Data and AI provenance

Data stewards, model-risk teams and content owners need to know which dataset, transformation or model version contributed to an output. Provenance can improve investigation and reproducibility, but it cannot establish that a source was accurate, representative or lawfully collected.

Each provenance record captures dataset references, versions, transformation steps, responsible parties and permission evidence. Source material and personal data remain under strict access control, and fingerprints let teams compare artefacts without exposing them.

A reference workflow approves a source, records a transformation, associates an output with its inputs and routes exceptions for review. Include missing lineage, revoked permission and changed datasets in evaluation tests. Decide whether a conventional metadata database already meets the requirement.

For AI agents, each workflow identifies the responsible person and service identity, limits permissions to the task and requires human approval for consequential actions. Signed action records capture the authorisation, model or policy version, inputs by reference and the result, without exposing sensitive content, so operators can reconstruct who authorised what.

Provenance makes AI workflows inspectable and accountable. It complements, rather than replaces, human review and task-specific evaluation of model quality and bias.