OSFI Guideline E-23: Model Risk Management and AI
OSFI's revised Guideline E-23 brings model risk management to every model a Canadian bank or insurer uses, including AI and machine learning, from 1 May 2027. This note explains its outcomes and principles, inventories and risk ratings, lifecycle controls, what it means for AI, and what vendors of AI models should expect.

Key takeaways
- E-23 was published on 11 September 2025 and takes effect on 1 May 2027.
- It applies to all FRFIs on a risk basis, including foreign branches.
- Institutions must inventory all models, including vendor and third-party models.
- Risk ratings drive the intensity of validation, monitoring and controls.
- AI/ML risks such as explainability, bias, drift and autonomy are expressly addressed.
What OSFI Guideline E-23 on model risk management requires
OSFI's revised Guideline E-23, Model Risk Management, published on 11 September 2025 and effective 1 May 2027, sets enterprise-wide expectations for how federally regulated financial institutions (FRFIs) develop, validate, deploy, monitor and retire models, including artificial intelligence and machine learning (AI/ML) models. It applies to all FRFIs, including foreign bank and foreign insurance branches to the extent consistent with their Canadian business, on a risk basis proportional to each institution's size, strategy, risk profile, complexity and interconnectedness.

The previous version of E-23 focused mainly on models used for regulatory capital at deposit-taking institutions. The 2027 version is broader. OSFI observes that institutions increasingly rely on models to support or drive decisions, including in business areas that did not traditionally use them, and that models now draw on more diverse data and complex techniques that heighten model risk. AI/ML is central to the update: the guideline addresses explainability, bias, autonomy, data quality and drift, and it expressly covers models sourced from vendors and other third parties.
E-23 identifies three expected outcomes of effective model risk management: model risk is well understood and managed across the enterprise; model risk is managed using a risk-based approach; and model governance covers the entire model lifecycle.
Enterprise governance and model inventories
Under Principle 1.1, effective reporting structures and proper resourcing should enable sound model governance. Consistent with OSFI's Corporate Governance Guideline, senior management is responsible for defining roles, accountabilities and expectations for model risk management across the enterprise and for ensuring appropriate resourcing. The framework should set out model inventory and risk rating requirements and lifecycle governance, provide transparent and consistent reporting of model risk at different levels of the enterprise, be reviewed periodically as technology evolves, and cover models or data sourced externally, including from third-party vendors under Guideline B-10.
Principle 1.3 adds a use test: models should be appropriate for their business purposes, and institutions should deploy them only where they meaningfully contribute to decision-making, risk assessment or other functions.
Principle 2.1 requires institutions to identify and track all models in use or recently decommissioned. They need defined processes to periodically identify models across the enterprise, including vendor and third-party models, for example through surveys of business units. For many FRFIs, building a complete inventory is the largest single task, because AI tools embedded in software, spreadsheets with complex calculations and analytics in marketing or HR may all qualify.
| E-23 principle | Requirement | Practical implication |
|---|---|---|
| 1.1 | Reporting structures and resourcing for model governance | Board and senior management oversight; dedicated model risk function |
| 1.3 | Models appropriate for business purposes | Use-case approval before deploying a model |
| 2.1 | Identify and track all models, including vendor models | Enterprise model inventory, refreshed periodically |
| 2.2 and 2.3 | Model risk rating; intensity commensurate with risk | Tiering that drives validation depth and monitoring frequency |
| 3.2 | Data suitable for intended use | Data lineage, quality and bias testing, including synthetic data |
| 3.3 and 3.4 | Development standards; independent review | Documentation standards; validation independent from developers |
| 3.5 and 3.6 | Controlled deployment; monitoring and decommissioning | Change control; drift monitoring; retirement process |
Risk rating and the model lifecycle
Principle 2.2 requires a model risk rating approach that assesses key dimensions of model risk, based on inherent risk reflecting the model's vulnerabilities and the materiality of its impacts, so that the most critical or complex models receive enhanced scrutiny. Under Principle 2.3, the scope, scale and intensity of model risk management should be commensurate with that risk: the inherent model risk rating drives the frequency, intensity and scope of lifecycle activities. Factors OSFI highlights for rating include the need for explainability, the need for alternative controls for "black box" or autonomous models, and the potential for biased outcomes, negative social and ethical implications, or privacy risks.
