ISO/IEC 23894: the ISO standard for AI risk management, explained

ISO/IEC 23894:2023 is a 26-page guidance standard from ISO and the IEC’s joint AI committee, ISO/IEC JTC 1/SC 42 ^[1]. The IEC dates its publication to February 6, 2023 ^[2]. ISO describes it as guidance on how organizations that develop, produce, deploy or use AI-based products, systems and services can manage AI-specific risk, and how to build that risk management into their AI-related activities ^[1].

It is not a standalone method. Its introduction says the document “is intended to be used in connection with ISO 31000:2018”, the general risk management standard. Where it adds to ISO 31000, it cites the matching ISO 31000 clause and then gives the AI-specific guidance ^[3]. In plain terms, ISO/IEC 23894 is ISO 31000 with an AI layer.

The standard follows ISO 31000’s clause structure and adds sub-clauses where AI needs them ^[3]. Clauses 1 to 3 are scope, normative references and terms; ISO/IEC 22989 supplies the AI terms ^[3]. The substance is in clauses 4 to 6 and three informative annexes.

Annex C is the part most teams use first. It gives an example of where risk management activities sit across an AI system’s life cycle ^[3], which helps turn an enterprise risk process into checkpoints engineering teams recognize.

These four are often lumped together as “AI risk frameworks”, but they do different jobs. ISO/IEC 23894 tells you how to run AI risk management. ISO/IEC 42001 requires you to run it inside a certifiable management system. The NIST AI RMF offers a voluntary US vocabulary for the same work. Article 9 of the EU AI Act makes a risk management system a legal duty for providers of high-risk AI systems.

The practical takeaway: one risk process can serve all four. Design it on ISO/IEC 23894, run it inside a 42001 management system, tag each step with its NIST function, and for any high-risk system add Article 9’s extras: fundamental-rights focus, use of post-market monitoring data, pre-market testing against defined thresholds, and attention to people under 18 and other vulnerable groups ^[14].

ISO/IEC 42001 says what an AI management system must do; ISO/IEC 23894 helps with one of the hardest parts of it. 42001 requires an AI risk assessment, AI risk treatment and an AI system impact assessment (clauses 6.1.2 to 6.1.4), and requires them to be carried out at planned intervals or when significant changes occur (clauses 8.2 to 8.4) ^[6][7]. It does not tell you how to design the risk method. That is what 23894 supplies, in ISO 31000 terms ^[3].

• Risk criteria and method. Use 23894’s clause 6.3 (scope, context and criteria) and 6.4 (assessment) to write the method that 42001’s auditors will ask for ^[3][7].
• Risk sources. 42001’s informative Annex C lists potential AI-related objectives and risk sources ^[6]; 23894’s Annexes A and B cover objectives and risk sources in more depth ^[3].
• Impact assessment. For the outward-looking assessment of effects on people and society, ISO points to ISO/IEC 42005:2025, which it says complements 42001 and 23894 ^[9].
• Life cycle. 23894’s Annex C gives a worked mapping of risk activities to life cycle stages ^[3], which you can line up with 42001’s AI system life cycle controls in Annex A.6 ^[18].

For the requirements themselves, see our ISO 42001 guide. If you already run an information security management system, our ISO 42001 vs ISO 27001 guide explains how the AI risk process differs from the ISMS one.

NIST’s AI Resource Center publishes a crosswalk between ISO/IEC 23894 and the AI RMF, revised by INCITS/AI on August 14, 2025; it supersedes NIST’s January 2023 crosswalk, which was based on the final draft ^[12]. NIST notes that listing a crosswalk does not imply endorsement ^[12]. The mapping is many-to-many:

Two differences matter. The AI RMF organizes risk around seven trustworthiness characteristics, from valid and reliable to fair with harmful biases managed ^[11]; 23894 organizes it around ISO 31000’s process ^[3]. And the AI RMF is being revised: NIST says a revised version is in progress, after the July 2025 White House AI Action Plan directed changes ^[11]. For the NIST side in depth, see our NIST AI RMF hub and NIST AI RMF vs ISO 42001.

