AI risk management for ISO 42001

AI systems carry specific risks. Model drift, bias, decisions that cannot be explained, data problems. ISO 42001 requires you to identify and control these risks systematically. The NIST AI Risk Management Framework offers additional guidance. On this page we explain how.

Why is AI risk management different under ISO 42001?

You know traditional IT risks: servers that fail, data that leaks, software with bugs. AI adds a layer on top of those. Alongside risk management, ISO 42001 also requires an AI impact assessment, which looks at the consequences for people.

AI learns. A traditional system does what you programme it to do. AI learns from data and can show unexpected behaviour.

AI is probabilistic. AI does not give definite answers but probabilities. There is always a chance of errors.

AI changes. Models can degrade over time (model drift) when the world changes but the model does not.

AI is hard to understand. Complex models are black boxes. You do not always know why a decision is made.

This calls for specific risk management practices.

The categories of AI risks

Technical risks

Model drift. The model performs worse over time because reality changes. The model was trained on last year’s data, but today’s world is different.

Example: A fraud detection system was trained on pre-COVID fraud patterns. After COVID, behaviour changed. The model misses new forms of fraud.

Overfitting. The model is trained too specifically on the training data and generalises poorly to new situations.

Data poisoning. Malicious actors manipulate training data so the model learns the wrong things.

Adversarial attacks. Attackers feed the model specific input to generate wrong output.

Technical debt. Poorly documented models, hardcoded parameters, no version control. Nobody knows anymore how it works.

Fairness and bias risks

Historical bias. The training data contains historical prejudices that the model learns.

Example: A recruitment system trained on historical decisions learns that men are hired more often, and starts scoring men higher.

Measurement bias. The features you measure are not representative or are themselves biased.

Proxy discrimination. The model discriminates indirectly through proxies. Postcode correlates with ethnicity, education correlates with socio-economic background.

Aggregation bias. The model works well on average, but poorly for specific subgroups.

Data risks

Data quality. Errors in the data lead to errors in the output. Garbage in, garbage out.

Representativeness. The training data is not representative of the real population.

Privacy. The training data contains personal data. Can the model leak it?

Data lineage. It is unclear where data comes from and what has been done with it.

Operational risks

Misuse. The system is used for purposes it was not designed for.

Over-reliance. Users trust the AI blindly and switch off their own judgement.

System outage. What happens if the AI is unavailable? Is there a fallback?

Scaling. The system works in testing but fails under production load.

Societal risks

Impact on employment. Automating tasks can cost jobs.

Reinforcing inequality. AI can reinforce existing social inequalities.

Public trust. AI incidents damage trust in your organisation and in AI in general.

The AI impact assessment goes deeper into societal risks. Risk management and impact analysis complement each other.

The risk management process

ISO 42001 requires a documented process for AI risk management. The process follows four steps. Below you can see for each step what you do and what it delivers.

Step	What do you do?	Result
1. Risk identification	Map all potential AI risks	Risk list per AI system
2. Risk analysis	Determine likelihood and impact per risk	Risk scores
3. Risk evaluation	Test the scores against your risk appetite	Overview of acceptable and unacceptable risks
4. Risk treatment	Choose a strategy: avoid, reduce, transfer or accept	Treatment plan with measures

We work out each step below with examples and techniques.

Tip: AI risks are more dynamic than traditional IT risks. Models can degrade over time (model drift) without you noticing. So plan periodic reassessments, not only after major changes.

Step 1: Risk identification

Map all potential risks. For each AI system, ask:

What can technically go wrong?
Which data is used, and what are the risks of that data?
How could the system be misused?
Who is affected if things go wrong?
Which external factors could affect the system?

Techniques:

Brainstorming sessions with diverse stakeholders
Checklist reviews against known AI risks
Analysis of incidents at comparable systems
Threat modelling

Step 2: Risk analysis

For each identified risk, you determine:

Likelihood: How likely is it that this risk will materialise?

High: Likely to happen within a year
Medium: Could happen, but not likely
Low: Unlikely

Impact: What are the consequences if the risk materialises?

High: Serious harm to people, major financial loss, legal consequences
Medium: Significant but manageable consequences
Low: Minimal consequences

The combination gives a risk score.

Step 3: Risk evaluation

Compare the risk scores with your risk appetite. Which risks are acceptable? Which are not?