The lifecycle principles then set expectations at each stage:
- Data (3.2). Data used to develop a model should be suitable for its intended use. OSFI notes that data may be structured, semi-structured or unstructured, synthetic or derived from proxies, and that flawed data, for example through inaccuracies, bias or missing records, has significant consequences.
- Development (3.3). Clear standards for performance and documentation, including, for AI/ML, proper training and parameter optimization.
- Independent review (3.4). A process, independent from development, to assess conceptual soundness and performance and confirm that models are properly specified, working as intended and fit for purpose. Internal reviewers or objective third parties may be used.
- Deployment (3.5). Deployment in an environment with quality and change controls, preserving model integrity in production, which OSFI notes is particularly important for AI/ML.
- Monitoring and decommissioning (3.6). Defined standards for monitoring, including detecting drift and performance breaches, and for retiring models.
Start with the inventory and the risk rating methodology. Without them, validation and monitoring cannot be prioritized. Include generative AI tools and vendor-embedded models from the outset, because they are the models most likely to be missed and most likely to attract supervisory questions after 1 May 2027.
What E-23 means for AI and machine learning
For AI/ML, model risk management under E-23 raises particular challenges. Complex models may be hard to explain, which affects both validation and the ability to justify decisions to customers and supervisors. Their performance can drift as data changes. They may produce biased outcomes in credit, insurance underwriting, fraud detection or marketing. And autonomous or agentic systems can take actions with limited human intervention. E-23 does not ban any technique, but it expects controls proportionate to these risks: stronger explainability requirements where decisions affect customers, compensating controls such as human review or output limits for black-box models, fairness and bias testing, and more frequent monitoring.
E-23 operates alongside other rules. Privacy law governs the personal information used to train and run models, and the federal privacy reform bill, Bill C-36, was before Parliament without royal assent as of the latest LEGISinfo record. Human rights and consumer protection laws apply to discriminatory outcomes. OSFI has also commented on international work on responsible AI adoption, and FRFIs should expect supervisory attention to AI governance well before the effective date.
What vendors of AI models to banks and insurers should expect
E-23 is addressed to FRFIs, but vendors will feel its effects through procurement and contracts, reinforced by Guideline B-10 on third-party risk and B-13 on technology and cyber risk. Vendors supplying AI models, scoring engines, fraud tools or AI-enabled software to Canadian banks and insurers should expect requests for:
- model documentation describing purpose, methodology, training data, limitations and known risks;
- performance metrics and testing results, including bias and fairness testing where relevant;
- information enabling the FRFI's independent validation, or third-party validation reports;
- notice of material model changes, retraining or new versions, and change control commitments;
- monitoring outputs, drift alerts and incident notification;
- data provenance, data location and use restrictions, including whether client data trains shared models; and
- audit and access rights, and exit and transition assistance.
Vendors that cannot explain their models in enough detail for validation may find their products rated as higher risk, subject to restrictions on use, or excluded. Building a "model card" and validation support package is becoming a commercial necessity for selling into Canadian financial institutions.
Illustration (hypothetical): a Canadian bank uses a vendor's machine learning model to prioritize fraud alerts and a generative AI tool to draft customer emails. Under E-23, both go into the model inventory. The fraud model is rated high-risk because of its customer impact, so it receives independent validation and monthly drift monitoring, and the contract requires notice of retraining. The email tool is rated lower-risk but subject to output controls and human review.
