Artificial intelligence is becoming part of everyday business operations. Organisations are using AI to automate processes, analyse customer information, support financial decisions, improve cybersecurity, and increase productivity.

 

However, as AI adoption increases, so do the risks associated with these technologies.

 

An AI system may produce inaccurate results, expose sensitive information, introduce unfair outcomes, or make decisions that are difficult to explain. These risks can affect customers, employees, business operations, and an organisation’s reputation.

 

Traditional risk management practices provide a useful foundation, but they may not fully address the unique characteristics of AI systems.

 

ISO/IEC 23894:2023 provides international guidance to help organisations identify, assess, treat, monitor, and communicate risks associated with artificial intelligence. It supports organisations in integrating AI risk management into their existing governance and business processes.

 

In this guide, we explain what ISO 23894 is, why it matters, its key principles, how it relates to other AI governance standards, and how organisations can implement its guidance effectively.

 

 

What is ISO 23894?

 

ISO/IEC 23894:2023, officially titled Information technology — Artificial intelligence — Guidance on risk management, is an international standard published in February 2023 by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC).

 

It provides guidance for organisations that develop, produce, deploy, or use AI-based products, systems, and services.

 

The purpose of ISO 23894 is to help organisations manage risks specifically associated with artificial intelligence while integrating those activities into their broader risk management practices.

 

The guidance can be adapted to organisations of different sizes, industries, and levels of AI maturity.

 

Unlike a regulation that imposes mandatory legal obligations, ISO 23894 provides a structured approach that organisations can adopt voluntarily or use to support applicable contractual and regulatory requirements.

 

ISO 23894 at a glance

 

Category	Details
Official designation	ISO/IEC 23894:2023
Standard title	Information technology — Artificial intelligence — Guidance on risk management
Publication date	February 2023
Developed by	ISO/IEC JTC 1/SC 42, Artificial Intelligence
Primary purpose	Guidance for managing AI-related risks
Applicable organisations	AI developers, providers, deployers, and users
Main focus	Integrating AI risk management into organisational activities
Certification	Guidance standard, not a standalone management-system certification
Related standards	ISO 31000, ISO/IEC 42001, ISO/IEC 42005

 

The standard helps organisations establish a more consistent approach to AI risk management without requiring them to replace their existing governance structures.

 

Why is ISO 23894 important for organisations?

 

AI systems create risks that can differ significantly from those associated with traditional software and business technologies.

 

For example, conventional software generally follows predefined instructions, while some AI systems generate outputs based on patterns learned from data. Their behaviour may be difficult to predict in every situation, especially when the data, environment, or user interactions change.

 

This creates additional challenges for risk assessment, monitoring, accountability, and regulatory compliance.

 

1. AI systems can produce unexpected outcomes

 

AI systems may generate inaccurate predictions, misleading information, or unreliable recommendations.

 

For example, an AI-powered financial assessment tool might incorrectly classify a legitimate customer as high risk because its training data does not adequately represent certain customer groups.

 

Without appropriate risk assessments and monitoring, such errors may remain undetected and affect business decisions.

 

ISO 23894 encourages organisations to consider risks throughout the AI lifecycle rather than focusing only on initial development or deployment.

 

2. AI introduces new data protection and security risks

 

Many AI systems depend on large datasets containing business information, customer records, or other sensitive data.

 

Poor data governance, inappropriate access controls, insecure integrations, and weaknesses in third-party AI services can create security and privacy risks.

 

Organisations need to understand how information is collected, processed, stored, shared, and used by their AI systems.

 

An effective AI risk management process helps identify these exposures and determine appropriate safeguards.

 

3. AI decisions can affect individuals and society

 

Some AI applications influence decisions about employment, financial services, healthcare, education, and access to essential services.

 

If these systems produce unfair or unreliable outcomes, the consequences may extend beyond the organisation operating them.

 

Risk management must therefore consider potential impacts on individuals, groups, and other affected stakeholders.

