ISO/IEC 42001 in practice: governing AI before scaling it

Most organizations did not begin their AI journey with a governance plan. They began with pilots: an assistant for the service desk, a model to classify tickets, an agent that reviews alerts. While those pilots are few and isolated, the risk is manageable. When they multiply, connect to production systems and start executing actions, the question changes: who is accountable for what the AI does?

ISO/IEC 42001 is the first international management system standard dedicated to AI. This article explains what it is, why it matters especially when deploying agents, and how to start without turning it into a paperwork exercise.

What ISO/IEC 42001 is

ISO/IEC 42001:2023 was published in December 2023 and specifies the requirements for establishing, implementing, maintaining and improving an artificial intelligence management system (AIMS). It applies to organizations that develop, provide or use AI systems, regardless of size or sector.

Three characteristics make it practical for anyone already working with ISO standards:

• It is certifiable. An organization can be audited by a certification body and obtain a certificate, as with ISO/IEC 27001.
• It shares the ISO management system structure. Context of the organization, leadership, planning, support, operation, performance evaluation and improvement. Anyone already running an ISMS will recognize the logic and can reuse processes such as internal audit, management review and document control.
• It includes reference controls and implementation guidance in its annexes, which the organization selects and justifies based on its risks, much like the ISO/IEC 27001 statement of applicability.

What the standard does not do is tell you which model to use or set technical thresholds. It requires the organization to define how it governs AI, assess its risks and impacts, apply proportionate controls and show that all of this works and improves.

Why it matters if you deploy AI agents

An AI agent does not just answer questions: it interprets a goal, decides on steps and executes actions on other systems (querying a database, opening a ticket, restarting a service, changing a configuration). That changes the risk profile in several ways:

• Access and privileges. The agent needs credentials and permissions. If they are oversized, a model error or a manipulation of its inputs can have real effects.
• Traceability. You need to be able to reconstruct what the agent saw, what it decided and what it did, both to audit and to correct.
• Human oversight. You need to define which actions it may execute alone, which require approval and which are forbidden.
• Third-party dependency. Many agents use models and platforms from external providers, whose changes affect behavior.
• Drift. An AI system's performance can change over time even if nobody touches the code, because the data or the underlying model changes.

ISO/IEC 42001 does not solve these problems on its own, but it forces you to address them systematically: with owners, documented assessments, controls and follow-up. That discipline is what lets you go from ten pilots to a hundred production use cases without losing control.

How it fits with ISO/IEC 27001 and the NIST AI RMF

A frequent question is whether 42001 replaces or duplicates frameworks the organization already uses. In practice, they complement each other:

Framework	Nature	Main focus	How it combines
ISO/IEC 27001	Certifiable management system standard	Information security: confidentiality, integrity and availability.	Covers protection of the data, access and infrastructure that AI systems use.
ISO/IEC 42001	Certifiable management system standard	Responsible use of AI: risks and impacts specific to AI across its lifecycle.	Adds what is specific to AI: impact on people, transparency, data quality, human oversight. Can be integrated with the existing ISMS.
NIST AI RMF 1.0	Voluntary framework, not certifiable	AI risk management, organized into the Govern, Map, Measure and Manage functions.	Provides a detailed method and common language to identify and measure AI risks within the management system.

For those operating in Spain or serving clients in the European Union, Regulation (EU) 2024/1689 on artificial intelligence is also relevant. A management system based on 42001 can help organize the response to its obligations, but it does not automatically equal compliance: the regulation must be analyzed case by case.

A practical roadmap to get started

A sequence we recommend for the first months, designed not to stall AI projects already underway:

1. Inventory of AI systems. List every use case: internal ones, vendor-provided ones and AI features built into software you already use. For each, record purpose, owner, data used, systems accessed, whether it executes actions and who provides it. It is common to find uses nobody had declared.
2. Scope and policy. Decide what the management system will cover (the whole organization or a set of services) and write a short AI policy approved by leadership: principles, permitted and prohibited uses, and responsibilities.
3. Roles. Define who approves new use cases, who owns each system, who assesses risk and who oversees. An AI committee or forum with technology, security, risk, legal and business participation is advisable.
4. Risk and impact assessment. For each system, assess risks to the organization (operational, security, legal) and impacts on individuals and groups (unfair decisions, errors with consequences, lack of transparency). Prioritize by criticality: not every use case needs the same depth.
5. Controls. Select controls proportionate to risk. For agents, the most common are: least privilege on credentials, an explicit list of permitted actions, human approval for high-impact actions, full logging of inputs and actions, testing before every model or instruction change, and a mechanism to stop the agent.
6. Monitoring. Define indicators per system: output quality, human intervention rate, errors, incidents and provider changes. Review them on a fixed cadence.
7. Internal audit and management review. Close the loop as in any management system: verify that what was defined is followed, record nonconformities and adjust.

Evidence worth having from the start

Even if certification is not the immediate goal, working with an auditor in mind helps organize the effort. This is the minimum evidence worth producing from the first cycle:

• An up-to-date AI system inventory, with owner and date of last review.
• An approved and communicated AI policy, with a record of its distribution.
• Risk and impact assessments per system, with the decision made and who approved it.
• A list of controls applied per system and the justification for those that do not apply.
• Activity logs for agents and records of human approvals.
• Minutes of the AI committee's periodic reviews, with actions and owners.
• Internal audit and management review results.

Certification can be a later goal. What matters at the start is having the inventory, the owners and the minimum controls working before you scale.

Common mistakes when implementing the standard

• Treating it as a documentation project. Lengthy policies nobody applies do not reduce risk. A few processes that are actually used are better.
• Leaving it only to security or data science. AI impacts are also legal, ethical and business matters.
• Forgetting third-party AI. AI features included in commercial platforms also belong in the inventory and the assessment.
• Assessing once and never looking again. AI systems change; the risk assessment should be revisited when the model, the data or the use changes.
• Duplicating structures. If an ISMS already exists, integrate both management systems instead of building a parallel one.

This article is informational and does not replace reading the standard or a formal assessment: validate scope and application in your organization with your compliance and legal teams.