ISO/IEC 42001: running an AI management system

0:00 ISO/IEC 42001

NIST gives you a way of thinking about AI risk. ISO slash IEC forty-two thousand and one gives you a way of running it, week after week, in a form an auditor can certify. For enterprise and public-sector buyers, a certificate or at least clear alignment is becoming a real differentiator. In this lesson, you'll learn what an AI management system is, the clause structure, how Annex A controls and the Statement of Applicability work, and whether a small business should certify.

0:36 Why a management system

Why does a management system matter when you already have policies? Because policies describe intentions, and management systems make them happen repeatedly: someone owns them, they're reviewed on schedule, problems are tracked and fixed, and improvements are recorded. For buyers, especially governments and large enterprises in the Gulf and Europe, a certificate or credible alignment with forty-two thousand and one is increasingly a shortcut for trust in tenders.

1:06 What it is

ISO slash IEC forty-two thousand and one, published in December twenty twenty-three, is the first international management system standard for AI. Think of it as the AI sibling of ISO twenty-seven thousand and one for information security. It sets requirements for establishing, running, and continually improving an AI management system. It applies to anyone who provides, develops or uses AI, of any size. And unlike NIST, you can be certified against it by accredited bodies. It doesn't decide which AI is acceptable. It makes sure you have a reliable system for deciding.

1:46 Analogy: a kitchen with inspections

An analogy: if the NIST framework is a checklist, ISO forty-two thousand and one is the operating rhythm of a well-run kitchen with regular inspections. There's a written policy, named roles, a routine for assessing risks, training records, procedures people follow, regular internal checks, a management meeting that reviews how things are going, and a habit of fixing problems at the root. An auditor doesn't just read the manual. They watch the kitchen work and check the records match reality.

2:21 The family

It sits in a family. Twenty-two nine eight nine defines AI concepts and terms. Twenty-three eight nine four gives guidance on AI risk management. Forty-two thousand and five, published in twenty twenty-five, guides AI system impact assessments. And forty-two thousand and six, also twenty twenty-five, sets requirements for the bodies that audit and certify AI management systems. Separately, European standards bodies are writing harmonized standards for the EU AI Act. Forty-two thousand and one isn't one of those, but it builds much of the organizational muscle you'll need.

2:59 Clauses 4 to 10

The standard follows the common structure of ISO management systems, clauses four to ten. Context: understand your situation, stakeholders, scope and AI roles. Leadership: top management commitment, an AI policy, and clear roles. Planning: AI risk assessment, risk treatment, AI system impact assessment and objectives. Support: resources, competence, awareness and document control. Operation: actually running the processes. Performance evaluation: monitoring, internal audit and management review. And improvement: fixing nonconformities and getting better.

3:30 Annex A and the SoA

Annex A is where the controls live, grouped into themes: AI policies, internal organization, resources, impact assessment, the AI system life cycle, data for AI, information for interested parties, use of AI, and third-party and customer relationships. You choose which controls apply based on your risk assessment, and record your choices and justifications in a Statement of Applicability. That's what makes the system proportionate. A fifteen-person agency and a bank will have very different statements.

4:03 Worked example: Karachi agency scope

Here's a scope statement for a forty-person digital agency in Karachi with UK clients. The AI management system covers using third-party generative AI tools and building AI automations for clients, for teams in Karachi and remote, serving clients in Pakistan, the UAE and the UK. It names their roles: deployer of third-party tools, and provider of client-specific chatbots and workflows. It states what's out of scope, and lists the interested parties: clients, end users, staff, vendors, regulators and the certification body. A tight, honest scope makes everything else easier.

4:42 Certify or align?

Should a small business certify? Certify if enterprise or public-sector clients ask for it, if you sell AI products or AI-heavy services, or if you already hold twenty-seven thousand and one and can integrate. Align without certifying if you mainly need internal discipline or your budget is tight. Alignment still gives most of the benefit, and you can certify later. If you already run an information security management system, reuse its document control, audits, management review and supplier management, and add the AI-specific pieces.

5:19 Example 2: Abu Dhabi consultancy

A simple example. A twelve-person consultancy in Abu Dhabi uses AI lightly: a chat assistant, transcription and an image tool. No client has asked for certification. So they align rather than certify: a one-page scope, a signed AI policy, a short risk register, a Statement of Applicability with most Annex A controls marked applicable at a light level, a yearly internal review and a management meeting twice a year. Total effort: a few days. If a government client later asks for certification, they're already most of the way there.

5:58 Common mistakes

Three mistakes. Writing policies that describe an ideal organization rather than the real one. Auditors test what you actually do, and gaps become nonconformities. Ignoring the impact assessment requirement, when ISO forty-two thousand and one expects AI system impact assessments, and forty-two thousand and five shows how to do them. And treating certification as the finish line. Surveillance audits check continual improvement, so the system has to keep running.

6:28 Watch me do it: Statement of Applicability

Watch me do it. I open a new tab called Statement of Applicability. Columns: Control theme, Control summary, Applicable, Justification, Implementation link, Owner. First row: policies related to AI. Applicable: yes. Justification: required baseline. Implementation: link to our signed AI policy. Owner: CEO. Internal organization: yes, several teams use AI, link to the RACI. Resources for AI systems: yes, link to the register itself. Assessing impacts: yes, because client chatbots affect end users, link to the impact assessment template. AI system life cycle: yes, we build automations for clients. Data for AI systems: yes, client data feeds chatbots. Information for interested parties: yes, link to the disclosure procedure. Use of AI systems: yes, everyone uses generative AI. Third-party relationships: yes, heavy vendor reliance, link to the due diligence tab. Now the key part: the scope statement above the table. I write one paragraph naming our offices, our roles as deployer and provider of client automations, what's out of scope, and who the interested parties are. That paragraph plus this table is the backbone of an aligned management system.

7:45 Recap and next step

Recap. ISO slash IEC forty-two thousand and one is a certifiable management system for AI, built on clauses four to ten, with Annex A controls selected through a Statement of Applicability. It fits alongside NIST and the EU AI Act. Watch for two traps: policies describing an imaginary company, and treating certification as the finish line. Your next step: write a one-paragraph scope statement for your organization, then fill in the Statement of Applicability starter from the lesson text.