An insurer uses a model to screen claims. When the hit rate suddenly drops, management asks which data the model was last trained on, which cloud region it runs in and who can retrain it. Nobody can answer all three questions. A.4 asks the organization to know and write down what each AI system depends on: components, data, tooling, computing power and people.
Resource documentation is the inventory behind the AI register. It makes dependencies visible before they turn into problems, and it gives the impact assessment the facts it builds on. Annex B.4.2 explicitly mentions aids such as data flow diagrams and architecture sketches for this purpose.
What does the standard require?
Annex A.4 of ISO/IEC 42001 contains five controls. The first asks for the documentation as a whole, and the other four go deeper into one type of resource each. The implementation guidance sits in Annex B.4.
- Document resources per life cycle stage (A.4.2). The organization works out which resources its activities need in each stage of an AI system and in other AI-related work, and records them. These include system components, data, tooling, system and computing resources, and people with their expertise. Resources may come from inside the organization, from customers or from third parties. If the documentation shows that something is missing, the organization adjusts the system’s specification or its deployment requirements.
- Data resources (A.4.3). The documentation covers, among other things, provenance, the date of the last change, for machine learning the split into training, validation, test and production data, the data categories, the labelling procedure, the intended use, quality, retention and disposal rules, known or potential bias, and data preparation.
- Tooling resources (A.4.4). This means algorithm and model types, tools for conditioning data, optimization and evaluation methods, provisioning tools, aids for model development, and the software and hardware used for design, development and deployment.
- System and computing resources (A.4.5). Records cover, for example, resource requirements, so that it is clear whether the system can run on constrained devices, the location (on premises, cloud, edge), processing including network and storage, and the effect of the hardware on the environment and on cost. Development, deployment and operation may each need different resources.
- Human resources (A.4.6). The organization documents the people and competences that carry the system, from development through deployment, operation, change and maintenance to transfer and decommissioning, plus verification and integration. It considers the need for diverse expertise, such as data science, oversight roles, specialists in security and privacy, AI research and domain knowledge.
In practice
One profile per AI system. Add a resource section with the five categories to every entry in the AI register. For a bought-in language model in customer service, a few lines often suffice: provider and model version, data centre region, the knowledge sources fed into it and the responsible business owner. A model trained in-house needs considerably more, above all on the data side.
Treat data sets as assets in their own right. Every data set gets an entry with provenance, usage rights, last update, category and owner. Write down known bias explicitly, even if it has not been fixed yet. A documented weakness can be treated, while an unknown one catches you by surprise in production.
Make third-party dependencies visible. Mark which tooling and computing resources come from suppliers: cloud platform, pre-trained model, libraries, labelling service providers. These entries are the basis for supplier management under A.10 and for a fallback plan in case a service is discontinued.
Name competence gaps early. Compare the competences each system needs with what is available internally. Where nobody can validate a model or judge its outputs from a domain perspective, record the gap and carry it into training plans or into engaging outside support. If a resource stays unavailable, the standard expects the design or the deployment conditions to be adjusted.
Typical audit evidence
For A.4, auditors typically expect:
- Resource section per AI system with components, data, tooling, computing resources and people (→ AI register in the AI Compliance Starter Kit)
- Data flow diagrams or architecture sketches for the main systems
- Data set descriptions with provenance, last update, category, intended use, quality, retention and known bias
- Principles on resources in the policy (→ AI policy, section 9)
- Resources and security of supply where the organization is a provider (→ AI provider policy, section 23)
- Competence overview with the roles per system and the status of training (→ Competence matrix)
- Link to the asset register for assets already managed in the ISMS (→ Asset register)
KPI
Share of AI systems in the AI register with complete resource documentation in all five categories
Complete means every category is filled in or marked as not applicable with a reason. The target is 100%. The figure usually drops for newly added systems and after a model change, when the details are not updated.
Supporting KPIs:
- Share of data sets with documented provenance and settled usage rights
- Number of systems with an open competence gap
- Months since the last review of the resource details per system (target: no more than 12)
Link to the AI Act
For providers of high-risk systems, A.4 largely overlaps with the technical documentation under Article 11 and Annex IV of the AI Act (Regulation (EU) 2024/1689), which describes among other things the system architecture, the computing resources used and the data requirements. Article 10 requires data governance for training, validation and test data, which is hard to demonstrate without documented data resources. Under Article 13(3)(e), the instructions for use state the computational and hardware resources needed and the expected lifetime.
For deployers, the human side matters most. Article 26(2) requires people with competence, training and authority for human oversight, and Article 4 obliges organizations to support the AI literacy of their staff. A.4.6 provides the overview of which competences each system needs (→ EU AI Act).
Related controls
A.4 provides the inventory that other controls draw on:
- A.3 Internal organization: assigns responsibility for assets and resources to named roles.
- A.5 Assessing impacts of AI systems: uses the resource documentation as its factual basis.
- A.7 Data for AI systems: goes deeper into data resources through acquisition, quality, provenance and preparation.
- A.5.9 Inventory of Information and Other Associated Assets: the ISO 27001 counterpart the AI register is linked to.
Sources
- ISO/IEC 42001:2023, Annex A.4 and Annex B.4: Resources for AI systems; clause 7.1
- Regulation (EU) 2024/1689 (AI Act), Articles 4, 10, 11, 13 and 26 and Annex IV
- ISO/IEC 22989:2022, AI concepts and terminology, named in Annex B.4.5 as a source on system resources
- ISO/IEC 42001 in the Cenedril wiki: structure of the standard and its relation to the AI Act