Skip to content

    Guide · Rules & regulators · Consumers & small business

    OSFI Guideline E-23: model risk management explained

    Short answer

    OSFI Guideline E-23 sets out how federally regulated banks, insurers and trust and loan companies should manage the risk of their models, explicitly including AI and machine learning. Published September 11, 2025, it takes effect May 1, 2027. It expects model inventories, risk ratings and lifecycle oversight, but it is not a complaint channel for customers.

    Figures as of November 3, 2026.

    Guideline E-23 is the main rulebook OSFI uses to expect federally regulated banks and insurers to manage the risks of their models, including AI. It was finalized in 2025 and takes effect in May 2027. Here is what it covers, how it fits with OSFI's other AI work, and what it does and doesn't do for customers.

    Reviewed

    Key takeaways

    • OSFI published the final Guideline E-23 on September 11, 2025. It takes effect on May 1, 2027, so as of October 2026 it is not yet in force.
    • E-23 explicitly covers AI and machine learning, including explainability, "black box" models, autonomous decision-making and model drift.
    • Institutions are expected to keep a model inventory, rate each model's risk, and govern models across their whole lifecycle, from design to retirement.
    • E-23 protects customers indirectly by setting expectations for how institutions build and oversee models. It is not a complaint channel for individuals.
    The model lifecycle E-23 expects institutions to govern
    1. 1Design and development

      Define the purpose and build the model on data that is accurate, representative and traceable, with attention to bias.

    2. 2Independent review

      Someone separate from the builders checks that the model is sound and fit for its purpose.

    3. 3Approval

      The model is approved for use under the institution's model risk framework.

    4. 4Deployment

      The model goes into use in real processes, such as credit, fraud or pricing.

    5. 5Monitoring

      Performance is tracked over time to catch problems such as model drift.

    6. 6Decommissioning

      Models that are no longer fit for use are retired in a controlled way.

    ↻ The cycle repeats while the model is in use

    What is Guideline E-23?

    The Office of the Superintendent of Financial Institutions (OSFI) is the federal regulator that supervises banks and federally regulated insurers for safety and soundness. Its guidelines set out what it expects those institutions to do.

    A model is a method that turns data into an estimate, score or decision, such as a credit score, a fraud alert or an insurance price. Model risk is the chance that a model is wrong, misused or poorly understood, and that this causes losses or harm.

    OSFI published the final version of Guideline E-23 – Model Risk Management on September 11, 2025. It takes effect on May 1, 2027. As of October 2026, it is published but not yet in force, which gives institutions time to adjust their practices.

    Who does E-23 apply to?

    E-23 applies to federally regulated financial institutions. It does not apply directly to provincially regulated institutions, such as most credit unions, and federally regulated pension plans are not mentioned in its scope. According to OSFI, the institutions covered include:

    • banks;
    • foreign bank branches, to the extent consistent with OSFI Guideline E-4;
    • life and fraternal insurers;
    • property and casualty insurers; and
    • trust and loan companies.

    Proportional, not one-size-fits-all

    OSFI says requirements are proportional to each institution's size, strategy, risk profile, complexity and interconnectedness. A large bank with many complex models is expected to do more than a small, simple institution.

    How does E-23 treat AI and machine learning?

    E-23's definition of a model explicitly includes AI and machine learning methods, and the guideline cites the OECD definition of AI. In other words, an AI system that produces scores, predictions or decisions is a model in OSFI's eyes and falls under the same framework as traditional statistical models.

    The guideline addresses risks that are especially relevant to AI:

    • Explainability: whether the institution can understand and explain how a model reaches its outputs.
    • "Black box" models: complex models whose inner workings are hard to interpret.
    • Autonomous decision-making: models that act with little or no human involvement.
    • Model drift: when a model's performance changes because the world, or the data, has changed since it was built.

    Vendor models count too

    Because the framework covers third-party models, an institution that uses an AI model from an outside vendor is still expected to manage that model's risk. Buying a model doesn't hand off the responsibility.

    Mitey the yeti at a desk with a laptop and a checklist
    Rules like E-23 work behind the scenes, so your best everyday tools are still your own records and a clear complaint.

    What does E-23 expect on governance, inventory and risk ratings?

    E-23 sets out several core expectations. Each one is about knowing what models an institution has, how risky they are, and who is accountable.

    • Enterprise-wide governance, with senior management accountable and reporting to the board.
    • A model risk management framework aligned with the institution's risk appetite. It covers third-party models too, not only those built in-house.
    • A model inventory at the enterprise level for models with non-negligible risk.
    • Risk ratings based on inherent risk, combining quantitative and qualitative factors such as complexity, autonomy and customer impact. The rating sets how intense the oversight should be.

