact assessment is central, not optional

The standard requires assessing impact on individuals, groups, and society, supported by ISO/IEC 42005:2025. For higher-risk systems it is the control an auditor probes hardest, and it reuses the impact work from earlier modules.

Certifiable buys verifiable trust, not legality

An accredited certificate lets a customer verify your governance once through a third party (as Anthropic demonstrated in January 2025). It does not make you EU AI Act compliant and does not guarantee good outcomes; it certifies you run a system, scoped to exactly what the certificate says.

Right-sizing is the skill in the title

Scope to a coherent, maintainable unit; select controls by risk; document only what a stranger needs to run the system. A bounded, honest certificate beats a broad, hollow one, and scope is a truth statement customers can read.

Reuse, do not rewrite

Your inventory, provenance file, evaluation report, and conformity file already satisfy much of the standard. Certification is largely pointing existing artifacts at the clauses they meet, not generating fresh paperwork.

Run all three instruments as layers

The EU AI Act is the legal floor, the NIST AI RMF is the risk engine, and ISO/IEC 42001 is the certifiable shell that holds the engine and produces the records. Choosing one over the others is a false choice.

A certificate certifies a running system, so plan the records

The journey runs from gap analysis, through building only the missing pieces and operating the system to generate records (including at least one internal audit and one management review), to a two-stage audit and ongoing surveillance. A same-month binder has no operating history and fails Stage 2.

Avoiding the standard is a failure too

The mirror of drowning is dodging: governing AI on instinct until a customer demands a certificate you do not have and a contract walks. The skill is to scope the standard to fit, neither drowning in it nor avoiding it.

Nonconformities are normal; closing them is the point

A major nonconformity blocks the certificate until fixed; a minor one is closed through a corrective-action plan. A system that surfaces and closes findings is stronger than one that never looks, and zero findings is more often under-auditing than perfection.

The third-party control carries the weight when you buy rather than build

For an organization that procures AI instead of training it, the Annex A third-party-and-customer-relationships objective is often the most load-bearing control, which is another reminder that right-sizing is about which controls matter for how you actually use AI, not just which systems you scope.

The risk assessment and the impact assessment are different jobs

The risk assessment (inward, drives control selection) tells you which Annex A controls to implement; the impact assessment (outward, on people) tells you what harm your system could do and whether your safeguards suffice. You need both, and calling an inward risk assessment your impact assessment skips the analysis the standard cares about most.

Read a vendor's certificate the way you scope your own

Scope, accreditation, and currency are where the truth lives, not the logo. A certificate scoped to an internal tool, issued by an unaccredited body, or lapsed, tells you little about the product you are buying, and a buyer who checks those three things is doing the same honest work as a certifier who scopes truthfully.

Listen to it

The same lesson, talked through by two hosts as episode 43 of the podcast.

Real cases

These examples show the standard applied to real organizations and decisions, with the reasoning stated explicitly. Where an event is centerpieced by another topic, it is named only in passing and pointed to its owner. Read them for the pattern rather than the names: in each, an organization chose a scope, selected controls by risk, and either reused governance it already had or paid the price of pretending. The examples span sectors and regions on purpose, because the standard is deliberately sector-neutral and jurisdiction-neutral, and the same moves recur whether the organization is a frontier lab, a hospital, a cloud provider, a public agency, or a five-person startup.

Example 1: Anthropic's certification as the anchor (ISO/IEC 42001:2023, certified January 2025). The clearest real example of a right-sized AI management system is a frontier AI lab choosing to be audited against ISO/IEC 42001 and scoping the certificate deliberately. Anthropic announced accredited certification in January 2025, issued by Schellman (an accredited certification body), positioning itself as one of the first frontier labs certified (web-verified, 2026-07-09). Two features are instructive for your own work. First, the certificate is scoped (to defined AI research, development, and services roles and locations), not claimed over "everything the company does," which is what a defensible scope looks like. Second, the company presented certification as building on governance work it already had (a public responsible-scaling policy, published safety research), not as a fresh paperwork project. That is the pattern to copy: certify a bounded scope on top of governance you can already evidence.

