A bank trains a fraud detection model on five years of transaction data. The data comes from three sources, has been cleaned twice and enriched once with purchased merchant data. When the supervisor asks why the model flags customers from certain postcode areas more often, nobody can say any longer which step introduced the bias. A.7 asks the organization to record the provenance, quality and preparation of its data well enough to answer any question of this kind.
Data determines what an AI system can do and where it fails. A.7 therefore works at five points: data management as a whole, acquisition, quality, provenance and preparation. The control goes beyond the inventory from A.4.3 and governs how the data itself is handled.
What does the standard require?
Annex A.7 of ISO/IEC 42001 contains five controls. The implementation guidance sits in Annex B.7.
- Data management for development (A.7.2). The organization defines, documents and applies processes for handling development data. Annex B.7.2 lists topics such as the privacy and security consequences of sensitive data, threats that arise from depending on data, transparency and explainability including provenance, how well the training data represents the later domain of use, and accuracy and integrity.
- Acquisition (A.7.3). The organization determines and documents how data is acquired and selected. Records cover, for example, the data categories and quantities needed, sources (internal, purchased, shared, open, synthetic), the nature of the source (static, streamed, gathered, machine generated), characteristics of the data subjects including possible bias, previous use, rights such as data protection or copyright, labelling metadata and provenance.
- Quality (A.7.4). The organization sets quality requirements and ensures that the data used for development and operation meets them. Where supervised or semi-supervised learning is used, the organization defines and measures the quality of each data split, from training through validation and testing to production, and improves it wherever it can. The organization considers how bias affects performance and fairness and adjusts model and data until both are acceptable for the use case.
- Provenance (A.7.5). A documented process records where data comes from for as long as the data and the system exist, from creation through updates and transformations to sharing and the transfer of control. Depending on source, content and context, the organization decides whether it also needs to verify the provenance.
- Preparation (A.7.6). The organization sets the criteria by which it selects preparation steps and the methods it uses. Typical steps are statistical exploration, cleaning, filling in missing values (imputation), normalization, scaling, labelling the target variables and encoding. Criteria and methods applied are documented for each AI task, because poor preparation can cause system errors.
In practice
Clear the rights before use. For each source, check whether the data may be used for training: legal basis and purpose compatibility for personal data, licence and copyright for third-party content, contract terms for purchased data. Anything unresolved stays out of the training set until the question is answered.
Check representativeness against the deployment. Compare the distribution of the training data with the people or cases the system will meet later. A model trained mostly on data from large cities can perform worse in rural areas. Document known gaps and decide whether to add data, reweight it or narrow the area of use.
Make quality measurable. Set criteria per data set, such as completeness, error rate in samples, agreement between labellers and currency. Check them before every training run and with every new data delivery. The same applies in production: if input data drifts noticeably from the training data, performance drops.
Keep preparation reproducible. Record the preparation steps as a versioned script or pipeline and keep the raw data. Derive the parameters for normalization and scaling from the training data alone and apply them unchanged to test and production data, so that no information from the test data leaks into training.
Typical audit evidence
For A.7, auditors typically expect:
- Data management procedure for development (→ AI provider policy in the AI Compliance Starter Kit, section 15)
- Datasheets per data set with source, rights, category, size and known bias
- Quality requirements and test results per data set and point of testing
- Provenance record with all processing steps and hand-overs
- Documented preparation with selection criteria and methods applied per AI task
- Data details in the technical documentation (→ Technical documentation template)
- Data-related risks such as data poisoning or lack of representativeness (→ AI risk register)
KPI
Share of training, validation and test data sets with documented provenance and a passed quality check
The target is 100% for every data set that went into an approved model. Data sets without proven provenance are a warning sign, because they leave later questions about rights and bias unanswerable.
Supporting KPIs:
- Share of data sources with a settled legal basis or licence
- Number of identified biases without a documented decision
- Deviation between training and production data for the most important features
Link to the AI Act
Article 10 of the AI Act (Regulation (EU) 2024/1689) is the direct counterpart for high-risk systems. It requires data governance practices covering, among other things, design choices, data collection and origin, preparation steps such as annotation, labelling, cleaning and enrichment, the examination of possible bias, and closing data gaps. Training, validation and test data have to be relevant, sufficiently representative and, as far as possible, free of errors and complete, and they have to take the intended setting of use into account.
The GDPR applies to personal training data, and copyright law to third-party content. Providers of general-purpose AI models need a copyright compliance policy under Article 53(1)(c) and publish a summary of the training content under point (d) (→ EU AI Act).
Related controls
A.7 reaches into several controls that work with data:
- A.4 Resources for AI systems: documents the data resources per system.
- A.6.1 Management guidance for AI system development: sets the rules for training data and data suppliers.
- A.6.2 AI system life cycle: includes the data details in the technical documentation and monitors drift in operation.
- A.5.34 Privacy and Protection of PII and A.8.33 Test Information: the ISO 27001 counterparts for personal data and test data.
Sources
- ISO/IEC 42001:2023, Annex A.7 and Annex B.7: Data for AI systems
- Regulation (EU) 2024/1689 (AI Act), Articles 10 and 53
- Regulation (EU) 2016/679 (GDPR)
- ISO/IEC 25024 (data quality), the ISO/IEC 5259 series (data quality for analytics and machine learning) and ISO 8000-2 (data quality vocabulary, including provenance), named in Annex B.7 as further sources
- ISO/IEC 42001 in the Cenedril wiki: structure of the standard and its relation to the AI Act