What does OSFI Guideline E-23 say about AI in banking?

What are the key points of OSFI Guideline E-23?

• Three outcomes: (1) model risk is well understood and managed across the enterprise, (2) model risk is managed using a risk-based approach, (3) model governance covers the entire model lifecycle. Principles are numbered 1.1 to 1.3, 2.1 to 2.3 and 3.1 to 3.6.
• Section A.4 defines a model as 'an application of theoretical, empirical, judgmental assumptions or statistical techniques, including AI/ML methods, which processes input data to generate results'. OSFI kept the definition deliberately broad after stakeholders asked for it to be narrowed.
• Principle 2.1 requires institutions to identify and track all models in use or recently decommissioned, including vendor and third-party models; only models with non-negligible inherent risk go into the model inventory, and Appendix 1 lists the minimum fields (for example model ID, risk rating, owner, developer, origin, version, dependencies, data sources, limitations, next review date).
• Principle 2.2 requires a model risk rating based on inherent risk, using quantitative and qualitative factors including 'model complexity or level of autonomy', reliability of data inputs and customer impacts; externally developed models are rated on a standalone basis.
• Principle 2.3 ties the frequency and intensity of review, documentation, approval authority, monitoring and re-rating to the risk rating, and says the 'extensive use of advanced AI/ML techniques should have correspondingly mature governance and oversight'.
• Principle 3.2 on model data requires data that is accurate, relevant and representative, compliant, traceable and timely, and flags that AI/ML models 'can easily mirror unwarranted data relationships'; Principle 3.3 requires explainability requirements that vary with purpose, autonomy, regulation and customer impact.
• Principle 3.4 requires review independent of development, with specific attention to AI/ML methods, explainability and third-party models; Principle 3.6 requires monitoring for AI/ML 'autonomous decision making, autonomous re-parametrization, and the elevated potential for model drift', plus contingency plans and decommissioning standards.
• Third-party models fall under the MRM framework (Principle 1.2, 'pursuant to our Guideline B-10') and OSFI's response letter says institutions must also ensure third-party models receive validation and monitoring commensurate to model risk; OSFI declined to add a grace period for validating third-party model updates.

What did OSFI Guideline E-23 change for banks?

The 2017 Guideline E-23 covered enterprise-wide model risk at deposit-taking institutions. The 2027 version applies to all models at all FRFIs, names AI/ML in the model definition, adds explainability, bias, data-governance and self-learning-model expectations, and moves to a risk-rating and inventory approach with proportional application. Canada's banks now have a published date (1 May 2027) to have AI/ML models inventoried, rated, reviewed and monitored under a single framework. OSFI had consulted on a draft from 20 November 2023 to 22 March 2024 and lengthened the proposed implementation period from twelve months to May 2027.

What does OSFI Guideline E-23 require of banks?

OSFI Guideline E-23 (Model Risk Management, 2027) expects each federally regulated bank to run an enterprise-wide model risk management (MRM) framework from 1 May 2027. The framework must identify every model in use, including AI/ML and vendor models, keep an inventory of those with non-negligible inherent risk, assign each a risk rating that drives the depth of review, documentation, approval and monitoring, and govern the whole lifecycle from design and data through independent review, deployment, monitoring and decommissioning. The guideline sets three outcomes and principles 1.1 to 3.6, applies proportionally to the institution's size and complexity, and singles out explainability, bias, self-learning behaviour, model drift and third-party black-box models for specific attention. It is guidance, not statute, and OSFI's own letter says only models that carry risk to the institution need full lifecycle governance.

E-23 is the model risk anchor of OSFI's AI position. OSFI has no separate AI rulebook for banks: its September 2024 OSFI-FCAC risk report says its frameworks on model risk, third-party risk, cybersecurity and operational resilience are technology-neutral and already cover AI, and its 2026 technology risk bulletins on frontier and generative and agentic AI point institutions back to E-23, B-10, B-13 and E-21. The July 2026 bulletin explicitly refers institutions to E-23 for 'enterprise-wide model risk management' expectations where AI models heighten model risk.

Scope is the main change from the 2017 version. OSFI says it left the definition of a model 'intentionally broad' and, asked whether low-risk generative AI uses such as document summarisation can be excluded from high-risk requirements, answered that institutions are 'empowered to make risk-intelligent decisions' when setting model risk ratings. Only models carrying non-negligible risk must be stored in the inventory and subjected to full lifecycle governance, so the practical workload depends on the rating methodology the bank builds. On third-party models, OSFI's letter says institutions should follow B-10 principles and ensure third-party models receive validation and monitoring commensurate to their risk, and that an exceptions policy may permit limited, specific use before validation is complete.

E-23 sits alongside, not above, the other OSFI guidelines it cross-refers to: B-10 for third parties, and B-13 and E-21 for technology, cyber and operational risks assessed before deployment (Principle 3.5). Internationally, it is the Canadian counterpart of the PRA's SS1/23 and, like SS1/23, treats AI as a model-risk question inside one framework, but unlike SS1/23 it names AI/ML in the definition.

WHAT THIS MEANS IN PRACTICE

• Build the inventory first: survey every business line, including areas that never used models before, and capture vendor and embedded generative AI tools, then triage for non-negligible risk.
• Define a model risk rating methodology that includes autonomy, explainability needs, data reliability and customer impact, and document how low-risk AI uses are exempted and tracked.
• Set explainability and bias-testing standards per rating tier, and decide in advance how self-learning models are judged to have 'materially changed'.
• Bring vendor models into scope: validate and monitor them commensurate with risk, link the work to B-10 third-party reviews, and settle an exceptions policy for models used before validation is complete.
• Plan against 1 May 2027: run a gap assessment now, prioritise high-risk models, and include model-failure contingency plans and decommissioning standards in the framework.