    Why customer impact matters

    Because customer impact is one of the rating factors, a model that affects people directly, such as one used in lending, would generally be expected to get closer oversight than a low-stakes internal tool.

    What does E-23 expect across the model lifecycle?

    E-23 expects governance to cover a model's whole life: design, independent review, approval, deployment, monitoring and decommissioning. The idea is that oversight doesn't stop once a model goes live.

    Data gets specific attention. OSFI expects data to be accurate, representative and traceable, with attention to bias. For AI, this matters a great deal: a model trained on data that doesn't reflect the people it will be used on can produce skewed results, even if nobody intended it.

    How does E-23 fit with OSFI's other AI work?

    E-23 sits alongside several reports. Those reports shaped the conversation, but they are not binding rules.

    OSFI and partner publications on AI
    PublicationDateWhat it is
    FIFAI report, "A Canadian Perspective on Responsible AI"2023-04-30Set out the EDGE principles: Explainability, Data, Governance, Ethics
    OSFI-FCAC Risk Report on AI uses and risks2024-09-24Joint report on AI adoption and risks at federally regulated institutions
    Guideline E-23 (final)2025-09-11Binding supervisory expectations for model risk, effective 2027-05-01
    FIFAI II final report2026-03-23Proposed the AGILE framework; states it is not regulatory guidance

    What does E-23 mean for you as a customer?

    E-23 is aimed at institutions, not individuals. It doesn't give you a new right to see a bank's model or to have a decision reviewed. Its protection is indirect: models that affect customers are expected to be inventoried, risk-rated, independently reviewed and monitored.

    It is reasonable to ask your bank plain questions about any decision that affects you: Was an automated model involved? Did a person review the outcome? What information about me was used, and how can I correct it if it's wrong? The bank may not share model details, but these questions help you understand what happened.

    If you think an automated decision was wrong, E-23 is not where you turn. Use your bank's complaint process and, if needed, the external ombudsman, as explained in how to complain about a bank AI decision. The Financial Consumer Agency of Canada (FCAC) supervises how banks handle complaints but does not resolve individual complaints. For questions about your personal information, see your rights when an algorithm denies you.

    For the wider legal picture, including the status of federal AI and privacy bills, see where Canada's AI law stands.

    From the Index: model-risk readiness for AI

    What the Canadian Bank AI Index shows on this topic, counted from the published data. Every number can be rechecked in the database.

    From the Index: model-risk readiness for AI: inputs, steps and result
    Institutions scored35Canadian Bank AI Index, 2026 Q4
    Supervised by OSFI19
    Say their model-risk framework covers AI8Each with a public source on its profile
    Share of all institutions8 of 35 (23%)ATB, BMO, CIBC, Coast Capital, Desjardins, National Bank, RBC, TD

    Open the scorecard →Computed November 1, 2026.

    What this means for you

    • You won’t see E-23 directly. Its protection is indirect: models that affect customers, such as those used in lending, would generally be expected to get closer oversight.
    • You can still ask plain questions about a decision: Was an automated model involved? Did a person review it? What information about you was used, and how can you correct it?
    • If you think an automated decision was wrong, use your bank’s complaint process and then the external ombudsman; see how to complain about a bank AI decision.
    • If you bank with a provincially regulated credit union, E-23 doesn’t apply to it directly, so check with its provincial regulator.

    Questions people ask

    Is OSFI Guideline E-23 in force?

    Not yet. OSFI published the final guideline on September 11, 2025, and it takes effect on May 1, 2027.

    Does E-23 cover AI?

    Yes. Its model definition includes AI and machine learning methods, and it addresses explainability, "black box" models, autonomous decision-making and model drift.

    Does E-23 apply to credit unions?

    E-23 applies to federally regulated financial institutions, such as banks, insurers and trust and loan companies. Provincially regulated credit unions are overseen by provincial regulators, so check with yours.

    Can I complain to OSFI about an AI decision?

    E-23 is not a consumer complaint channel. Start with your bank's complaint process, then the external ombudsman if the issue isn't resolved.

    Are the FIFAI EDGE and AGILE frameworks binding?

    No. They come from industry forum reports; the final FIFAI II report states it is not regulatory guidance. E-23 is the binding guideline.

    Educational information only, not financial, legal or privacy advice. Rules and bank features change; check the linked official pages and talk to your financial institution or a qualified professional about your situation.

    Keep reading

    Related guides

    Get the quarterly Canadian Bank AI Index update

    New scores, new deployments and rule changes, once a quarter from the MITEY Network. Unsubscribe any time.

    Educational information only, not financial, legal or privacy advice. Index scores measure public evidence, not service, safety or fairness. If you think you have been scammed, contact your financial institution right away and report it to the Canadian Anti-Fraud Centre at 1-888-495-8501. Last reviewed November 1, 2026.