ISO 23894, risk management built for AI
Apply ISO 31000 to the risks that are special to AI systems.
Official sources
Use these official places to read the text and find the latest version.
What ISO 23894 is
ISO/IEC 23894 is guidance on managing risk for AI. It follows the ISO 31000 process and adds what is different about AI: bias, drift, opaque decisions and data quality.
It is used alongside ISO 42001 to run the risk side of an AI programme.
How ISO 23894 got to where it is today
The milestones that changed what organisations have to do.
ISO/IEC 23894:2023
Published as AI-specific guidance based on ISO 31000.
Dates are for orientation. The official website holds the authoritative record.
Do you need it?
If any of these sound familiar, probably yes.
You build or deploy AI and need a risk method.
You follow ISO 31000 already.
You are preparing for ISO 42001.
Leadership asks about AI risk.
What it asks of you
The requirements fall into six areas.
Principles
Carry ISO 31000 principles over to AI.
Framework
Leadership, integration and resources for AI risk.
Context and scope
Purpose, users, data and environment.
Assessment
Identify, analyse and evaluate AI-specific risks.
Treatment
Choose and plan responses.
Monitoring
Watch for drift and new risks.
ISO/IEC 23894 in 4 parts, 9 topics and 29 guidance points
Guidance on applying the ISO 31000 risk approach to artificial intelligence. It keeps the same principles, framework and process, and adds AI-specific considerations such as data, model behaviour and the AI life cycle. It is guidance, not a certifiable standard.
9 topics · 29 guidance points
PRPrinciples (Clause 4)The ISO 31000 principles, read for AI.1 topic · 8 guidance points
PR-corePrinciples of AI risk managementPrinciples that make AI risk management effective.8 guidance points
- Cl. 4 principle 1
AI risk management is integrated into all organisational activities that involve AI.
What you do
- Add risk gates to AI project approval
- Include AI in enterprise risk reporting
- Cl. 4 principle 2
The approach is structured and comprehensive across AI uses.
What you do
- Use one method for all AI systems
- Cover built and bought AI
- Cl. 4 principle 3
The approach is customised to the organisation's AI uses, context and objectives.
What you do
- Scale controls to your AI risk tiers
- Reflect your sector rules
- Cl. 4 principle 4
The approach is inclusive, involving stakeholders affected by AI.
What you do
- Consult users, staff and affected groups
- Record concerns raised
- Cl. 4 principle 5
The approach is dynamic, as AI systems and their risks change over time.
What you do
- Review risks when models, data or use change
- Watch for drift after launch
- Cl. 4 principle 6
The approach uses the best available information, recognising limits of data and models.
What you do
- Document data quality and test results
- State assumptions and uncertainty
- Cl. 4 principle 7
The approach takes human and cultural factors into account.
What you do
- Consider how people trust or over-rely on AI output
- Train staff on appropriate use
- Cl. 4 principle 8
The approach is continually improved.
What you do
- Learn from AI incidents and near misses
- Update controls from lessons
Evidence to keep
- AI risk policy
- AI risk register
- Management reports
FWFramework (Clause 5)Building AI risk management into governance and operations.3 topics · 10 guidance points
FW-leadLeadership and integrationManagement commitment and embedding AI risk in existing processes.2 guidance points
- Cl. 5.2
Leadership demonstrates commitment and assigns accountability for AI risk.
What you do
- Appoint an executive owner for AI risk
- Approve an AI risk policy
- Cl. 5.3
AI risk management is integrated into governance, strategy and existing processes.
What you do
- Link it to information security, privacy and quality processes
- Include AI in enterprise risk management
Evidence to keep
- AI governance charter
- Leadership statement
- Integration plan
FW-designDesign of the frameworkUnderstanding context and setting out how AI risk will be handled.5 guidance points
- Cl. 5.4.1
Organisational context, including AI objectives, legal duties and stakeholder expectations, is understood.
What you do
- List AI uses, laws and expectations
- Review when strategy changes
- Cl. 5.4.2
Commitment to managing AI risk is set out in a policy or statement.
What you do
- Publish the policy and make it visible
- Align it with AI principles you have adopted
- Cl. 5.4.3
Roles, authorities and accountabilities for AI risk are assigned.
What you do
- Define owners for each AI system
- State who can approve deployment
- Cl. 5.4.4
Resources, including skills in AI, data and ethics, are allocated.
What you do
- Budget for testing and monitoring
- Train or hire specialists
- Cl. 5.4.5
Communication and consultation about AI risk are established, internally and externally.
What you do
- Provide channels for reporting AI concerns
- Set a reporting calendar
Evidence to keep
- Context analysis
- AI risk policy
- Role assignments
- Resource plan
- Communication plan
FW-runImplementation, evaluation and improvementPutting the framework into practice, checking and improving it.3 guidance points
- Cl. 5.5
The framework is implemented with a plan, owners and timing.
What you do
- Roll out in phases starting with high-risk systems
- Brief teams building or buying AI
- Cl. 5.6
The framework's performance is evaluated against its purpose.
