ISO/IEC 23894: Information Technology — Artificial Intelligence — Risk Management

· About ISO/IEC 23894: Information Technology — Artificial Intelligence — Risk Management

Key Takeaways

  • ISO/IEC 23894 is a voluntary international guidance standard, not a law, and it applies to organizations that develop, produce, deploy, or use AI systems across their lifecycle[2][3].
  • The standard was published in February 2023 and remains current in 2026; no 2025–2026 amendment was identified in the sources reviewed, while regional adoptions such as EN ISO/IEC 23894:2024 and DIN EN ISO/IEC 23894:2025-07 were issued without changing the base ISO text[2][14][5].
  • Organizations using ISO/IEC 23894 are expected to integrate AI risk management into existing enterprise risk processes, identify and analyze AI risks, treat them, and monitor them continuously[2][8].
  • Because ISO/IEC 23894 is a standard and not a statute, it does not set statutory fines or criminal penalties; penalties arise only if a contract, procurement rule, certification scheme, or linked law incorporates it[2][8].
  • The standard is designed to complement ISO/IEC 42001 and other management-system controls, and it aligns well with ISO 27001, NIST CSF 2.0, and ISO 42001 for governance, security, and assurance work[2][8].

What It Is

ISO/IEC 23894:2023, Information technology — Artificial intelligence — Guidance on risk management, is an international guidance standard for managing risks associated with AI systems across the full lifecycle of a product, system, or service[2][3]. It was published on 6 February 2023 and is currently listed as published/current by ISO and national standards bodies that carry the text forward[2][10].

The standard is issued through the ISO/IEC process and is maintained as an international standard rather than a regulator-led code. In practice, implementation is driven by internal governance, procurement, certification, and market expectations rather than direct enforcement by a public authority[2][8].

A visible downstream milestone is European adoption as EN ISO/IEC 23894:2024, approved by CEN on 12 February 2024 and published in 2024 without modification to the ISO text[14]. Another downstream publication is DIN EN ISO/IEC 23894:2025-07, which reflects the European version and shows the document remained current through mid-2025[5].

Who Must Comply

ISO/IEC 23894 is written for organizations that develop, produce, deploy, or use AI-enabled products, systems, or services[8]. It is therefore broad enough to cover model developers, software vendors, integrators, deployers, enterprise users, and operators that rely on AI in business processes[8].

There are no statutory applicability thresholds, territorial triggers, or turnover tests in the standard itself because it is voluntary guidance rather than a legal regime[2][3]. There is also no built-in exemption list in the ISO text as surfaced by the available sources; instead, applicability depends on whether an organization has AI-related risks to manage[2][8].

Extraterritorial reach is a contract and adoption issue, not a legal one: the standard can be required globally by customers, public buyers, auditors, or certification schemes, but it does not itself create cross-border enforcement powers[2][5]. Any sanctions come from the incorporating instrument, not from ISO/IEC 23894 itself[2].

Core Requirements

  1. Integrate AI risk management into the organization’s broader risk framework by embedding AI-specific controls into existing governance, risk, and compliance processes rather than treating AI as a standalone silo[2][8].
  2. Identify AI risks across the lifecycle for design, development, deployment, operation, and decommissioning, including risks linked to data, model behavior, human oversight, and misuse[2][8].
  3. Assess likelihood and impact using a structured method so that AI risks are prioritized consistently with the organization’s appetite and decision rules[2][8].
  4. Treat and mitigate risks through control selection, redesign, safeguards, oversight, and residual-risk acceptance where justified and documented[2][8].
  5. Monitor and review risks continuously because AI systems can change through updates, data drift, context shifts, or new misuse patterns after deployment[2][8].
  6. Document decisions and communicate responsibilities so that risk ownership, escalation paths, and accountability are clear across technical and business teams[2][8].

