What Annex A.4 Is Actually About
If you have been working through ISO 42001 and you reach Annex A.4, you might initially think it looks straightforward. Data documentation, tools, people. Surely every organisation already tracks these things?
On this page
In practice, Annex A.4 is one of the sections where AI management systems most commonly fall short during audits. The reason is not that organisations lack data or tools or staff. It is that they have not documented these resources in a way that connects back to their AI systems and the risks those systems carry. There is a meaningful difference between having resources and being able to demonstrate that those resources are appropriate, controlled, and understood.
This article walks through what Annex A.4 requires across its three main areas, what auditors look for when they examine this section, and where organisations tend to struggle. Whether you are an internal auditor preparing to assess your own AI management system or a lead auditor conducting a third party certification audit, understanding A.4 in depth will sharpen your approach.
For context on the broader internal organisation controls that sit alongside A.4, the article on internal organisation controls in ISO 42001 Annex A.3 covers the governance structures that A.4 builds upon.
The Three Areas of Annex A.4
Annex A.4 addresses three distinct but related resource categories for AI systems. Each one has its own documentation expectations and its own audit footprint.
- A.4.1 Data for AI systems
- A.4.2 AI system tools and infrastructure
- A.4.3 AI system human resources
The logic connecting all three is that an AI system is only as trustworthy as the resources that power it. Poorly documented training data introduces bias and traceability gaps. Undocumented tools create accountability blind spots. Unqualified or unaware staff introduce operational risk. Annex A.4 asks organisations to get these three resource categories under documented control.
Exemplar Global Certified Training ProviderRTP No. 310970A.4.1: Data for AI Systems
What the Control Requires
Control A.4.1 requires that the organisation establish and document policies and procedures related to the data used in its AI systems. This covers data acquisition, data quality, data provenance, and data management throughout the AI system lifecycle.
The intent is clear. AI systems are fundamentally data dependent. The outputs of an AI model are shaped by the data it was trained on, the data it is tested against, and the data it processes in operation. If that data is poorly understood, poorly documented, or poorly controlled, the AI system itself cannot be trusted.
ISO 42001 is not prescriptive about exactly how data documentation must look, but auditors will expect to see evidence that the organisation has addressed at minimum the following questions:
- What data sources feed each AI system?
- How is data quality assessed before use?
- How is data provenance recorded and maintained?
- Are there controls over data labelling, annotation, and preprocessing?
- How is data protected from unauthorised modification?
- Are there retention and disposal policies for training data?
Common Audit Findings in A.4.1
The most frequent finding in this area is that data documentation exists at a technical level inside the development team but has never been formalised into a policy or procedure that sits within the AI management system. A data scientist might have a notebook or a version controlled repository with data lineage information, but if that information is not accessible to those responsible for the AIMS and is not reviewed as part of governance, it does not satisfy A.4.1.
Another common issue is that organisations document their initial training data but fail to maintain records of data updates. AI systems that are retrained or fine tuned over time need ongoing data documentation, not just a one time record from initial development.
A third finding relates to third party data. Many AI systems use data sourced from external providers, public datasets, or data brokers. Organisations frequently cannot demonstrate that they have assessed the quality, provenance, or terms of use of these external data sources. This is a genuine gap that auditors should probe specifically.
What Good Looks Like
A well controlled organisation will have a documented data governance framework that applies to each AI system in scope. This framework will identify data sources, describe quality criteria, record how data is validated before use, and specify who is responsible for data management decisions. It will be reviewed when AI systems are updated or when new data sources are introduced.
The data documentation will be version controlled and linked to the relevant AI system records, so an auditor can trace from the AI system to its data and back again without having to chase multiple people across departments.
A.4.2: AI System Tools and Infrastructure
What the Control Requires
Control A.4.2 shifts focus from data to the tools and infrastructure used to develop, deploy, and operate AI systems. This includes development environments, training platforms, model management tools, monitoring systems, and any third party AI services or APIs that the organisation relies on.
The requirement is that the organisation identify, document, and manage these tools in a way that supports the objectives and risk profile of the AI management system. This is not simply an asset inventory exercise. It is about understanding whether the tools being used are appropriate for the intended purpose and whether their use introduces risks that need to be controlled.
Auditors will look for evidence that the organisation has considered:
- What tools and platforms are used at each stage of the AI lifecycle?
- Are third party tools and services assessed before adoption?
- Are there version control and change management processes for AI tooling?
- How are tool related risks identified and addressed?
- Are there controls over access to AI development and deployment environments?
The Third Party Tool Problem
This area of A.4.2 deserves specific attention because it catches many organisations off guard. A significant proportion of AI systems today are built on top of third party foundations. Organisations use commercial machine learning platforms, pre trained foundation models accessed via API, cloud based AI services, and open source frameworks maintained by external communities.
Each of these introduces a dependency that the organisation cannot fully control. If the third party changes its model, deprecates an API version, or introduces a bias in an update, the organisation's AI system is affected. Annex A.4.2 asks organisations to have visibility over these dependencies and to have considered the risks they create.
During an audit, a useful question is to ask the organisation to walk you through what would happen if a key third party AI tool became unavailable or changed its behaviour. If the answer is confusion or uncertainty, that tells you the tool dependency has not been adequately documented or risk assessed.
Connecting Tools to the Risk Treatment Plan
One of the more sophisticated things an auditor can do in this section is trace from the tool inventory to the risk treatment plan. If a particular tool is identified as critical infrastructure for an AI system that carries significant risk, there should be corresponding controls in the risk treatment. If those controls are absent, that is a finding worth raising.
The article on checking the Statement of Applicability against the AI risk treatment plan explains how this traceability works in practice and is worth reading alongside this section.