Example 2: A cloud provider certifying its AI services for its customers. Major cloud and enterprise-software providers have pursued ISO/IEC 42001 certification for their AI offerings. Their customers, especially in regulated sectors, ask "how is the AI you are selling me governed." A certificate scoped to the AI services line lets thousands of customers verify governance once, through an accredited third party, instead of each running its own vendor audit. The management-system value here is leverage: one audited shell answers a question that would otherwise arrive as a thousand separate procurement questionnaires. (The broader story of buyers borrowing teeth on voluntary frameworks is developed in (see Topic 6.2).)

Example 3: A hospital scoping the management system to clinical AI only. A health system running dozens of algorithms could try to certify all of them and drown. The right-sized move is to scope the AMS to clinical decision-support AI, where the impact on patients is highest, and to select Annex A controls accordingly: the impact-assessment control and the human-oversight control are plainly applicable, while a control aimed at, say, generative marketing content might be marked not applicable within this scope. The impact assessment at the center of the standard is the same rigor the external validation of a deployed clinical model demanded. (see Topic 4.6) A narrow, honest clinical scope produces a certificate that means something to patients and regulators; an everything-scope produces a certificate the hospital cannot maintain.

Example 4: A startup that reused its inventory instead of starting over. A small company with two AI features (a support chatbot and a recommendation model) can reach a credible Statement of Applicability in weeks rather than months by pointing existing artifacts at the clauses. Its AI systems inventory (see Topic 0.2) supplies the Clause 4 context and the Annex A resource documentation; its data provenance file (see Topic 2.6) supplies the data control; its evaluation and monitoring records (see Topic 4.6) supply Clause 9 performance evaluation. The startup's scope is small and defensible, its documentation is thin but real, and its certificate covers exactly its two features. This is the anti-drowning path: scope tight, reuse artifacts, document only what a stranger needs to run the system.

Example 5: The certificate that covered nothing that mattered. Consider, as a cautionary pattern rather than a named organization, a company that scoped its AI management system to an internal, low-risk tool while marketing "ISO 42001 certified" as if it covered its flagship customer-facing AI. The certificate is real; the scope is a shell game. An informed customer reads the scope statement on the certificate, sees that the flagship product is outside it, and trusts the company less, not more. The lesson is that scope is a truth statement: certifying the wrong boundary and implying a wider one is a governance failure the standard's own transparency (the published scope) is designed to expose.

Example 6: Bolting the AMS onto an existing ISO 27001 system. Organizations that already hold ISO/IEC 27001 for information security frequently find ISO/IEC 42001 far cheaper to reach, because the two share the Harmonized Structure: the same Clause 4 context, Clause 5 leadership, Clause 9 internal audit, and Clause 10 improvement machinery are already running. They add the AI-specific pieces (the impact assessment, the AI lifecycle controls, the data-for-AI controls) rather than building a management system from scratch. This is the practical payoff of the shared skeleton: an AI management system is not a new building, it is a new wing on a building many organizations already own.

Example 7: A public-sector agency and the third-party control. A government agency that does not build AI itself but procures it from vendors finds that the Annex A third-party-and-customer-relationships objective is the most load-bearing control in its whole Statement of Applicability. The agency's real risk is not in a model it trains but in what its suppliers do, so its management system leans heavily on vendor governance: requiring impact-assessment evidence from suppliers, contractual rights to audit, and documented handling of the AI value chain. This is a globally common pattern, not a single-jurisdiction one; an agency in Ottawa, Nairobi, or Seoul that buys rather than builds AI will center the same third-party control. It shows that right-sizing is not only about which systems you scope but about which controls carry the weight for the way your organization actually uses AI.