What you do
- Define indicators such as incidents and overdue treatments
- Review annually
- Cl. 5.7
The framework is adapted and continually improved.
What you do
- Update after regulation or technology change
- Track actions to closure
Evidence to keep
- Implementation plan
- Framework review
- Improvement tracker
PRCProcess (Clause 6)Applying the risk process through AI systems' life cycles.4 topics · 8 guidance points
PRC-setupCommunication, scope, context and criteriaSetting up the assessment for an AI system.2 guidance points
- Cl. 6.2
Communication and consultation take place throughout the process.
What you do
- Involve developers, operators and affected users
- Record decisions and feedback
- Cl. 6.3
Scope, context and risk criteria are established for the AI system.
What you do
- Define purpose, data, users and boundaries
- Set criteria covering harm to people as well as business impact
Evidence to keep
- Scope statement
- Risk criteria
- Consultation records
PRC-assessRisk assessmentIdentifying, analysing and evaluating AI risks.3 guidance points
- Cl. 6.4.2
AI risks are identified, drawing on sources such as data, model behaviour and use context.
What you do
- Run structured workshops across the life cycle
- Check for bias, security, privacy and safety issues
- Cl. 6.4.3
Risks are analysed for likelihood and consequence, with attention to uncertainty.
What you do
- Rate using test and monitoring evidence
- Note where evidence is weak
- Cl. 6.4.4
Risks are evaluated against criteria to decide on treatment.
What you do
- Rank and escalate those above tolerance
- Document who accepted what
Evidence to keep
- AI risk register
- Impact assessment
- Test results
PRC-treatRisk treatmentSelecting and implementing responses.1 guidance point
- Cl. 6.5
Treatment options are selected, planned and implemented, and residual risk is recorded.
What you do
- Choose controls such as data checks, human review and testing
- Assign owners and dates
Evidence to keep
- Treatment plan
- Control evidence
- Residual risk sign-off
PRC-followMonitoring, review, recording and reportingKeeping AI risk under watch.2 guidance points
- Cl. 6.6
AI systems and their risks are monitored and reviewed over time.
What you do
- Track performance and drift in production
- Reassess after incidents or changes
- Cl. 6.7
The process and results are recorded and reported.
What you do
- Keep a risk log and report to management
- Retain records for audit
Evidence to keep
- Monitoring reports
- Incident records
- Management reports
ANAnnexesSupporting material on objectives, sources of risk and the AI life cycle.1 topic · 3 guidance points
AN-suppInformative annexesReference material for applying the process.3 guidance points
- Annex A
Examples of AI-related objectives guide what risk management should protect.
What you do
- Map your AI objectives, such as fairness and robustness, to risks
- Use them to shape risk criteria
- Annex B
Common AI risk sources, such as data quality and model complexity, help risk identification.
What you do
- Use the list as a prompt in assessments
- Add sources specific to your systems
- Annex C
Risk management activities are mapped onto the stages of the AI system life cycle.
What you do
- Run risk checks at design, development, deployment and retirement
- Record outputs at each stage
Evidence to keep
- Objective and risk source mapping
- Life cycle risk checklist
Nothing matches. Clear the search or pick another part.
Outcomes and implementation examples are from ISO/IEC 23894:2023 Information technology - Artificial intelligence - Guidance on risk management, published by NIST. Descriptions of what to do and the evidence lists are our own plain-English guidance. Check the official source for the authoritative text.
Your path to getting started
Five steps from first look to ready. Effort depends on your size and starting point.
- 1
List AI systems
And their purposes.
- 2
Define AI risk scales
Including harm to people.
- 3
Assess each system
Use a short template.
- 4
Treat top risks
Owners and dates.
- 5
Monitor
Re-assess on change.
How Optimus AI helps
Each area maps to a part of the platform, so the work and the evidence sit together.
| What it asks | What you do in Optimus AI |
|---|---|
| Principles | Risk policy |
| Framework | Roles and ownership |
| Context and scope | AI system records |
| Assessment | AI risk register |
| Treatment | Treatment plans |
| Monitoring | Review reminders |
Free ISO 23894 readiness checklist
A one-page PDF you can print or share with your team. No sign-up needed.
Download the checklist (PDF)- AI risk assessment template
- AI risk scale
- Treatment plan
- Review log
Share or cite this page
Writing about ISO 23894? You are welcome to link to this guide. Copy the link below.
<a href="/frameworks/iso-23894/">ISO 23894 explained in plain English - Optimus CyberX</a>Common questions
Is it certifiable?
No. It is guidance, not a set of requirements, so there is no certificate. Use ISO 42001 if you need a certifiable AI standard.
Do we need ISO 31000 first?
It is not required, but knowing ISO 31000 helps, because ISO 23894 follows the same risk process and adds what is different about AI.
How does it relate to ISO 42001?
ISO 42001 asks for AI risk assessment. ISO 23894 explains how.
Is it a law?
No. It is a voluntary international guide. It can support compliance with laws such as the EU AI Act but is not itself a legal requirement.
Related frameworks
See how ready you are for ISO 23894
Answer questions across every area. We email you a report showing where you comply and where you do not.