Model governance is the framework of policies, roles, controls, and documentation that an organisation applies across the lifecycle of its machine learning and statistical models to manage their risk, ensure their fitness for purpose, and satisfy regulatory and ethical obligations. It establishes accountability for model development, validation, approval, deployment, and ongoing monitoring, supported by artefacts such as model cards, audit trails, and validation reports. By imposing oversight, documentation, and review gates, it ensures models remain accurate, fair, explainable, and compliant rather than opaque or unaccountable.
Overview
- Model governance matured first in regulated finance, where supervisory guidance on model risk management established formal validation and oversight expectations, and has since broadened to all AI-driven decision systems.
- It assigns clear ownership across the three lines of defence — model developers, independent validators, and internal audit — to avoid the conflicts of self-assessment.
- Governance gates require that models pass documented review before deployment and remain under continuous oversight afterwards.
- Emerging AI regulation, including risk-tiered regimes, is making structured model governance a legal as well as a prudential requirement.
Key aspects
Lifecycle controls
- Defined stage gates for development, Model Validation, approval, deployment, and decommissioning, each with required evidence.
Documentation
- Model Documentation and Model Cards capture purpose, data, limitations, performance, and intended use for review and reuse.
Independent validation
- Model Validation independently challenges assumptions, data, and performance, testing for robustness and Fairness.
Ongoing oversight
- Model Monitoring and periodic revalidation detect drift and degradation, with an Audit Trail recording decisions and changes.
Mechanisms
Model inventory
- A central register (often a Model Registry) records every model, its risk tier, owner, and approval status.
Review and approval
- Risk-tiered review boards approve high-impact models and set conditions, monitoring frequency, and escalation paths.
Explainability and challenge
- Explainability techniques support effective challenge by validators and meaningful accountability to affected parties.
Applications
- Credit and capital models — banks govern rating and provisioning models under prudential model-risk expectations.
- Healthcare AI — clinical models require validation, documentation, and monitoring before and after deployment.
- Hiring and lending decisions — governance enforces Fairness testing and adverse-impact review for consequential decisions.
- Generative AI controls — governing prompts, fine-tunes, and evaluations of large models extends the framework to new model classes.
- Regulated reporting — Audit Trail and documentation evidence satisfy auditors and supervisors, supporting Regulatory Compliance.