Not on its own. Article 9 requires providers of high-risk AI systems to establish, implement, document and maintain a risk management system, run as a continuous iterative process across the system’s life cycle ^[14]. ISO/IEC 23894’s process covers much of the method, but Article 9 adds legal specifics: it focuses on risks to health, safety and fundamental rights, requires risks to be eliminated or reduced through design where technically feasible, requires testing against predefined metrics and probabilistic thresholds, and asks providers to consider people under 18 and other vulnerable groups ^[14].

ISO/IEC 23894 is not a harmonised standard under the Act. EVS lists no directive or regulation against its European adoption, EN ISO/IEC 23894:2024 ^[4]. The European standard written for Article 9 is prEN 18228, AI risk management, from CEN-CENELEC JTC 21; it was under public enquiry until the end of July 2026 ^[16]. The first harmonised AI Act standard, EN 18286:2026 on quality management, was cited in the Official Journal on October 14, 2026 ^[17]. After the Digital Omnibus on AI, Regulation (EU) 2026/1744, high-risk obligations apply from December 2, 2027 for Annex III systems and August 2, 2028 for AI in regulated products ^[15].

• Teams implementing ISO/IEC 42001 that need a defensible AI risk method for their auditors ^[7].
• Risk functions that already use ISO 31000 and want to extend enterprise risk management to AI without a second vocabulary ^[3].
• Providers of high-risk AI systems for the EU, as a starting method for an Article 9 risk management system while prEN 18228 is finished ^[14][16].
• Deployers and buyers of third-party AI. The guidance covers organizations that deploy or use AI, not only those that build it ^[1].
• US organizations aligned to the NIST AI RMF that need an international reference, using the INCITS/AI crosswalk ^[12].

You probably don’t need it if your only goal is a certificate and your consultant or platform already supplies an AI risk method mapped to 42001. Even then, it is a cheap way to check that method.

An illustration of clause 6 applied to a hypothetical resume-screening model bought from a vendor. The clause numbers are the standard’s ^[3]; the content is ours.

1. 01

   6.2 Communicate and consult

   Bring in HR, legal, the vendor and a representative of the hiring managers who will use the scores.

2. 02

   6.3 Set scope, context and criteria

   Scope: screening for three job families. Context: employment law in each hiring location. Criteria: what level of score disparity between groups is unacceptable, and who signs off on residual risk.

3. 03

   6.4 Assess the risk

   Identify risks such as biased training data, drift as job requirements change and over-reliance by recruiters. Analyze likelihood and consequence, then evaluate each against the criteria.

4. 04

   6.5 Treat the risk

   Choose options: a bias test before go-live and quarterly, human review of every rejection, contract terms requiring the vendor to report model changes.

5. 05

   6.6 Monitor and review

   Track selection rates by group and recruiter override rates; re-run the assessment when the vendor ships a new model version.

6. 06

   6.7 Record and report

   Log decisions, owners and residual risk in the AI risk register and report to the governance committee.

Hiring tools can also be regulated directly in some places; for one example, see our NYC Local Law 144 page.

ISO and the IEC list ISO/IEC 23894:2023 at CHF 155 for 26 pages ^[1][2]. The European adoption, EVS-EN ISO/IEC 23894:2024, costs €24.80 including tax as a PDF from Estonia’s standards body ^[4]. Our ISO 42001 PDF guide compares sellers and explains what you can read free.

We found no revision. ISO’s catalog shows edition 1 at the published stage, with no revision, amendment or replacement listed, and the IEC Webstore lists no amendment or corrigendum, both as of November 5, 2026 ^[1][2]. The related work to watch is elsewhere: ISO/IEC 42005:2025 for impact assessment ^[9], prEN 18228 in Europe ^[16] and the NIST AI RMF revision ^[11].

Most teams start with a spreadsheet register and move to software when the number of AI systems grows. Our free AI risk register comes with 14 AI risks mapped to ISO/IEC 42001, the NIST AI RMF, the OWASP LLM Top 10 and EU AI Act Article 9, and exports to XLSX, CSV or Markdown. When you outgrow it, compare platforms in our independent ranking of AI risk management software.

• AI risk register →Free, editable register with 14 mapped AI risks.
• AI risk management software →Platforms for AI risk registers, assessments and controls.
• NIST AI RMF →The voluntary US framework, its four functions and the revision.
• NIST AI RMF vs ISO 42001 →When to use each, and how teams use both.
• ISO 42001 vs ISO 27001 →Extending an ISMS into an AI management system.