Purpose
Reduce privacy, security, and compliance risk by requiring teams to justify every data attribute used for AI and to strip unnecessary data before storage or processing.
Scope
Applies to collection, ingestion, transformation, and retention of any personal or sensitive data used for AI development, testing, or inference, regardless of source.
- Raw databases, feature stores, and derived datasets
- Temporary staging areas, notebooks, and scratch space
- Vendor-supplied datasets integrated into AI pipelines
Definitions
- Essential Attribute: Data element proven necessary to achieve a defined AI outcome.
- Minimization Review: Structured assessment validating scope, data volume, and retention.
- Pseudonymization: Processing that reduces identifiability while maintaining analytical utility.
Policy
Teams must collect the minimum data required for the approved AI purpose, strip extraneous attributes before storage, and enforce retention schedules. Any request to retain additional fields must include a documented minimization justification and privacy review.
Roles and Responsibilities
Privacy Engineering maintains minimization checklists and tooling. Data Stewards execute minimization reviews. Product Owners confirm business necessity. Security validates technical controls (masking, tokenization).
Procedures
For every dataset used in AI workflows:
- Map data elements to the business purpose and mark optional vs. essential.
- Apply transformation (masking, hashing, aggregation) to non-essential elements.
- Enforce column- and row-level filtering at ingestion.
- Document retention period and implement automated purge jobs.
- Review minimization justifications annually or when the AI purpose changes.
Exceptions
Temporary retention of optional attributes for exploratory analysis requires DPO approval and must be purged within 30 days unless formalized into the data use register.
Review Cadence
Quarterly audits sample datasets to verify minimization controls. Findings feed into remediation plans tracked by Privacy Engineering.
References
- EU AI Act Article 10 (Data and data governance)
- GDPR Article 5(1)(c) (Data minimization)
- Internal documents: Data Minimization Checklist, Retention Catalog, Masking Standards