Use this checklist when designing or reviewing a data product, platform, or AI application. Items should be applied in proportion to the sensitivity of the information and the impact of the use case.
Ownership and Governance
- An accountable business owner and data steward are named.
- Decision rights for definitions, quality thresholds, access, and change are documented.
- Critical business terms are defined in the enterprise glossary.
- Applicable policies and obligations are identified and mapped to controls.
- Exceptions are recorded with compensating controls and an expiry date.
Metadata, Lineage, and Quality
- Required metadata (owner, description, classification, freshness, quality status) is complete.
- Lineage is captured from source to consumption for critical information.
- Quality expectations are measurable and tied to business use.
- Observability covers freshness, volume, schema, distribution, and pipeline health.
- Incidents are routed to accountable owners and reviewed for systemic causes.
Security and Privacy
- Every human, workload, and agent identity is unique and authenticated.
- Access follows least privilege and is reviewed periodically.
- Data is classified at creation, and classification propagates to copies and derived data.
- Sensitive data is encrypted, and masked or tokenized where full values are not needed.
- Processing purposes are documented, and data use is minimized to the purpose.
- Retention and deletion apply to replicas, backups, logs, caches, indexes, and derived datasets.
- Non-production environments use masked or synthetic data where appropriate.
- Third-party processing, including AI services, is assessed for retention, location, and use of data.
AI-Specific Controls
- The AI use case, its users, and its consequences are documented and risk-assessed.
- Retrieval enforces user permissions before information reaches the model.
- Tools are narrowly scoped; write actions require stronger authorization and, where appropriate, approval.
- Authorization is enforced outside the model.
- Prompt injection and tool misuse are included in testing.
- Evaluation covers relevance, grounding, policy compliance, and failure behavior.
- Prompts, outputs, and traces are logged with appropriate protection and retention.
- A named owner can suspend the system if risk conditions are not met.