Which ISO/IEC AI Standard Should a Regulated Enterprise Adopt First?

Certifiable requirements versus guidance

The first distinction that matters is between a standard that states requirements and a standard that offers guidance. ISO/IEC 42001 specifies requirements for establishing, implementing, maintaining and continually improving an Artificial Intelligence Management System (AIMS). ISO describes it as a management system standard and applies it to organizations that develop, provide or use AI systems. Certification is performed by independent certification bodies, and ISO notes that adopting the standard delivers value even without certification.

The other standards in this group play different roles. ISO/IEC 23894:2023 is guidance on AI-related risk management and describes processes for implementing and integrating it. ISO/IEC 22989 establishes AI terminology and concepts, and ISO/IEC 23053 provides a framework for describing a generic AI system that uses machine learning. ISO/IEC 38507:2022 gives guidance to members of an organization's governing body on enabling and governing the use of AI. Treat these four as guidance or reference material, not as programs you certify against.

One further item is worth tracking. ISO/IEC 42007, a high-level framework for developing conformity assessment schemes for AI systems, is at draft stage (DIS). It should not be treated as available for certification schemes today.

The main standards at a glance

Use this comparison to decide what each document can and cannot do for your program.

• ISO/IEC 42001: AI management system requirements. Certifiable by independent certification bodies. Applies to organizations that develop, provide or use AI.
• ISO/IEC 23894:2023: Guidance on AI risk management processes. Informs risk assessment and treatment. Treat as guidance.
• ISO/IEC 22989: AI terminology and concepts. Gives teams a shared vocabulary.
• ISO/IEC 23053: Framework for describing a generic AI system using machine learning. Supports consistent system descriptions.
• ISO/IEC 38507:2022: Guidance for governing-body members on enabling and governing AI use. Anchors board-level oversight.
• ISO/IEC 42007 (DIS): Draft framework for conformity assessment schemes for AI systems. Monitor, but do not plan around it yet.

A decision framework for choosing the first standard

Start with two questions. First, do any stakeholders, such as customers, regulators or your board, expect independent assurance of how you manage AI? Second, do you already operate a management system, such as an ISO 27001 program, whose scoping, internal audit and management review practices you can extend? If the answer to either is yes, ISO/IEC 42001 is the logical first standard. It gives you an auditable structure, and ISO describes it as an integrated approach to managing AI projects, from risk assessment to effective treatment of those risks. Risk treatment sits inside the management system rather than in a separate standard.

If you have no need for certification, you can still adopt 42001 as an alignment target. ISO states that the standard delivers value on its own and that certification adds stakeholder confidence. Decide early which path you are on, because it changes scope documentation, audit preparation and the evidence you must retain.

Adopt ISO/IEC 23894 alongside 42001 when your risk function needs a defined method for AI risk. It is guidance, so it shapes how you perform risk assessment and treatment rather than adding auditable requirements. Whether to adopt it before or together with 42001 is a judgment that depends on how mature your existing risk processes are. ISO/IEC 22989 is inexpensive to adopt early because it aligns vocabulary across compliance, engineering and legal teams. ISO/IEC 38507 belongs with whoever owns board-level oversight of AI use.

Fitting standards to regulatory obligations

ISO states that ISO/IEC 42001 does not replace laws or regulations. It provides a management framework that helps organizations meet compliance obligations more effectively. A regulatory obligations register therefore stays necessary, and each legal duty needs its own operational control and record, which you then map to the management system.

The EU AI Act illustrates the point. As of November 2026, Regulation (EU) 2024/1689 is in force. The original high-risk application date cited by the Commission is 2 August 2026. On 19 November 2025 the Commission proposed the Digital Omnibus on AI, which would move application to 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for those embedded in products. The Commission cites delays in standards, national competent authorities and conformity assessment bodies. It also says CEN and CENELEC did not deliver standards for the high-risk requirements by the requested August 2025 timeline, and that standards are voluntary but decisive for legal certainty.

For planning, this means two things. Do not assume that adopting an ISO/IEC standard settles Act compliance, and keep your EU timeline flexible while the proposal and the standards work progress. Treat NIST AI RMF alignment as a further mapping target for the same operational evidence rather than a separate program.

Where runtime enforcement and evidence fit

A management system describes what you intend to do. An auditor then asks whether production AI systems actually behaved that way. Policies on acceptable use, data handling and human oversight carry weight only if something enforces them when a model, agent or tool call executes, and if a durable record shows the decision and its outcome. Controls that are reviewed after the fact can show intent and detect problems, but they cannot show that a prohibited action was prevented.

Which records a certification body or regulator will expect depends on your scope, your risk treatment decisions and the obligations in play, so define them with your auditor rather than assuming a standard list. As a design principle, aim for records that are generated by the control itself, retained for a period you set, and attributable to a specific interaction and policy decision. Evidence assembled manually before an audit is harder to trust and expensive to maintain.

Trussed addresses this operational layer. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions such as allow, redact, block, rate limit and escalate for human approval, applied to individual interactions and tool calls. The audit-ready evidence generated by that enforcement is mapped to frameworks including NIST and the EU AI Act. Trussed does not deliver ISO/IEC certification, which is issued by independent certification bodies, and it does not cover model or dataset provenance.