Fintech

OSFI and AI model risk in Canada: what financial firms should know

The Office of the Superintendent of Financial Institutions (OSFI), Canada's federal prudential regulator, expects federally regulated financial institutions to manage the risk of the models they rely on, and its model risk guideline, E-23, is the central reference. AI and machine learning raise the same core questions as other models, with added attention to explainability, data and change. This is a Canadian regime; comparable expectations exist in other jurisdictions, such as model risk guidance from the Office of the Comptroller of the Currency (OCC) and the Federal Reserve in the United States and expectations from United Kingdom and European Union supervisors.

Published October 27, 2026 | For chief risk officers, chief information officers and heads of model risk and compliance

The model lifecycleStep 1: Identify (Inventory and risk rating). Step 2: Develop (Data, method, documentation). Step 3: Validate (Independent challenge). Step 4: Approve and deploy (Limits of use). Step 5: Monitor (Drift, overrides, incidents). Step 6: Retire (Decommission safely). The last step feeds back into the first.IdentifyInventory and riskrating1DevelopData, method,documentation2ValidateIndependentchallenge3Approve anddeployLimits of use4MonitorDrift, overrides,incidents5RetireDecommission safely6The model lifecycleStep 1: Identify (Inventory and risk rating). Step 2: Develop (Data, method, documentation). Step 3: Validate (Independent challenge). Step 4: Approve and deploy (Limits of use). Step 5: Monitor (Drift, overrides, incidents). Step 6: Retire (Decommission safely). The last step feeds back into the first.IdentifyInventory and risk rating1DevelopData, method, documentation2ValidateIndependent challenge3Approve and deployLimits of use4MonitorDrift, overrides, incidents5RetireDecommission safely6
Fig. 1 Monitoring findings return to identification and development, so model risk is managed across the whole life of a model.

What is OSFI, and who does it supervise?

The Office of the Superintendent of Financial Institutions supervises federally regulated banks, insurers, trust and loan companies and certain pension plans. Its guidelines set out supervisory expectations rather than prescribing exact methods, and institutions are expected to apply them in proportion to their size, complexity and risk.

What is Guideline E-23 about?

E-23 addresses how institutions govern and control the risks that come from using models, meaning methods that turn data into estimates or decisions. Its central idea is that model use creates risk of error or misuse, and that risk should be managed across the model's whole life. OSFI has revised the guideline over time and its scope has been extended to cover AI and machine learning. Read the current version, including its effective date and scope, directly on OSFI's website and confirm interpretation with counsel.

What does model risk management cover in general?

Lifecycle stageQuestion to answerTypical evidence
IdentificationWhich models and AI systems do we use?Model inventory with owners and risk ratings
DevelopmentIs the purpose, data and method sound?Design documents, data lineage, test results
ValidationHas someone independent challenged it?Validation report and issue log
Approval and deploymentWho agreed it is fit for use, within what limits?Sign-off, use limits, change control
MonitoringIs it still performing as expected?Performance, drift and override reports
RetirementHow do we decommission safely?Retirement plan and records

What is different about AI models?

  • Explainability. Complex models can be harder to explain, so firms need methods and documentation suited to the decision's importance.
  • Data dependence. Performance depends heavily on training data quality, representativeness and bias.
  • Change. Models that retrain or are updated by a vendor change behavior, so change control must define what triggers revalidation.
  • Third-party models. Vendor and foundation models may be opaque, so institutions need due diligence and contractual access to information.
  • Generative AI. Outputs can be inconsistent or invented, which calls for testing, guardrails and human review.

What practical steps can you take now?

  1. Build a model inventory that includes AI, spreadsheets with embedded logic and vendor tools.
  2. Rate each model by risk using materiality and impact, and scale controls accordingly.
  3. Assign ownership. Separate model developers, users and independent validators.
  4. Define validation standards for AI, including data, performance, stability, bias and explainability.
  5. Set monitoring thresholds and triggers for review or retraining.
  6. Document limits of use so models are not stretched beyond their purpose.
  7. Brief senior management and the board so oversight is informed.

How can the platform support model risk controls?

Many model risk tasks become easier when the data platform records evidence automatically. A model registry holds each version with its owner, purpose and approval, while a data catalog and lineage show what trained and fed each model. Feature stores keep training and production inputs consistent, and MLOps pipelines run validation tests on every change. Monitoring then reports drift, overrides and performance to the model risk team. For generative AI, evaluation suites, guardrails and human-in-the-loop review give validators something concrete to test. These capabilities can be built on AWS, Microsoft Azure or Google Cloud, and the institution's own model risk framework still defines what is required.

How does this connect to privacy and other regimes?

Model risk management sits alongside other obligations, including privacy law such as the Personal Information Protection and Electronic Documents Act (PIPEDA) and provincial equivalents in Canada, third-party risk, technology and cyber risk, and conduct expectations from other regulators. A single AI system may touch several, so coordinate teams instead of running separate reviews.

Because guidance is updated, confirm current requirements with OSFI and counsel, and treat this article as orientation only.

What are the common gaps?

  • The inventory covers statistical models but misses AI tools adopted by business units.
  • Validation checks accuracy at launch but not stability, bias or behavior under stress.
  • There is no clear trigger for revalidation when a vendor updates a model.
  • Documentation is weak, so validators cannot reproduce results.
  • Monitoring exists but is not reviewed or acted on.
  • Ownership is unclear between developers, business users and risk functions.

How should you treat vendor models?

Outsourcing a model does not outsource accountability. Ask vendors for documentation of purpose, data sources, testing and known limitations, and negotiate rights to information, notice of material change and support for your validation. Where a vendor cannot provide enough transparency, compensate with stronger input and output testing, tighter limits of use and more human review.

What would senior management and the board want to see?

A concise view of the number and tier of models in use, the highest-risk items, validation status and overdue reviews, material incidents or overrides, and planned changes. Reporting should highlight decisions that need attention, such as accepting a known limitation or approving a new use case. Training helps oversight bodies ask good questions of AI systems even when they are not technical specialists.

Where ARCHR fits

ARCHR can help institutions stand up a model and AI inventory, design validation and monitoring processes, and build the supporting data and logging so evidence is produced as part of normal operations. Work is designed to support alignment with an institution's own model risk framework, which it owns and is accountable for.

This article is for general information only and is not legal, financial or investment advice.

FAQ

Frequently asked questions

What is OSFI Guideline E-23?

It is OSFI's guideline on model risk management for federally regulated financial institutions. Its scope and effective dates have been updated over time, so read the current version on OSFI's website.

Does model risk management apply to AI and machine learning?

In general, yes. Models that use AI or machine learning raise the same questions of governance, validation and monitoring, with added attention to explainability and change. Confirm the current wording with OSFI.

We are not federally regulated. Does this matter?

Many firms adopt similar practices as sound governance, and clients or partners may expect them. Your own regulators set the binding requirements.

What is a model inventory?

A maintained list of models in use with owner, purpose, risk rating, data, validation status and review dates. It is the foundation for all other model risk controls.

Let's talk about your data and AI priorities

Tell us what you are trying to solve. We aim to reply within one business day.