A.4.3: AI System Human Resources
What the Control Requires
Control A.4.3 addresses the people who work with AI systems. It requires that the organisation identify the roles involved in AI system development, deployment, operation, and oversight, and ensure that those individuals have appropriate competence for their responsibilities.
This control sits alongside but is distinct from the competence requirements in Clause 7.2 of ISO 42001. Clause 7.2 addresses the general competence requirements for anyone whose work affects the AIMS. A.4.3 goes further by requiring the organisation to specifically document the human resource requirements for each AI system and to demonstrate that those requirements are being met.
In practical terms, this means the organisation needs to be able to answer:
- What roles are involved in each AI system and what are the competence requirements for each role?
- How are individuals assessed against those requirements?
- What training or development is provided where gaps exist?
- Are there clear accountability structures for AI system decisions?
- How is human oversight of AI system behaviour maintained?
Why Human Resources Documentation Matters in AI
The reason ISO 42001 dedicates a specific Annex A control to human resources in AI is that AI systems introduce competence requirements that are genuinely different from traditional technology systems. Understanding how a machine learning model behaves, recognising signs of model drift, interpreting confidence scores, and making appropriate decisions about when to override or escalate AI outputs all require specific knowledge that general IT or business competence does not provide.
Organisations that deploy AI systems without ensuring that the people operating those systems understand their limitations are creating real risk. Annex A.4.3 is designed to make that risk visible and controlled.
For auditors, this is one of the more interesting areas to explore because it sits at the intersection of technical competence and organisational accountability. It is not enough for the data science team to be competent. The people who use AI outputs to make decisions, the managers who oversee AI system performance, and the executives who are accountable for AI risk all need appropriate understanding of the systems they are responsible for.
What Auditors Look for in A.4.3
When auditing A.4.3, look for a documented skills or competence framework that is specific to each AI system in scope. This should go beyond a generic job description. It should identify the particular knowledge and capabilities required to perform each AI related role safely and effectively.
Then look for evidence that the organisation has assessed its people against that framework. Training records, competence assessments, and qualification records are all relevant. If the organisation has identified gaps, look for evidence that those gaps are being addressed through training, recruitment, or interim controls.
One area that is often underdeveloped is the documentation of human oversight roles. Many organisations have technical staff who build and maintain AI systems but have not clearly documented who is responsible for ongoing oversight of AI behaviour in operation. This is a gap that A.4.3 is specifically designed to close.
The article on building AI competence under Clause 7.2 provides useful background on how competence requirements are established and evidenced in an AI management system context.
How A.4 Connects to the Rest of the AIMS
Annex A.4 does not stand alone. Its requirements connect directly to several other parts of ISO 42001 and the broader AIMS structure.
The data documentation in A.4.1 feeds into the AI impact assessment process. If the organisation cannot demonstrate that its training data is well understood and controlled, the impact assessment conclusions become unreliable. Similarly, the tool documentation in A.4.2 informs the risk assessment, because tools that are poorly understood or inadequately controlled represent operational risks that need to be addressed.
The human resources documentation in A.4.3 connects directly to the operational controls in Clause 8. If the people responsible for operating AI systems are not competent for their roles, the operational controls that depend on human judgement cannot be relied upon.
When auditing A.4, it is worth tracing these connections explicitly. Ask to see how the data documentation informs the impact assessment. Ask how tool risks are reflected in the risk register. Ask how competence gaps are addressed before staff are assigned to AI system roles. These traces reveal whether A.4 is genuinely integrated into the AIMS or whether it is a set of documents that were created for compliance purposes and then set aside.
Practical Audit Approach for Annex A.4
When planning your audit of A.4, consider structuring your approach around three questions for each control area.
- Is the requirement documented? Does a policy, procedure, or framework exist that addresses the control?
- Is the documentation implemented? Can the organisation show that the documented approach is actually being followed in practice?
- Is it effective? Does the documentation and its implementation actually reduce the risks it is intended to address?
For A.4.1, request the data governance documentation and then trace it to a specific AI system. Ask to see how data quality was assessed for the most recent training dataset. Ask what happened when a data quality issue was identified.
For A.4.2, request the tool inventory and then ask about the most recently adopted AI tool or service. How was it assessed before adoption? What controls govern its use? What would happen if it were discontinued?
For A.4.3, request the competence framework for AI roles and then select two or three individuals in those roles. Ask to see their competence assessments and training records. Ask what training they received before taking on their AI responsibilities.
This approach moves you from document review to implementation verification to effectiveness assessment, which is the right audit trajectory for any Annex A control.
Exemplar Global Certified Training ProviderRTP No. 310970Documented Information Requirements
ISO 42001 requires organisations to maintain documented information as evidence of conformance with Annex A controls. For A.4, this means retaining records of:
- Data governance policies and procedures, including version history
- Data quality assessments for each AI system
- Tool inventories and third party assessments
- Competence frameworks for AI roles
- Training records and competence assessments for individuals in AI roles
- Evidence of human oversight arrangements
During an audit, the absence of these records is itself a finding. An organisation cannot claim conformance with A.4 on the basis of verbal assurance alone. The standard requires documented evidence, and auditors should request it specifically rather than accepting descriptions of what the organisation does.
Training for Auditors Working in This Space
Auditing Annex A.4 effectively requires auditors who understand both the management system requirements and the technical realities of AI development and deployment. This is a relatively new competence area for most auditors, and the gap between what the standard requires and what auditors can confidently assess is real.
If you are building your capability in this area, the ISO 42001 Lead Auditor and Internal Auditor courses at Audit Workshop are designed specifically to develop practical auditing skills for AI management systems. The training covers the full Annex A control set, including A.4, and gives you the tools to audit data governance, tooling, and human resources in a way that goes beyond surface level compliance checking.













