Description
Managed changes to AI systems including model updates, data changes, and configuration modifications with impact assessment and testing.
Implementation Guidance
Change Classification
Classify AI system changes by risk: routine (hyperparameter tuning, minor bug fixes), significant (model retraining, new data sources, feature changes), and major (architecture changes, new use cases, scope expansion). Each category requires different approval levels and testing depth.
Impact Assessment
Before implementing changes, assess impact on: model performance, fairness metrics, security posture, regulatory compliance, and downstream systems. Document the assessment and obtain approval from the appropriate authority.
Change Execution
Follow a controlled change process: document the change request, perform impact assessment, execute in non-production environment, run regression tests (accuracy, fairness, security), obtain approval, deploy with monitoring, and validate post-deployment. Maintain full audit trail.
Evidence Requirements
- Change request records with classification
- Impact assessments for significant and major changes
- Regression test results pre- and post-change
- Change approval records
- Post-deployment validation reports