Example 8: The certification body itself, and why accreditation is the anchor. When Anthropic was certified in January 2025, the certificate was issued by Schellman, a certification body accredited by a national accreditation authority (web-verified, 2026-07-09). That accreditation is the reason a customer can trust the certificate at all: the certification body's own credibility is staked on having audited the system honestly, and the accreditation authority oversees the certification body. The lesson for your own procurement is directional: when a vendor claims ISO/IEC 42001, the first question is who issued the certificate and whether that body is accredited, because an unaccredited "certificate" or a self-declared "aligned" claim carries none of the independent verification that makes the standard worth anything.

Example 9: The 27001 shop that reused its internal-audit machinery. An enterprise that had held ISO/IEC 27001 for years did not stand up a new audit function for its AI management system; it extended the internal-audit program and management-review calendar it already ran for security to cover the new AI-specific controls. The Clause 9 activities that trip up first-time certifiers (internal audit and management review) were already habitual, so the AI certification added only the AI-specific evidence, not a whole new governance rhythm. This is the shared-skeleton payoff seen from the Check phase: the hardest-to-fake parts of a management system, the proof that it is actually operated and reviewed, transfer directly from one standard to the next.

Example 10: The vendor questionnaire that a certificate answered once. A mid-sized AI vendor was fielding dozens of bespoke security-and-governance questionnaires a year from enterprise prospects, each demanding evidence of how its AI was governed. After certifying a scoped AI management system, it could answer the recurring "how is your AI governed" question with a single accredited certificate and its published scope, collapsing weeks of per-deal questionnaire work into one verifiable artifact. The management-system value here is the same leverage a cloud provider gets at larger scale: one audited shell answers, once and verifiably, a question that would otherwise arrive one deal at a time.

