# OSFI Guideline E-23 and AI agents: what Canadian banks and insurers need to do OSFI released the final version of E-23, its model risk guideline, on 11 September 2025. Banks and insurers have until 1 May 2027 to comply, and AI and machine learning models are clearly in scope. Here is what that means for the AI agents you are building now. ## Key facts - **Guideline:** OSFI Guideline E-23: Model Risk Management (2027), published 11 September 2025 - **Effective:** 1 May 2027 - **Applies to:** Every federally regulated bank and insurer, foreign branches included - **Scope:** All models, including AI and machine learning, across their full lifecycle - **Related rules:** OSFI B-13 Technology and Cyber Risk Management; PIPEDA; Quebec Law 25 ## What the rules mean for AI agents - **Model inventory:** Register every agent that makes or supports decisions, with its owner, purpose, data, tools and the foundation models it relies on. - **Risk rating:** Rate each agent by the materiality of its decisions and its level of autonomy, and scale controls to match. - **Lifecycle governance:** Cover design, review, approval, deployment, monitoring and retirement, including changes to prompts, tools and underlying models. - **Independent review:** Give validators the documentation and test results they need to challenge the agent's design and limits. - **Monitoring:** Track accuracy, drift, overrides and incidents in production, with thresholds that trigger review. - **Third-party models:** Apply the same discipline to vendor and foundation models, with due diligence and contractual rights to information. ## Readiness checklist - Add agents and copilots to the model inventory. - Define what counts as a model for agentic systems at your institution. - Rate each agent by decision materiality and autonomy. - Document purpose, data, tools, prompts and limits. - Test against prompt injection and excessive agency. - Keep human approval on material actions. - Set monitoring thresholds and review triggers. - Review third-party and foundation model dependencies. - Map controls to B-13 and privacy rules. ## FAQ ### Is an AI agent a model under E-23? In most cases, yes. E-23 applies to all models, including AI and machine learning, and an agent that makes or supports decisions with a model at its core should be inventoried, risk-rated and governed. ### Does E-23 apply to foundation models we buy from vendors? Yes. Third-party models remain in scope. You need enough information about them to assess and monitor the risk, even if you cannot see inside them. ### How does E-23 relate to B-13? E-23 covers model risk; B-13 covers technology and cyber risk. AI agents touch both, so their identity, access, logging and resilience controls should satisfy B-13 as well. ### What should we do before May 2027? Inventory your agents now, agree how you will rate them, and build documentation and monitoring into new agent projects so they are compliant on day one. ## Sources - [OSFI: Guideline E-23 Model Risk Management (2027)](https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027) - [OSFI: Letter on Guideline E-23](https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027-letter) - [OSFI: Guideline B-13 Technology and Cyber Risk Management](https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/technology-cyber-risk-management) - [Office of the Privacy Commissioner of Canada: PIPEDA](https://www.priv.gc.ca/en/privacy-topics/privacy-laws-in-canada/the-personal-information-protection-and-electronic-documents-act-pipeda/) - [Office of the Privacy Commissioner of Canada: Artificial intelligence](https://www.priv.gc.ca/en/privacy-topics/technology/artificial-intelligence/) - [OWASP Top 10 for LLM Applications (OWASP GenAI Security Project)](https://genai.owasp.org/llm-top-10/)