Separate training, evaluation and operational data.
A supplier's training information, your test examples and the data entered during use serve different purposes. Map those roles before asking for evidence. Identify which datasets or services your organisation controls and which depend on an external party.
Keep the business purpose and affected people visible. A technically valid data feed can still be unsuitable for the way a system is being used.
Use the applicable data requirements carefully.
Article 10 addresses data governance and quality for relevant high-risk systems, including suitability for intended purpose, provenance, preparation and bias-related review. It distinguishes systems involving model training from those using other techniques.
These requirements do not create a general permission to collect any information. Privacy and other applicable rules need their own assessment, and a governance record should identify the responsible reviewers.
Record quality and limitations in context.
Ask how data was obtained, which population or conditions it represents and what known gaps may affect the intended use. Document relevant transformations and evaluation assumptions. Avoid claims such as 'unbiased data' without explaining the evidence, context and limits of that conclusion.
- Source, purpose and responsible owner.
- Preparation and relevant quality checks.
- Known gaps, limitations and evaluation context.
- Access, retention and change responsibilities.
Make data questions actionable.
Assign each unresolved issue to someone who can obtain evidence or change the process. State how the issue affects the proposed use and what would be needed to resolve it. Where data comes from a vendor, connect the question to supplier review rather than leaving it as an unowned technical note.
EUAIC can help organise the evidence and decision record. Data cleansing, statistical evaluation, privacy assessment and technical controls require a defined scope and appropriate expertise; a completed record does not prove dataset fitness.