Model Governance
The processes, controls, and documentation frameworks ensuring AI models are developed, validated, deployed, and retired in a controlled and auditable manner.
Model governance is the operational implementation of AI governance specifically focused on the model lifecycle — from initial development through production deployment and eventual retirement.
Core components of model governance:
Model Inventory / Registry — A centralised catalogue of all models in use across the organisation, including metadata (purpose, owner, training date, version), performance history, and governance status. Without a registry, organisations cannot answer basic questions regulators ask: "What models do you use? Who owns them? When were they last validated?"
Validation and Approval Workflow — A defined process for independent model review before production deployment, including back-testing, stress testing, and sensitivity analysis. Typically involves a Model Risk function separate from model developers.
Change Management — Controlled processes for model updates, retraining, and parameter changes. Every change must be documented, tested, and approved before taking effect in production.
Performance Monitoring — Ongoing tracking of model accuracy, drift (PSI), fairness, and explanation quality in production. Automated alerts for threshold breaches.
Retirement and Archival — Controlled decommissioning of models, with full audit trail preserved for regulatory inspection after retirement.
Regulatory context: SR 11-7 (OCC/Federal Reserve) in the US, RBI FREE-AI in India, and SS1/23 (PRA) in the UK all require documented model governance frameworks for regulated financial institutions.