← All themes

Data Governance across ISO 42001, NIST AI RMF and the EU AI Act

// theme · data-governance

Data Governance

Open in explorer →

Training/validation/test data quality, bias, provenance, representativeness.

// Do once → satisfies all three
ONE dataset datasheet per training/validation/test set - source, licence, consent, representativeness, bias check, drift baseline.

Art.10 prescribes the substance; ISO A.7 demands the controls; NIST MAP/MEASURE want the evidence. A single datasheet is the artefact that proves all three.

ISO 42001
Annex A.7
NIST AI RMF
MAP 2.3 · MEASURE 2.10 · MEASURE 2.11
EU AI Act
Art.10
// Evidence auditors expect
  • Dataset datasheet / data card (source, licence, consent, date)
  • Bias and representativeness analysis per protected attribute
  • Data-quality metrics (completeness, duplicates, label noise)
  • PII inventory + lawful basis for training data
// Common pitfalls
  • Assuming 'we bought the data so it's fine' - EU Art.10 demands representativeness and bias examination regardless of source.
  • No record of training-data lineage when a regulator asks 12 months later.
  • Bias check only at launch, never re-run after retraining.