Typically:

High risk: Not acceptable, must be treated
Medium risk: Treat, or accept with monitoring
Low risk: Accept with documentation

Step 4: Risk treatment

For each risk that needs treatment, you choose a strategy:

Avoid: Stop the activity that causes the risk. Do not use this AI system.

Reduce: Implement controls that lower the likelihood or impact.

Transfer: Pass the risk on, for example through insurance or outsourcing.

Accept: Accept the risk consciously and document this.

Controls for AI risks

The Annex A controls provide concrete measures. Here are the most important ones per risk type.

Against technical risks

Monitoring and alerting. Monitor model performance continuously. Alerts when deviations occur.

Validation and testing. Test thoroughly before deployment. Retest regularly.

Version control. Keep all model versions. Make rollback possible.

Documentation. Document models: architecture, training data, parameters, limitations.

Against bias risks

Bias detection. Test models for bias before and after deployment. Use tools such as IBM AI Fairness 360 or Amazon SageMaker Clarify.

Diverse training data. Make sure the training data is representative of all groups.

Fairness metrics. Define and measure fairness metrics.

Human review. Have people check decisions, especially where the impact is high.

Against data risks

Data governance. Policy for data collection, storage and use.

Data quality checks. Automatic checks for completeness, consistency and timeliness.

Data lineage. Document the origin and transformations of data.

Privacy by design. Minimise personal data and anonymise where possible.

Against operational risks

Fallbacks. Plan what happens if the AI fails. A manual process as backup.

User training. Train users in responsible use. See AI literacy.

Scope definition. Define clearly what the system is and is not intended for.

Load testing. Test whether the system can handle production load.

Bias detection in practice

Bias is one of the biggest AI risks. How do you deal with it?

Before deployment (pre-deployment)

Analyse training data. Is every group sufficiently represented? Are there historical prejudices?

Test on subgroups. Measure performance separately for each demographic group. Are there large differences?

Calculate fairness metrics.

Statistical parity: Do all groups get a positive result equally often?
Equal opportunity: Is the true positive rate the same for all groups?
Disparate impact: Is the ratio between groups within acceptable limits?

During operation (in production)

Monitor outcomes. Keep measuring whether outcomes are distributed fairly.

Collect feedback. Give people the option to challenge decisions.

Periodic audits. Have a bias audit carried out regularly.

Afterwards (when problems arise)

Mitigation techniques. Pre-processing (adjust the data), in-processing (adjust the model), post-processing (adjust the output).

Transparency. Communicate honestly when bias has been found and what you are doing about it.

Fully bias-free AI does not exist. The goal is to minimise and control bias, not to eliminate it. Be honest about the limitations.

Tools for AI risk management

You do not have to reinvent the wheel. There are common tools for each purpose.

Purpose	Example tools
Bias and fairness	IBM AI Fairness 360 , Microsoft Fairlearn , Google What-If Tool
Model monitoring	MLflow , Evidently AI , Arize
Explainability	SHAP , LIME , InterpretML

ISO 42001 documentation for AI risk management

ISO 42001 requires documented risk management. Record:

The risk management process
The identified risks per AI system
The analysis (likelihood, impact)
The chosen treatment
The implemented controls
The residual risks after treatment
Reviews and updates

A risk register (a spreadsheet or tool) is the central document.

ISO 42001 within the ISO/IEC AI standards family

ISO 42001 does not stand on its own. The committee ISO/IEC JTC 1/SC 42 develops a whole family of AI standards. A few of them are especially relevant for risk management. ISO 42001 is the only certifiable management system standard; the others are guidance that helps you fill in the requirements.

Standard	Name	Role in risk management
ISO/IEC 42001:2023	AI management system (AIMS)	Certifiable framework: requires a documented risk process
ISO/IEC 23894:2023	Guidance on AI risk management	Elaborates the risk process, specifically for AI
ISO/IEC 22989:2022	Concepts and terminology	Common terms such as bias and model drift
ISO/IEC TR 24028:2020	Overview of trustworthiness in AI	Background on robustness, safety and explainability
ISO/IEC 42005:2025	AI system impact assessment	Guidance for the AI impact assessment

A full overview is available on ISO standards for AI. For risk management, ISO/IEC 23894 is the most practical addition: it describes how to fill in the risk process, while ISO 42001 requires that you have one.

And the US NIST AI RMF

Outside the ISO family there is the NIST AI Risk Management Framework . This American framework is complementary and is built on four functions: Govern, Map, Measure and Manage. If you implement ISO 42001, you can use the NIST AI RMF as additional guidance. The concepts overlap, and its approach can help.