E-23

Implement E-23 Compliance using ​ValidMind.

Follow this detailed, step-by-step guide which will:

1. Walk you through the practical steps required to implement end-to-end compliance using ​ValidMind.

2. Link to supporting documentation and training for specific product features.

3. Focus on actionable implementation rather than just explaining requirements.

Overview

OSFI Guideline E-23 (Enterprise-Wide Model Risk Management for Deposit-Taking Institutions)¹ contains the most comprehensive treatment of AI among major MRM regulations, reflecting the 2027 effective date and the anticipated evolution of AI. By emphasizing outcomes over prescribed processes, E-23 ensures that institutions retain flexibility in how they achieve regulatory objectives.

¹ Office of the Superintendent of Financial Institutions:
Guideline E-23 – Model Risk Management (2027)

² Model risk management

This guide is organized around the expectations for federally regulated deposit-taking institutions in Canada to manage model risk effectively by covering traditional model risk management (MRM)² requirements while incorporating forward-looking considerations for AI and machine learning models.

Traditional MRM requirements

E-23 aligns with established MRM frameworks, requiring:

• Model inventory — Comprehensive registry of all models
• Development standards — Clear documentation of model design and implementation
• Validation framework — Independent validation and effective challenge
• Ongoing monitoring — Performance tracking and periodic review
• Governance — Board oversight and clear accountability

AI/ML-specific considerations

E-23 introduces enhanced requirements for AI and ML models in addition to traditional MRM requirements. ​ValidMind helps you address these considerations:

Explainability

Enhanced requirements for model interpretability ensure stakeholders can understand AI-driven decisions.

Steps

1. Configure documentation templates³ to capture explainability information.

2. Run interpretability tests to generate explanations.⁴

3. Document the level of explainability appropriate for model risk tier.⁵

4. Establish processes for explaining decisions to affected parties.⁶

³ Working with document templates

⁴ Run tests and test suites

⁵ Working with model documentation

⁶ Setting up workflows

Alternative controls

Compensating controls for “black box” approaches where full interpretability is not achievable.

Steps

1. Document the rationale for using less interpretable models.⁷

2. Implement enhanced monitoring⁸ as a compensating control.

3. Establish human oversight requirements.

4. Configure alerts for unexpected model behavior.

⁷ Working with model documentation

⁸ Ongoing monitoring

Bias assessment

Evaluation of ethical implications and discrimination risks in AI models.

Steps

1. Run fairness and bias tests⁹ during development.

2. Document bias assessment methodology.¹⁰

3. Establish thresholds for acceptable bias levels.

4. Implement ongoing bias monitoring.¹¹

⁹ Run tests and test suites

¹⁰ Working with model documentation

¹¹ Ongoing monitoring

Autonomous decision-making

Governance for self-learning capabilities and automated decisions.

Steps

1. Document the scope of autonomous decision-making.

2. Establish human oversight requirements.

3. Configure review workflows¹² for autonomous decisions.

4. Implement safeguards and intervention mechanisms.

¹² Setting up workflows

Model drift

Monitoring for performance degradation over time.

Steps

1. Configure ongoing monitoring¹³ for drift detection.

2. Set up alerts for performance degradation.

3. Establish thresholds for re-validation triggers.

4. Document drift monitoring methodology.

¹³ Ongoing monitoring

Dynamic learning

Management of continuously updating models.

Steps

1. Document model update frequency and methodology.

2. Establish change control for model updates.

3. Configure validation requirements¹⁴ for dynamic models.

4. Implement rollback procedures.

¹⁴ Manage validation guidelines

Implementation checklist

Use this checklist to track your E-23 implementation progress:

• Model inventory established with AI/ML classification
• Documentation templates configured for AI-specific requirements
• Validation workflows include AI/ML considerations
• Explainability requirements documented by risk tier
• Bias assessment processes implemented
• Ongoing monitoring configured for drift detection
• Human oversight controls in place
• Dynamic learning governance established