Agile InsightsBriefing
A closed reference volume, three white dice and a pair of headphones on a dark desk, labelled ISO/IEC 23894, Reference Series and Reference 7. A yellow bar along the bottom reads Reference series, number 7.

Reference 7 · Primer · About 5 minutes

ISO/IEC 23894: the AI risk management guidance

The international guidance for managing AI risk: it extends ISO 31000 to AI, and it is guidance, not a certificate.

#rules

Download the MP3 · Leonard, an AI voice, reads a script written from the sources below.

In this episode

01Guidance, not a certificateISO/IEC 23894 is international guidance on managing AI risk, built on ISO 31000. The 2023 edition is current as at October 2026. No auditor will certify you against it, and no law requires it.
02What ordinary risk registers missMost risk processes were built for floods, fraud and supplier failure. A model can drift, be biased in ways nobody planned and be hard to explain. The guidance tells you what to ask.
03Four ideas, and monitoring never stopsYou add AI to your existing risk management, led from the top with named roles. Models change after launch, so monitoring and review never stop. A risk assessment filed and forgotten is the classic failure.
04How it fits with 42001 and NISTISO/IEC 42001 is certifiable and requires you to manage AI risk. This guidance explains how to do that well. The NIST AI Risk Management Framework has the same intent, so choose one as your working method.
05Three steps to get startedFind out whether you already run ISO 31000 style enterprise risk management, and extend it if you do. Build a register of AI systems with named owners. Check your risk process against the guidance and fix the gaps, especially monitoring.
Sources and links (8)
The standard · ISO catalogue page
ISO catalogue page for ISO/IEC 23894:2023 (may need a browser).
Australian adoption · Standards Australia
Standards Australia store entry for AS ISO/IEC 23894:2023, listed as current.
The parent standard · ISO catalogue page
ISO 31000:2018, the general risk management guidelines this builds on.
NIST AI Risk Management Framework · NIST
NIST's page for the framework, noting version 1.0 is being revised.
NIST crosswalk · NIST AI Resource Center
NIST's mapping between ISO/IEC 23894 and the AI RMF.
European adoption · SIS
Swedish standards body entry showing the 2024 European version replacing the 2023 national one.
A plain explainer · Mindgard
Practitioner overview of the guidance and how it pairs with 42001.
Related: AI impact assessment · IEC webstore
IEC webstore entry for ISO/IEC 42005, published 2025.

Written from public sources and checked on 10 October 2026. The standard is paraphrased, not reproduced. Check the current text before relying on it.

How this maps to what we do

The standards and frameworks that describe how to run AI governance as a system, and the practical work each one calls for.

Standard or rule
What an organisation needs
Where we help
ISO/IEC 42001 (and the ISO/IEC 42000 family)Episode: ISO/IEC 42001: the AI management system standard
A working AI management system: policy, roles, a risk process, internal audit and evidence, ready for certification if wanted.
AI Operating Model DesignAI Governance Framework AlignmentISO/IEC 42001 Implementation (with partners)
The four functions (Govern, Map, Measure, Manage) put into practice and cross-walked to the standards you already follow.
AI Governance Framework AlignmentAI Risk & Compliance
ISO/IEC 23894: AI risk management
A risk method that fits AI: identification, assessment, treatment and monitoring across the lifecycle.
AI Risk & Compliance
AI brought into the security and privacy management systems the organisation already runs, instead of a parallel set.
AI SecurityAI Governance Framework AlignmentCISO Advisory (with partners)

These are Agile Insights AI Strategy & Governance offerings. Some are delivered with partners.

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Leonard If your organisation uses AI, someone must own the risk. I S O, I E C, 23894 is the international guidance on how to manage that risk. By the end of this episode, you will know what it is, how it differs from the certifiable AI standard, and what to do first. This is current as at October 2026.

Leonard First, what it is. It is published by the International Organization for Standardization and the International Electrotechnical Commission, through their joint AI committee. It is guidance, not a certification standard. That means no auditor will certify you against it, and no law requires it. It was first published in February 2023, and it builds directly on I S O 31000, the general standard for risk management.