Roles, documentation and evidence
Supervisors assess model risk management largely through documents and records, so the framework needs to produce evidence as a by-product of everyday work. Most institutions organize roles around three groups:
- Model owners and developers in the business or analytics teams, who define the use case, build or procure the model, document it, and monitor its performance in production.
- An independent model risk function, typically in the second line, which maintains the inventory and rating methodology, performs or commissions validation, sets standards and reports aggregate model risk to senior management.
- Internal audit, which tests whether the framework is designed and operating effectively.
For each model, the core documentation set usually includes a statement of purpose and intended use, a description of methodology and data, assumptions and limitations, development testing results, the risk rating with its rationale, validation reports and findings, approval records, monitoring thresholds and results, and change logs. For AI/ML models, institutions are adding records of training data sources, feature selection, explainability analysis, fairness testing and human-in-the-loop controls.
Issue management is the other half of the evidence. Validation findings should be logged, rated, assigned and tracked to closure, with escalation where deadlines slip. Models with open high-severity findings may need restrictions on use or compensating controls until they are fixed. Senior management reporting should show trends in inventory size, risk ratings, overdue validations and open findings, so that model risk can be compared with the institution's risk appetite.
Finally, the framework should cover decommissioning. Retired models should be removed from production, with records retained, and recently decommissioned models should remain visible in the inventory, as Principle 2.1 contemplates, so that their legacy effects on decisions and data can be traced.
Preparing for 1 May 2027
With the effective date approaching, FRFIs should be well into implementation. A typical roadmap includes a gap assessment against each principle, an updated model risk management policy approved by senior management, an enterprise inventory with risk ratings, a validation plan prioritized by risk, monitoring standards for AI/ML, revised vendor due diligence and contract templates, and targeted training for model owners, developers and business users. Smaller FRFIs can apply the guideline proportionately, but should document why their approach fits their size and risk profile, and revisit that rationale as their use of AI grows.
How TRW & Co helps with model risk management
We advise FRFIs, AI vendors and fintech partners on model risk management under OSFI Guideline E-23. That includes gap assessments, policy and framework drafting, inventory and risk-rating methodologies, governance of AI and generative AI use cases, vendor due diligence and AI contract terms, and alignment with B-10, B-13 and privacy obligations. Where an opinion or filing requires a Canadian-licensed lawyer, we work with licensed counsel. Our financial services regulatory and cybersecurity work covers related issues; see also our notes on OSFI Guideline B-10 and OSFI licensing. To discuss your program, contact us. This note is general information, not legal advice.
Frequently asked questions
When does the revised E-23 take effect?
On 1 May 2027. OSFI published the final revised Guideline E-23 on Model Risk Management on 11 September 2025. Institutions should use the time before the effective date to build inventories, risk ratings, validation plans and AI governance.
Does E-23 apply to insurers?
Yes. The revised guideline applies to all federally regulated financial institutions, including banks, trust and loan companies, life and fraternal insurers, property and casualty insurers, and foreign bank and insurance branches to the extent consistent with their Canadian business, applied proportionately.
Are generative AI tools models under E-23?
E-23 is designed to capture AI/ML models used to support or drive decisions, and it expects institutions to identify all models in use, including vendor models. Generative AI tools used in business processes should be assessed for inclusion in the inventory and risk-rated according to their use and impact.
Do vendors have to comply with E-23?
Not directly; the guideline applies to FRFIs. But FRFIs must cover vendor and third-party models in their frameworks and manage them under B-10, so vendors should expect requests for documentation, testing results, change notifications, monitoring data and audit rights in contracts.
How does E-23 address bias in AI models?
The guideline lists the potential for biased outcomes and negative social and ethical implications as factors in rating model risk, and highlights that flawed data, including biased data, has significant consequences. Higher-risk models therefore need data quality controls, bias testing and closer monitoring.
Sources
General information only, not legal advice. References to statutes, courts and authorities describe published law and practice at the date shown.