 

This requires collaboration between technical teams, business leaders, compliance professionals, and relevant subject matter experts.

 

4. Organisations need stronger AI accountability

 

AI governance becomes difficult when responsibility for managing AI risks is unclear.

 

For example, a technology team may develop an AI application while another department uses it to support important business decisions. Without clearly assigned responsibilities, issues involving accuracy, security, or compliance may not receive appropriate attention.

 

ISO 23894 supports integrating AI risk management into organisational governance so that risk decisions are documented, responsibilities are understood, and appropriate oversight can be maintained.

 

 

What are the key principles of ISO 23894?

 

ISO 23894 builds on established risk management concepts and adapts them to the characteristics of artificial intelligence.

 

The following principles explain how organisations can apply its guidance in practice.

 

Integrating AI risk management into existing processes

 

AI risk management should not operate as an isolated activity managed only by data scientists or technology teams.

 

Organisations should connect AI-related risks with their existing enterprise risk management, cybersecurity, compliance, and governance processes.

 

For example, an organisation introducing an AI-powered customer service platform should consider operational risk, data privacy, third-party dependencies, and potential customer impacts within its existing risk management structure.

 

This approach helps leadership maintain a consistent view of organisational risks.

 

Considering the complete AI lifecycle

 

AI risks can emerge at different stages of a system’s lifecycle.

 

During development, risks may arise from training data quality, model design, or testing limitations. During deployment, risks may involve integration failures, inappropriate access, or unexpected system behaviour.

 

Once the system is operating, changes in data patterns, user behaviour, or external conditions may affect its performance.

 

Organisations should therefore review risks throughout development, deployment, operation, modification, and retirement.

 

Applying risk management according to context

 

Not every AI system presents the same level of risk.

 

An AI application used to recommend internal meeting times generally has a different risk profile from an AI system used to support medical decisions or evaluate loan applications.

 

ISO 23894 allows organisations to adapt their risk management approach according to the nature of the AI system, its intended purpose, the operating environment, and the potential consequences of failure.

 

This helps organisations direct resources towards the systems and risks that require greater attention.

 

Involving relevant stakeholders

 

AI risks often involve technical, legal, operational, and ethical considerations.

 

Effective assessments may require input from data scientists, cybersecurity professionals, compliance officers, legal advisers, business owners, and individuals who understand how the system affects its users.

 

Including relevant perspectives can help organisations identify risks that may be overlooked when assessments are conducted by a single department.

 

Continuously reviewing AI risks

 

An AI risk assessment should not be treated as a one-time document completed before deployment.

 

Models may be updated, data quality may change, new vulnerabilities may emerge, and organisations may begin using AI systems for purposes that were not originally anticipated.

 

Regular monitoring and reassessment help organisations identify changes in risk exposure and evaluate whether existing safeguards remain effective.

 

What types of AI risks does ISO 23894 help organisations manage?

 

ISO 23894 provides guidance that organisations can apply across different categories of AI risk.

 

The following categories represent practical areas to consider when developing an AI risk management programme. They are not presented as an exhaustive classification from the standard.

 

AI security and cybersecurity risks

 

AI systems can introduce vulnerabilities through insecure integrations, unauthorised access, malicious inputs, and weaknesses in their supporting infrastructure.

 

For example, a generative AI application connected to internal business systems may be exposed to prompt injection attacks that attempt to manipulate its behaviour.

 

Organisations should assess these threats and implement appropriate security controls, access restrictions, testing procedures, and monitoring activities.

 

Data privacy and confidentiality risks

 

AI applications may process personal information, confidential business records, or commercially sensitive data.

 

Risks can arise when information is collected without appropriate controls, retained unnecessarily, or shared with external AI services without sufficient oversight.

 

Risk assessments should consider data classification, access permissions, retention practices, third-party processing, and applicable privacy obligations.

 

Bias and fairness risks

 

AI systems trained on incomplete or unrepresentative datasets may produce outcomes that disadvantage certain individuals or groups.

 

For example, an AI recruitment tool might consistently favour applicants with particular backgrounds because of patterns present in historical hiring data.

 

Organisations should evaluate whether their AI systems produce inappropriate or unfair outcomes and establish controls to reduce identified risks.

 

Accuracy and reliability risks

 

AI-generated results may be incorrect, inconsistent, or unsuitable for a particular business decision.

 

This is particularly important when organisations use AI to support activities involving financial reporting, legal analysis, healthcare, or safety-critical operations.

 

Testing, validation, human oversight, and performance monitoring can help organisations identify and manage these risks.

 

Transparency and explainability risks

 

Some AI systems are difficult to interpret, making it challenging to explain why a particular result was produced.

 

This can create governance problems when employees, customers, regulators, or auditors require an explanation of decisions.

 

Organisations should evaluate the level of transparency and explanation needed for each AI use case and establish appropriate documentation and review processes.

 

Third-party AI risks

 

Many organisations rely on external AI providers, cloud services, APIs, and pre-trained models.

 

These dependencies may create risks involving service availability, security, data processing, contractual responsibilities, and changes to model behaviour.

 

Third-party risk assessments should therefore consider how external AI services affect the organisation’s overall risk exposure.

 

How to implement ISO 23894: A practical step-by-step guide

 

Implementing ISO 23894 requires organisations to translate risk management principles into clear responsibilities, repeatable assessments, and ongoing monitoring activities.

 

The following implementation approach is a practical workflow informed by the standard’s purpose and established risk management practices. It is not a reproduction of the standard’s official clauses.

 

Step 1: Establish an AI risk management policy

 

Start by defining how your organisation intends to manage AI-related risks.

 

The policy should explain the scope of AI risk management, governance responsibilities, reporting expectations, and how risk decisions will be approved.

 

It should also clarify how AI risk management connects with existing enterprise risk management, information security, privacy, and compliance policies.

 

Senior leadership should support the policy and ensure that appropriate resources are available.

 

Step 2: Create an inventory of AI systems

 

Organisations cannot effectively manage AI risks without understanding which AI systems they use.

 

Create an inventory covering internally developed applications, third-party AI tools, generative AI services, and AI features embedded within existing software.

 

For each system, document its purpose, business owner, provider, data sources, affected processes, and level of business importance.

 

This inventory helps organisations identify systems that require detailed assessments and ongoing oversight.

 

Step 3: Define the scope and context of each assessment

 

Before assessing an AI system, determine how it will be used and what could happen if it performs incorrectly.

 

Consider the intended users, business objectives, operating environment, affected stakeholders, and applicable legal or regulatory requirements.

 

For example, an AI system supporting loan approvals may require a more detailed assessment than an internal tool used to summarise non-sensitive meeting notes.

 

Defining the context helps ensure that risk assessments reflect real-world consequences.

 

Step 4: Identify AI-related risks

 

Work with relevant stakeholders to identify potential threats, weaknesses, failures, and harmful outcomes.

 

Consider cybersecurity vulnerabilities, data privacy concerns, inaccurate predictions, biased decisions, operational dependencies, and misuse of AI capabilities.

 

Risk identification should also examine how people interact with the system and whether inappropriate reliance on AI-generated outputs could create additional risks.

 

Record each identified risk in a structured register so that it can be assessed, assigned, and monitored.

 

Step 5: Analyse and evaluate identified risks

 

After identifying risks, assess their potential consequences and likelihood using a method appropriate to the organisation and AI use case.

 

Consider both the probability of an undesirable outcome and the severity of its impact.

 

For AI systems, organisations may also need to account for uncertainty, limited testing data, or difficulty estimating the likelihood of certain failures.

 

Compare assessment results with the organisation’s risk criteria and tolerance levels to determine which risks require treatment.

 

Step 6: Develop AI risk treatment plans

 

Once risks have been prioritised, determine the appropriate response.

 

Depending on the circumstances, organisations may reduce risks through technical safeguards, strengthen human oversight, restrict certain AI capabilities, modify the system, or decide not to proceed with a particular use case.

 

Some risks may be accepted when they fall within approved tolerance levels, but these decisions should be documented and authorised.

 

Each treatment plan should identify responsible owners, required actions, target dates, and methods for checking whether the controls are effective.

 

Step 7: Document decisions and maintain evidence

 

Documentation is essential for accountability and effective governance.

 

Maintain records of risk assessments, identified risks, treatment decisions, testing results, approvals, control activities, and reassessment outcomes.

 

These records help organisations explain how they manage AI risks and provide useful evidence during internal reviews, audits, and regulatory assessments.

 

Documentation should be maintained as systems change rather than created only when an audit is approaching.

 

Step 8: Monitor AI systems and review risks regularly

 

AI risk management continues after deployment.

 

Organisations should monitor relevant performance indicators, security events, incidents, changes in system behaviour, and the effectiveness of risk controls.

 

Reassess systems when significant changes occur, such as model updates, new data sources, expanded use cases, or changes in applicable requirements.

 

Regular reviews help organisations identify emerging risks and adjust their controls before problems become more serious.

 

Practical example: Applying ISO 23894 to an AI recruitment system

 

Consider an organisation introducing an AI application to screen job applications and recommend candidates for interviews.

 

The objective is to reduce administrative work and help recruitment teams review applications more efficiently.

 

However, the system may introduce several risks that need to be assessed before deployment.

 

Risk area	Example risk	Possible treatment
Fairness	Recommendations disadvantage certain applicant groups	Test outcomes, review data quality, and introduce human oversight
Privacy	Candidate information is accessed or retained inappropriately	Apply access controls and retention policies
Accuracy	Qualified applicants are incorrectly rejected	Validate recommendations and require appropriate human review
Transparency	Recruitment teams cannot explain recommendations	Maintain documentation and provide suitable explanations
Third-party risk	External AI provider changes the model without adequate notice	Establish vendor review and change management procedures
Accountability	No department owns the system’s risk decisions	Assign a business owner and formal approval responsibilities

 

The organisation should document these risks, evaluate their significance, implement appropriate controls, and monitor the system after deployment.

 

This example demonstrates how ISO 23894 can support more responsible AI adoption by connecting technical risks with business decisions and governance responsibilities.

 

ISO 23894 vs ISO 42001: What is the difference?

 

ISO 23894 and ISO/IEC 42001 are related standards, but they serve different purposes.

 

ISO 23894 provides guidance on managing AI-related risks, while ISO 42001 specifies requirements for establishing, implementing, maintaining, and continually improving an AI management system.

 

ISO/IEC 42001:2023 is a management-system standard that supports responsible organisational governance of AI systems.

 

Comparison	ISO 23894	ISO 42001
Main purpose	AI risk management guidance	AI management system requirements
Focus	Identifying, analysing, evaluating, and treating AI risks	Establishing and maintaining an organisational AI management system
Type	Guidance standard	Requirements standard
Certification	Not a standalone certifiable management-system standard	Supports third-party management-system certification
Main audience	AI developers, users, risk managers, and governance teams	Organisations developing, providing, or using AI systems
Practical role	Helps organisations manage AI-related risks	Establishes a structured system for governing AI

 

Can organisations use ISO 23894 and ISO 42001 together?

 

Yes. The two standards can complement each other.

 

An organisation implementing ISO 42001 can use ISO 23894 as supporting guidance when developing and improving its AI risk management processes.

 

For example, ISO 42001 can provide the management-system structure for policies, responsibilities, objectives, and continual improvement, while ISO 23894 can inform how AI-related risks are identified, evaluated, and treated.

 

Using both can help organisations connect their AI governance policies with practical risk management activities.

 

ISO 23894 vs NIST AI RMF: How do they compare?

 

The NIST AI Risk Management Framework (AI RMF) is another widely recognised approach to managing artificial intelligence risks.

 

Developed by the US National Institute of Standards and Technology, the framework helps organisations manage risks associated with AI systems and promote trustworthy AI development and use.

 

NIST AI RMF 1.0 organises its core around four functions: Govern, Map, Measure, and Manage.

 

Both ISO 23894 and NIST AI RMF encourage organisations to establish governance, assess AI-related risks, implement appropriate responses, and maintain ongoing oversight.

 

However, their structures and approaches differ.

 

Area	ISO 23894	NIST AI RMF
Publisher	ISO and IEC	NIST
Main objective	Guidance for managing AI-related risks	Framework for managing AI risks and supporting trustworthy AI
Structure	Risk management guidance and integration	Govern, Map, Measure, Manage
Application	Adaptable to organisational context	Voluntary and adaptable across sectors
Use alongside other standards	Can complement ISO 42001 and ISO 31000	Can complement ISO standards and other risk frameworks

 

Organisations do not necessarily need to choose one approach exclusively.

 

They can use ISO 23894 to strengthen their AI risk management methodology while drawing on NIST AI RMF resources to support governance, measurement, and risk treatment activities.

 

The appropriate combination depends on organisational needs, applicable requirements, and the maturity of existing governance processes.

 

Is ISO 23894 mandatory or certifiable?

 

ISO 23894 is a guidance standard, which means organisations are not generally required by the standard itself to obtain certification or demonstrate conformity through a dedicated ISO 23894 certification scheme.

 

It does not establish a standalone set of management-system certification requirements.

 

However, its guidance may help organisations support obligations arising from laws, regulations, contracts, and internal governance policies.

 

For example, an organisation subject to AI-related regulatory requirements may use ISO 23894 to inform its risk assessment and control processes.

 

Implementing ISO 23894 does not automatically establish compliance with the EU AI Act or any other law. Organisations must separately assess the legal requirements applicable to their systems, roles, jurisdictions, and activities.

 

If an organisation is seeking certification for an AI management system, ISO/IEC 42001 is the more relevant standard to evaluate.

 

What are the benefits of implementing ISO 23894?

 

Organisations that apply ISO 23894 effectively can strengthen their ability to identify AI-related risks and make informed decisions about AI adoption.

 

Better visibility into AI risks

 

A structured risk management process helps organisations understand where AI systems are being used, which business activities depend on them, and what could happen if those systems fail.

 

This visibility can help leadership prioritise resources and address significant risks more effectively.

 

Improved governance and accountability

 

Clearly defined responsibilities help ensure that AI risks are not left entirely to technical teams.

 

Business owners, risk managers, compliance professionals, and senior leaders can work together to evaluate risks and approve appropriate responses.

 

Stronger compliance readiness

 

Documented assessments, control decisions, and monitoring activities can help organisations demonstrate how they manage AI-related risks.

 

These records may support compliance reviews and audits, although additional evidence will be needed to satisfy specific legal or certification requirements.

 

More consistent risk assessments

 

Using a common approach allows organisations to assess different AI systems against established risk criteria.

 

This helps reduce inconsistent decision-making and makes it easier to compare risks across business units and applications.

 

Greater confidence in AI adoption

 

Effective risk management can help organisations identify which AI applications are appropriate for their business and what safeguards are needed before deployment.

 

This supports innovation while reducing the likelihood of unmanaged risks affecting operations, customers, or other stakeholders.

 

Common challenges when implementing ISO 23894

 

Although ISO 23894 provides useful guidance, organisations may encounter several practical challenges when putting it into operation.

 

Limited visibility into AI applications

 

Employees may use AI tools without formal approval or centralised oversight.

 

This makes it difficult for governance teams to understand which systems process organisational data and whether appropriate risk assessments have been completed.

 

Maintaining an AI inventory and establishing clear approval processes can help improve visibility.

 

Disconnected risk and compliance processes

 

AI risk assessments may be maintained separately from cybersecurity risk registers, vendor assessments, compliance obligations, and internal controls.

 

This fragmentation makes it harder to identify relationships between risks and maintain consistent reporting.

 

Integrating AI risk management into existing GRC processes can help organisations reduce duplication and improve accountability.

 

Lack of clear risk ownership

 

AI systems often involve multiple departments, including IT, legal, security, operations, and business management.

 

Without clear ownership, risk treatment activities may be delayed or overlooked.

 

Organisations should assign responsibility for each AI system and establish appropriate escalation and approval procedures.

 

Difficulty maintaining continuous oversight

 

AI systems may change more frequently than traditional business applications.

 

Manual assessments and spreadsheet-based tracking can make it difficult to maintain accurate records of risks, controls, and remediation activities.

 

Structured workflows, scheduled reviews, and appropriate automation can help organisations maintain more consistent oversight.

 

How CyberArrow GRC supports AI risk management

 

Implementing ISO 23894 requires more than identifying potential AI risks. Organisations also need a reliable way to document assessments, assign responsibilities, manage controls, track remediation, and maintain oversight as their AI systems evolve.

 

At CyberArrow, we understand that managing these activities through disconnected spreadsheets and manual processes can create unnecessary complexity.

 

CyberArrow GRC helps organisations bring governance, risk management, and compliance activities together in one centralised platform.

 

Our platform supports several activities that are important when operationalising AI risk management guidance.

 

Centralised risk assessments and risk registers

 

CyberArrow GRC helps organisations conduct structured risk assessments and maintain centralised records of identified risks.

 

Teams can document AI-related risks, assign ownership, track their status, and manage risk treatment activities alongside other enterprise risks.

 

This helps organisations avoid maintaining separate risk management processes for every technology or business function.

 

Automated compliance workflows

 

Our platform helps reduce repetitive administrative work through automated compliance workflows, task management, and evidence collection.

 

These capabilities support organisations that need to coordinate AI governance activities across multiple departments and maintain consistent documentation.

 

Policy and control management

 

CyberArrow GRC helps organisations centralise policies, manage controls, and maintain accountability for important governance activities.

 

By connecting risks with relevant controls and responsibilities, organisations can improve visibility into how their risk management decisions are implemented.

 

Continuous control monitoring and audit readiness

 

AI risk management requires ongoing oversight as systems, data, and business requirements change.

 

CyberArrow GRC supports continuous internal control monitoring, compliance tracking, and evidence management to help teams identify issues and maintain audit readiness.

 

These capabilities can support an organisation’s ISO 23894 implementation programme, although using CyberArrow GRC does not automatically establish conformity with ISO 23894 or compliance with AI regulations.

 

Conclusion: Strengthen AI risk management with CyberArrow GRC

 

Artificial intelligence offers significant opportunities for organisations to improve efficiency, support decision-making, and develop new products and services. However, these benefits must be supported by effective governance and risk management.

 

ISO 23894 provides valuable guidance for organisations seeking to identify, assess, treat, and monitor AI-related risks throughout the AI lifecycle.

 

By adopting a structured approach, organisations can improve accountability, strengthen oversight, and make more informed decisions about how AI technologies are developed and used.

 

However, effective AI risk management requires more than policies and periodic assessments. Organisations need consistent processes, reliable documentation, clear ownership, and ongoing visibility into risks and controls.

 

At CyberArrow, we help organisations simplify these challenges through CyberArrow GRC, our integrated governance, risk, and compliance platform.

 

Our platform helps teams automate compliance workflows, manage risk assessments, centralise policies and controls, collect evidence, track remediation activities, and maintain continuous visibility into their risk and compliance processes.

 

CyberArrow is trusted by the world’s biggest brands across the US, Europe, Africa, Asia and the Middle East. We support organisations in strengthening governance, simplifying compliance, and building more efficient risk management programmes.

 

Whether your organisation is beginning its AI governance journey or improving existing AI risk management processes, CyberArrow GRC can help you establish a more consistent and accountable approach.

 

Ready to strengthen your AI risk management and simplify GRC? Book a free CyberArrow GRC demo and discover how we can help your organisation manage risks and compliance more effectively.

 

 

FAQs