Leonard Now the problem it solves. Most organisations already have a risk process. It was built for floods, fraud and supplier failure. AI behaves differently. A model can drift as the world changes. It can be biased in ways nobody planned. Its reasoning can be hard to explain. Picture a lending model that quietly worsens for one group of customers over six months. Your standard risk register may never ask the right question. This guidance tells you what to ask.

Leonard Who uses it? Risk and compliance leads use it to extend the enterprise risk framework to AI. Chief information and data officers use it when approving a new AI system. Boards meet it as the method behind a paper on AI risk. It also comes up in procurement, when a customer asks how you manage the risks of the AI you supply, and when an auditor asks how risk assessment under the AI management system is actually done.

Leonard What is in it? Four ideas carry it. The first is that it follows the shape of I S O 31000: principles, a framework, and a process. So you add AI to your existing risk management. You do not build a second, separate system.

Leonard The second idea is the framework. Risk management must be led from the top, with named roles, resources and accountability, and joined to how the organisation is governed. Without that, the process is paperwork.

Leonard The third is the process itself, applied to AI. You set the scope, the context and your risk criteria, meaning what level of harm you will accept. You identify risks. You analyse how likely and how serious they are. You evaluate them against your criteria. You treat the ones that matter, and you accept what remains knowingly. Throughout, you consult the people affected.

Leonard The fourth idea is where AI really differs. Monitoring, review, recording and reporting never stop. Models change after launch, so a risk assessment filed and forgotten is the classic failure. The guidance also includes annexes with AI-specific objectives and risk sources, such as fairness, transparency, data quality and the level of automation. Treat those as prompts for your own thinking.

Leonard How does this connect to the rest of the stack? Think of three layers. Data foundations sit underneath, with the data quality standards. The AI management system standard, I S O, I E C, 42001, sits in the middle. It is certifiable, and it requires you to manage AI risk. This guidance explains how to do that well. Above them sit local and regulatory rules, such as the European Union AI Act. The National Institute of Standards and Technology in the United States publishes the AI Risk Management Framework, which has the same intent. NIST publishes a crosswalk between the two, so choose one as your working method rather than running both. Version 1.0 of the NIST framework is currently being revised, according to NIST.

Leonard Getting started takes three steps. One, find out whether you already run I S O 31000 style enterprise risk management. If you do, extend it. Two, build a register of your AI systems, each with a named owner. Three, run a readiness check of your current risk process against the AI risks in the guidance, and fix the gaps, especially monitoring. The primary sources are the I S O catalogue entry, the Standards Australia store entry, and the NIST framework page. Our show notes link to each.

Leonard Now currency, as at October 2026. The 2023 edition is the current international edition. I found no published revision or announced replacement. Standards Australia lists it as current, as an identical adoption, A S I S O, I E C, 23894:2023. Europe has its own 2024 adoption. Related AI standards from the same committee arrived in 2025, including one on AI impact assessment and one on certification bodies for 42001. The standard is copyright, so this episode paraphrases it. Check the current text before relying on any of this.

Leonard To recap, this guidance turns the standard risk process into a practical method for AI, and the key habit is never stopping. If you'd like to know more, sign up for our podcast at agile-insights.com.au.

More on #rules

Reference 15 · PrimerThe quiet obligations: records, data ethics, digital experience and cloudReference 14 · PrimerSecurity for government AI: the PSPF, the ISM and the frontier AI eraReference 13 · PrimerThe Commonwealth AI rulebook: the Policy, the technical standard and the APS AI PlanReference 12 · PrimerThe DTA AI Reference Architecture: one blueprint for government AIEpisode 21 · DiscussionTrust cannot be rushed: how much authority will you give AI?Reference 11 · PrimerAustralian sector rules for AI: what applies to your industry

Every episode tagged #rules

All of the reference series (16) · Follow AI Governance Reference in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series