Deadlines and Penalties

| milestone | date | what applies | |---|---:|---| | ISO/IEC 23894 published | 6 February 2023 | First edition of the international guidance standard was released[2][3]. | | EN ISO/IEC 23894 adopted | 12 February 2024 | CEN approved the European standardization of the ISO text without modification[14]. | | EN ISO/IEC 23894 published | 11 April 2024 | European publication of the standard became available[14]. | | DIN EN ISO/IEC 23894 issued | July 2025 | German national adoption of the European text was published[5]. |

Maximum fines: None are set by ISO/IEC 23894 itself because it is not a legal instrument[2][3]. Other sanctions: The standard does not create direct sanctions, but contractual remedies, procurement disqualification, certification failure, audit findings, or liability exposure may follow if a party is required to comply and does not[2][5]. No 2025–2026 amendment or delay affecting the base ISO publication was identified in the sources reviewed[2][5][14].

How to Comply

  1. Map AI use cases and owners by inventorying every system that develops, embeds, deploys, or materially relies on AI, and assign accountable business and technical owners.
  2. Adopt a common risk taxonomy using ISO/IEC 23894 for AI-specific risks, then map the results into ISO 27001 controls for information security and governance.
  3. Build an AI risk assessment workflow that covers data quality, model robustness, bias, explainability, human oversight, security abuse, and downstream harm.
  4. Tie controls to ISO 42001 if the organization is implementing an AI management system, because ISO/IEC 23894 provides risk-management guidance that fits naturally inside AI governance.
  5. Use NIST CSF 2.0 where security is a major driver by aligning AI threat identification, protection, detection, response, and recovery activities to the framework’s functions.
  6. Set approval and escalation rules for high-risk deployments, including legal, security, privacy, and product sign-off before production release.
  7. Monitor post-deployment behavior with incident triggers, drift checks, misuse reporting, and periodic reassessment after model, data, or context changes.
  8. Test evidence and auditability so that documentation, decisions, residual-risk acceptances, and remediation actions can be shown to customers, auditors, or certifiers.

Related Regulations

  • ISO/IEC 42001 overlaps closely because it is the AI management-system standard; ISO/IEC 23894 supplies the risk-management method, while ISO/IEC 42001 supplies the broader management-system structure[2][8].
  • ISO/IEC 27001 overlaps on governance, access control, supplier management, and incident handling, but it is security-focused rather than AI-risk-specific.
  • NIST AI RMF 1.0 conflicts only in terminology and structure, not substance; many organizations map its govern, map, measure, and manage functions to ISO/IEC 23894 workflows.
  • EU AI Act may impose binding legal duties where AI systems fall into regulated categories, while ISO/IEC 23894 remains a voluntary implementation standard that can help demonstrate structured risk management.
  • ISO 31000 overlaps as a general enterprise risk framework, but ISO/IEC 23894 is narrower and more detailed for AI-specific risks.

FAQ

Does ISO/IEC 23894 apply to companies outside Europe?

Yes. ISO/IEC 23894 is an international standard, so it can be used by organizations anywhere in the world[2][3]. Its application depends on business adoption, contracts, procurement requirements, or certification schemes rather than geography[2].

Is ISO/IEC 23894 mandatory?

No. The standard itself is guidance and does not create legal compulsion or standalone penalties[2][3]. It becomes mandatory only if a contract, policy, regulator, or procurement rule requires it[2].

Does ISO/IEC 23894 have fines or penalties?

No fines are set by the standard itself[2][3]. Any sanctions come indirectly through contracts, certification conditions, audit findings, or another law that incorporates the standard[2][5].

What changed in 2025 or 2026?

The sources reviewed did not identify a 2025 or 2026 amendment to the base ISO/IEC 23894:2023 text. What did appear were national and regional adoptions, including EN ISO/IEC 23894:2024 and DIN EN ISO/IEC 23894:2025-07, which reflect the same underlying ISO edition[14][5].

How does ISO/IEC 23894 relate to ISO/IEC 42001?

ISO/IEC 42001 is the management-system standard for AI, while ISO/IEC 23894 provides the risk-management guidance used inside that system[2][8]. Many organizations use both together: 42001 for governance and 23894 for the practical risk process.

Sources

Put it into practice

More compliance guides