AI risk management is the coordinated work of identifying, assessing and treating the risks that come from developing or using AI systems. ISO/IEC 23894 is the international guidance for it. It takes the familiar approach of ISO 31000 and adds what is different about AI. This guide explains the standard and how to apply it.
Key points
- ISO/IEC 23894:2023 gives guidance on AI risk management. It is not a certifiable standard.
- It follows ISO 31000: principles, framework and process.
- Its annexes list AI-related objectives and risk sources and map the process to the AI system life cycle.
- ISO/IEC 42001 requires AI risk assessment and treatment; ISO/IEC 23894 shows how to do them.
- AI risk changes after deployment, so monitoring matters as much as the first assessment.
Why AI risk needs its own guidance
Most organizations already manage risk. AI adds features that ordinary methods miss:
- Behaviour is learned from data, so poor or unrepresentative data becomes a risk in itself.
- Results can be hard to explain, which complicates accountability.
- Performance can drift as the world changes, without anyone changing the system.
- Systems may act with a high level of automation, leaving less room for human correction.
- Harm can fall on people outside the organization, who never agreed to the risk.
How ISO/IEC 23894 is organized
The standard mirrors ISO 31000 clause by clause and comments on each from an AI point of view:
- Principles: risk management should be integrated, structured, customized, inclusive, dynamic, based on the best available information, attentive to human and cultural factors, and continually improved.
- Framework: leadership and commitment, integration into the organization, design, implementation, evaluation and improvement.
- Process: communication and consultation, scope and context, risk assessment, risk treatment, monitoring and review, recording and reporting.
Three annexes do much of the practical work. Annex A lists objectives to consider, such as accountability, fairness, privacy, robustness, safety, security, transparency and the quality of training data. Annex B lists risk sources. Annex C maps risk management activities to the stages of the AI system life cycle.
AI risk sources to look for
Annex B gives a useful checklist for workshops:
- complexity of the environment the system operates in
- lack of transparency and explainability
- the level of automation
- risk sources related to machine learning, such as data quality and continuous learning
- system hardware issues
- system life cycle issues, from design flaws to poor decommissioning
- technology readiness, when a method is not yet mature
AI risk management step by step
- Set the context. Decide which AI systems and uses are in scope, who the stakeholders are, and what criteria you will use to judge risk.
- Identify risks. For each system, work through the objectives and risk sources. Include risks to people as well as to the business; the AI impact assessment feeds this step.
- Analyse. Estimate likelihood and consequence. For AI, add how many people are affected and whether the effect can be reversed.
- Evaluate. Compare with your criteria and decide what needs treatment.
- Treat. Options include improving data, adding human review, restricting the use case, testing more, changing supplier, or not going ahead.
- Monitor. Track performance, incidents and drift in operation, and reassess when the model, data or use changes.
- Record and report. Keep the results so that decisions can be explained to management, auditors and regulators.
How it relates to other frameworks
- ISO/IEC 42001 is the management system. It requires AI risk assessment and treatment and points to ISO/IEC 23894 for guidance.
- ISO 31000 is the general risk standard underneath. If your enterprise risk process follows it, AI risk fits inside.
- NIST AI RMF covers similar ground with its Govern, Map, Measure and Manage functions. See ISO/IEC 42001 vs NIST AI RMF.
- EU AI Act requires a risk management system for high-risk AI systems. See ISO/IEC 42001 vs EU AI Act.
Training
Lead AI Risk Manager is the PECB course for professionals who lead this work: identifying, assessing, treating and monitoring AI risk. If you manage risk more broadly, ISO 31000 Risk Manager covers the general method. To build the full management system, see ISO/IEC 42001 Lead Implementer.
Sources: ISO/IEC 23894:2023, Guidance on risk management and ISO 31000:2018 (ISO).
Frequently asked questions
Can an organization be certified to ISO/IEC 23894?
No. It is a guidance standard. Organizations are certified to ISO/IEC 42001, and ISO/IEC 23894 helps them meet its risk requirements.
What is the difference between ISO/IEC 23894 and ISO 31000?
ISO 31000 is general risk management guidance for any organization. ISO/IEC 23894 follows the same structure and adds AI-specific guidance, objectives and risk sources.
Is AI risk management the same as an AI impact assessment?
No. Risk management looks at effects on the organization’s objectives. An impact assessment looks at effects on individuals, groups and society. The impact assessment is an input to the risk assessment.
Who should own AI risk?
A named owner for each AI system, supported by a central function that sets the method. In many organizations that function sits with risk, compliance or information security.
