What an auditor, or Sekit's evidence engine, asks for.
AI dataset record: provenance, permitted use and quality
For each dataset an AI system uses, the record of where it came from, what the company is allowed to do with it (licence, consent, restrictions), whether it contains personal data, and the quality and preparation checks run before using it (cleaning, labelling, data split).
From the Sekit evidence catalog
In practice
Provenance means tracing data through every hand it passes, not only where a training set began: which company files got uploaded into which vendor tool, when, by whom, and whether that use still matches the original consent or client contract. A company that buys a customer list, merges it with support data, then feeds the result into a model has several handoffs to track. An auditor follows a single dataset from its source to where it ended up, a trained model or a vendor's tool, checking the AI dataset record shows the chain, not only the final state. The typical gap: data copied, merged or uploaded informally, with the record covering only the original source.
Common gaps
Company files get uploaded into a vendor tool's knowledge base, but nobody records when, by whom, or whether that still matches the original client consent.
If you train your own model, nobody tracks which dataset version trained which model version, so a quality problem cannot be traced to its origin.
Data provenance stops once a dataset is exported from its source system, losing the trail through any cleaning or transformation step that follows.
Questions your auditor will ask
Can you trace this dataset back to its original source?
The AI dataset record documents provenance for each dataset, including source, transformations, and where it ended up, whether that is your own model or a vendor tool's knowledge base.
What happens to the record when a dataset gets transformed or merged?
Each transformation is logged against the dataset record so the chain from original source to where the data ended up, a vendor tool's knowledge base or your own trained model, stays intact, not only the starting point.
Is the current use of this data still within its original authorization?
Provenance tracking ties data use back to what was originally permitted, flagging any use that has drifted from the original licence or consent basis.
How current is your inventory of data feeding AI systems?
The data inventory is updated whenever flows change, extended here to cover dataset versions and transformations used specifically for AI.
Where regulation demands it
NIS2 art. 12.4 (Asset inventory) must track assets as they currently exist, so a dataset that gets merged or transformed still needs tracking, not only its original source.
Related controls
Via the shared Sekit CSF topic, not the framework's own index.