Where people go wrong

  • "We have to implement all thirty-eight Annex A controls." No. The standard is risk-based. You run your AI risk assessment, select the controls that treat the risks you found, and record every include-or-exclude decision in the Statement of Applicability with a justification. Implementing all thirty-eight without scoping is the classic drowning failure and, ironically, produces a weaker certificate because none of the selections were reasoned.
  • "ISO/IEC 42001 certification makes us compliant with the EU AI Act." It does not. The standard is voluntary; the Act is binding law. Running the management system generates much of the evidence the Act wants and demonstrates diligence, but a certificate is not legal compliance. Treating the certificate as a shield against the law is a serious category error. (see Topic 5.3)
  • "A management system is software we buy." A management system is people, policies, roles, processes, and records, not a product. Governance tools can help you keep the records, but buying a tool does not stand up a management system, and no tool passes a Stage 2 audit for you: the audit checks that humans are actually operating the system.
  • "The wider the scope, the better the certificate." The opposite is usually true. A too-wide scope is unmaintainable and fails surveillance audits; a bounded, coherent scope that a small team can keep current, and that covers what stakeholders actually care about, is the mark of maturity. Scope is a truth statement, and certifying the wrong boundary while implying a wider one is a governance failure.
  • "Certification is a one-time achievement." Certification runs on a multi-year cycle with surveillance audits between the initial certificate and recertification. If the system stops running, the certificate is at risk. It certifies a living system, not a past event; a certificate with a dead system behind it is the same binder failure the NIST topic warned about. (see Topic 6.2)
  • "ISO/IEC 42001 competes with the NIST AI RMF; we pick one." They operate at different layers. NIST is a voluntary risk method (the engine); ISO/IEC 42001 is a certifiable management system (the shell). Mature organizations run the NIST functions inside the ISO management system. Choosing between them is a false choice.
  • "The impact assessment is optional detail." The AI system impact assessment (impact on individuals, groups, and society) is central to the standard, not a footnote, and ISO/IEC 42005:2025 exists specifically to support it. For higher-risk systems it is the control an auditor will probe hardest. Skipping it guts the standard's purpose.
  • "A Statement of Applicability where everything is 'applicable' is the safe, thorough answer." It is a red flag. It signals that nobody did the risk-based selection the standard requires, and an auditor will ask for evidence on every one and find gaps. An honest "not applicable" with a one-line justification is stronger than an unevidenced "applicable."
  • "We should write all our documentation from scratch for the audit." You should reuse. Your AI systems inventory, data provenance file, evaluation report, and conformity file already satisfy much of the standard. Rewriting them fresh is wasted effort and creates conflicting records. Right-sizing is largely the act of pointing existing artifacts at the clauses they already meet.
  • "An accredited certificate and a self-declared 'ISO 42001 aligned' claim are basically the same." They are not. Accredited certification means an independent, accreditation-overseen body audited and vouched for your system. "Aligned" or "certified" claims with no accredited body behind them are marketing, and a knowledgeable buyer checks who issued the certificate and whether they are accredited.
  • "We can write the whole system and get certified the same month." You cannot. The auditor needs evidence the system is operating, which means it must run for a period and generate real records, including at least one internal audit and one management review. A system with no operating history has nothing for a Stage 2 audit to examine, no matter how polished the documents look.
  • "An internal audit and a management review are optional overhead." They are the Check phase of the standard and among the most common reasons a well-documented system fails Stage 2. The internal audit is your own check that the system works before the external auditor arrives; the management review is leadership formally examining performance and deciding what to improve. Skipping them signals a system that was written, not run.
  • "The number of Annex A controls tells you how big the job is." It does not; your risk assessment and scope do. Thirty-eight is the size of the menu, not the size of your meal. A tightly-scoped organization may implement a modest subset with strong evidence and hold a stronger certificate than one that implemented far more controls thinly across an unmaintainable scope.
  • "The AI risk assessment and the AI system impact assessment are one document." They are related but distinct. The risk assessment looks inward at risks to the organization and drives which controls you select; the impact assessment looks outward at effects on individuals, groups, and society. Collapsing them into one inward-facing document skips the outward-facing analysis the standard cares about most for higher-risk systems.
  • "A nonconformity at audit means we failed." Nonconformities are the normal output of audits. A major one must be fixed before the certificate issues; a minor one is handled through a corrective-action plan. What matters is whether you close them, and a system that never surfaces a nonconformity is more likely under-audited than perfect.
  • "Certification proves our AI is safe and unbiased." It does not. The standard certifies that you run a system for managing AI risk within a defined scope, not that any particular model is safe or fair. Marketing a certificate as a safety or fairness guarantee overclaims and will not survive an informed customer or a regulator who reads what the certificate actually attests.
  • "Once we pick the controls, the Statement of Applicability never changes." It is a living document. As your systems, risks, and scope change, controls move in and out and evidence updates, and the surveillance audits expect to see it maintained. A Statement of Applicability frozen at the initial certificate is a sign the management system stopped running.

Questions people ask

What is AI management system (AMS, sometimes AIMS)?
The set of policies, roles, responsibilities, processes, and records through which an organization directs and controls its development and use of AI across the system's whole life. It is organizational machinery, not software, and it is the concern ISO/IEC 42001 certifies.
What is ISO/IEC 42001:2023?
The first international standard specifying requirements for an AI management system, published jointly by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) in December 2023. It is voluntary but certifiable by an accredited third party. More on ISO/IEC 42001:2023
What is certifiable / accredited certification?
The property that an independent certification body, itself overseen by a national accreditation authority, can audit an organization against the standard and issue a certificate. Accreditation is what makes the certificate trustworthy rather than a self-applied label.
What is Harmonized Structure (formerly High-Level Structure or Annex SL)?
The common seven-clause skeleton (Context, Leadership, Planning, Support, Operation, Performance evaluation, Improvement) shared by modern ISO management-system standards, which is why ISO/IEC 42001 bolts onto systems like ISO/IEC 27001 and ISO 9001.
What is plan-Do-Check-Act (PDCA)?
The continuous improvement cycle underlying the management-system clauses: Plan (Context, Leadership, Planning, Support), Do (Operation), Check (Performance evaluation), Act (Improvement). The same never-finishing loop the NIST framework describes as an operating system. (see Topic 6.2) More on Plan-Do-Check-Act (PDCA)

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →