logo

Blog & Insights

ISO/IEC 42001 Certification: Who Is Accountable for Your AI?

ISO/IEC 42001 Certification: Who Is Accountable for Your AI?

ISO Certification

ISO/IEC 42001 certification assesses an artificial intelligence management system, or AIMS. It concerns how an organisation governs the development, provision or use of AI within a defined boundary. The practical starting point is accountability: which AI systems are involved, what people rely on them to do and who can approve, change or stop their use.

ISO/IEC 42001:2023 is the current published edition, confirmed by ISO’s standard record. Certification is about the assessed management system. It does not establish that every model is accurate, every output is safe, every dataset is unbiased or every applicable law has been satisfied. Buyers need to understand the use case and the evidence behind any more specific assurance claim.

Identify your role before choosing the certification boundary

An organisation can develop AI, integrate third-party models, provide an AI-enabled service, or use AI in internal decisions. It may perform several roles at once. Those roles affect what it controls and what evidence it needs from others. Buying an external AI service does not remove the need to decide how it may be used within your organisation.

Describe the intended scope in operational language. Identify the legal entity, teams, AI-enabled activities and relevant lifecycle stages. If the system covers one customer-support application, do not present it as certification of every AI activity in the corporate group. Conversely, an enterprise-wide claim needs a method for identifying AI use beyond the central technology department, including tools acquired through ordinary business subscriptions.

Build an inventory that answers management questions

A useful inventory records more than model names. Consider the business purpose, owner, affected users, source or provider, data involved, integration points and decision authority. Distinguish a trial from a production service and identify uses that could materially affect people or essential operations. Keep sensitive technical and personal information appropriately protected.

Ask how the inventory stays current. Procurement, software changes, employee experimentation and supplier updates can introduce new uses. A one-time survey may miss these routes. Establish a practical trigger for review when a tool’s purpose, model, data or user population changes. The inventory should help management find responsibilities and risks, rather than become a catalogue nobody consults before deployment.

Separate organisational risk from impacts on people

An AI application can create business risks and affect individuals or groups. A performance error, inaccessible interface or inappropriate use of personal information may matter differently to each affected party. Evaluate the intended purpose and foreseeable circumstances of use, including how people may misunderstand or over-rely on the output.

Involve relevant perspectives: technical teams, operational owners, legal or privacy specialists and people familiar with the affected users. This is especially useful when an application influences access, recommendations or important decisions. A risk score should support a reasoned decision, with assumptions and limitations recorded. It should not be used to make a troubling impact disappear by selecting a convenient numerical rating.

Translate responsible AI principles into operating choices

Transparency, oversight and accountability become useful when connected to actions. Decide what users need to know, which outputs require review, what evidence supports release and when escalation is necessary. Specify who has authority to suspend use. Avoid a policy that promises human oversight while providing no time, information or means for a person to intervene.

The NIST AI Risk Management Framework is a complementary voluntary resource covering governance and risk across the AI lifecycle. It is not an ISO certification scheme. Teams may use such resources to improve discussions, while keeping a clear distinction between supporting guidance, the requirements standard and legal obligations applicable to the actual use.

Use an accountability-and-evidence map

This original table is a planning aid. It does not prescribe a universal set of AI controls.

Lifecycle decision

Accountable question

Evidence to consider

Selecting a use case

Why is AI appropriate here?

Purpose, alternatives and relevant constraints

Choosing data or a provider

What limitations do we accept?

Provenance, permissions and supplier information

Approving release

What establishes acceptable performance?

Evaluation results and authorised decision

Operating the service

Who detects and responds to problems?

Monitoring, escalation and user feedback

Changing or retiring it

What happens to affected users and records?

Change review, communications and retention decisions

 

Take one live use case through the table. If a supplier provides the underlying model, distinguish the supplier’s responsibilities from your integration, configuration and deployment decisions. Responsibility should remain understandable when a problem crosses organisational boundaries.

Evaluate performance for the actual setting

An impressive demonstration may not represent the conditions in which the application will operate. Choose evaluation data and methods relevant to the intended users, tasks and risks. Document important limitations, including circumstances where the system should not be relied on. Consider whether performance changes across different groups or operating conditions where relevant and lawful to evaluate.

Define a practical response to unexpected results. Who investigates, who decides whether use can continue and how are affected users informed? Monitoring should be connected to a decision, rather than generating dashboards without follow-up. Avoid claiming that a single aggregate accuracy measure establishes suitability for every purpose. The significance of an error depends on the use and its consequences.

Manage external models and changing services

AI services can change through model updates, revised interfaces, new settings or altered provider terms. Establish how changes are identified and reviewed. Determine which evidence the provider can supply and what remains for your organisation to evaluate. A supplier’s AIMS certificate does not automatically certify your use case or prove that a specific model meets your performance needs.

Contracts and operating arrangements may need to address data handling, permitted use, notification, support, incident communication and exit. Ask what happens if the service becomes unavailable or no longer meets your needs. The appropriate safeguards depend on the context; do not assume that all AI suppliers offer the same evidence, contractual commitments or control over updates.

A fictional example: an internal assistant becomes customer-facing

Imagine a fictional company that trials an AI assistant for drafting internal help notes. A business team later connects it to a customer portal. The original evaluation did not consider external users, disclosure, unsuitable advice or escalation to a person. The technical tool is similar, but the purpose and exposure have changed.

A useful response would pause the unsupported expansion where appropriate, review the new use, evaluate risks and impacts, assign responsibility and establish suitable release evidence. This fictional illustration shows why change management matters. It does not claim a real deployment failure or prescribe a legal outcome. The organisation should assess its particular application and obligations with competent advice.

Implement an AIMS through real decisions

Obtain the requirements standard, assign leadership and define scope. Review existing governance, procurement, security, privacy and development arrangements. Establish the necessary AI-specific responsibilities, assessments and controls without duplicating useful processes unnecessarily. Where the standard requires control applicability documentation, connect it to risks and the reasons for chosen controls rather than treating an annex as a shopping list.

Train people for their roles, operate the arrangements and retain decision evidence. Conduct internal audits that examine live use cases and supplier relationships. Management review should address performance, changes, resources and lessons from incidents or complaints. Readiness for certification depends on evidence of a functioning management system, not a policy document claiming responsible AI in broad terms.

Ask the certification body about AI assessment competence

Explain your roles, use cases, technologies and operating boundary to CB Name. Ask how the team will assess the organisation’s responsibilities and what expertise supports the proposed audit. Confirm accreditation for AIMS certification rather than assuming approval for information security extends automatically to AI management systems.

ISO/IEC 42006:2025 sets additional requirements for bodies auditing and certifying AIMS. Use the certification-body directory for discovery and check the relevant approval independently. Clarify evidence confidentiality, audit coverage and the exact issuer. A general statement that a body “offers AI certification” is insufficient to establish the nature or recognition of its service.

Budget for governance and maintain a clear assurance claim

Costs may include internal time, evaluation, data work, specialist advice, supplier assurance, training and external assessment. The effort depends on the use cases and existing maturity. A company operating one limited internal application faces different work from a provider managing multiple AI-enabled services. Avoid universal implementation deadlines or fixed certification prices without a defined scope.

After assessment, review the system as technologies, uses and responsibilities change. Explain the certified boundary accurately in procurement and marketing. AI legislation may introduce separate product, role and deployment obligations; certification does not automatically establish compliance. For certificate status checks, use the verification guide, then examine application-specific assurance relevant to the intended decision.

Decide what users can reasonably rely on

For each important application, describe the intended user and the task the system supports. Explain the limits in language that person can understand. A warning hidden in technical documentation may not help a customer who encounters the output without context. Consider how interface design, defaults and workflow influence reliance on the result.

Ask whether a person can recognise a questionable output and obtain appropriate help. Human review is difficult when reviewers lack subject knowledge, see only the final answer or face an unrealistic volume of work. Evaluate the arrangement as it will operate, including busy periods and exceptions. A policy describing oversight should match the authority and resources provided.

Retain the reasoning behind the chosen approach and revisit it when use expands. An internal drafting assistant may later influence customer advice or operational action, changing the consequences of error. The organisation should recognise that shift before relying on the original evaluation. This user-focused review is a practical complement to technical testing: it examines how the application participates in a real decision. It does not establish legal acceptability or replace assessment of particular impacts, but it helps management identify where stronger information, controls or expert review may be needed.

References and information note

·         ISO/IEC 42001: AI management system requirements

·         ISO/IEC 42006: AIMS certification-body requirements

·         NIST: AI Risk Management Framework

·         NIST: AI RMF Playbook

This guide is general information, not an assessment of an AI product or legal compliance. The example is fictional. Certification does not establish that a model is bias-free or suitable for every use. Submit factual corrections through the contact page.

Frequently Asked Questions

Can an organisation that only uses third-party AI apply ISO/IEC 42001?

Yes. Its management system should reflect its responsibilities as a user and any other roles it performs.

Is an AIMS certificate a certificate for one model’s accuracy?

No. Model performance claims need evidence specific to the model, task and operating context.

Does information security certification automatically cover AI governance?

No. Related processes can support one another, but the standards and certification scopes are distinct.

Should an experimental tool appear in the AI inventory?

Establish an inventory approach that identifies relevant trials and their status, especially where they may affect people or operations.

Does human oversight mean adding a final approval button?

Not necessarily. The person needs appropriate information, competence, time and authority to make meaningful decisions.

Can a supplier’s certificate remove our deployment responsibilities?

No. Assess your own configuration, integration, use and monitoring responsibilities.

Should a model update trigger review?

Evaluate changes according to their significance. An update can affect assumptions, performance, data handling or control effectiveness.

Does certification automatically demonstrate compliance with AI legislation?

No. Determine applicable legal duties for the organisation’s role and use case separately.

Is every AI use equally risky?

No. Purpose, users, information, operating conditions and consequences affect the assessment.

Can the scope cover one business unit?

A defined boundary is possible, but public statements must not imply coverage of unassessed units or uses.

Explore AI management system certification

Find potential AIMS assessors, then confirm the exact service, relevant accreditation and team competence for your AI roles and use cases.

Explore Certification Bodies
View All Articles