OSFI E-23 model risk management, built for AI and ML
OSFI Guideline E-23 sets supervisory expectations for how federally regulated financial institutions manage model risk across the full lifecycle, and the 2027 version folds AI and ML models directly into scope. Effective May 1, 2027. VerifyWise gives you the tiering, validation, monitoring and inventory to meet those expectations.
A principles-based, technology-neutral supervisory guideline. No fixed fines, gaps surface through OSFI supervision.
What is OSFI Guideline E-23?
OSFI Guideline E-23 is the Office of the Superintendent of Financial Institutions' supervisory framework for model risk management (MRM). The final 2027 version, published September 11, 2025, is principles-based and technology-neutral. It sets expectations for identifying, assessing, managing and monitoring the risk that models introduce across their full lifecycle.
The guideline was broadened from earlier deposit-taking-focused guidance to cover all models used by federally regulated financial institutions, and it explicitly names artificial intelligence and machine learning models. Model risk is defined as adverse financial, operational or reputational impact arising from the design, development, deployment or use of a model.
Full lifecycle
Design, review, deployment, monitoring, decommission
Technology-neutral
Same rigor for a regression or a large AI model
Aligns with model risk principles in the EU AI Act and SR 11-7.
Who needs to comply?
- Banks and foreign bank branches
- Trust and loan companies
- Life insurers and fraternal benefit societies
- Property and casualty insurers and insurance company branches
- Applied on a risk basis, proportional to each institution's size, strategy, risk profile and complexity
Excluded: federally regulated pension plans. Earlier E-23 drafts covered FRPPs, but the final version removed them. Plan administrators are directed to CAPSA Guideline No. 10 instead.
Scope covers all models with non-negligible risk, including those that fall outside regulatory approval.
Tier every model, then calibrate the rigor
E-23 expects every model to get a risk rating from quantitative factors (portfolio size, financial impact) and qualitative factors (complexity, data reliability, customer impact). Ratings drive how much governance each model gets and are reviewed on trigger events. AI and ML models typically warrant higher ratings and stronger controls.
Most intensive governance, most frequent monitoring, tightest revalidation triggers
Large portfolios, high financial impact, complex or AI/ML models
Proportionate validation and scheduled monitoring
Moderate impact, moderate complexity and data reliability
Lighter-touch validation, periodic review
Small portfolios, low impact, simple and well-understood models
VerifyWise tiers each model into one of three tiers by materiality drivers, and a tier increase can automatically open a revalidation task.
What E-23 expects you to do
Eight principles-based expectations, from an enterprise-wide framework to AI and ML controls.
Enterprise-wide model risk framework
Risk-based policies to identify, assess, manage and monitor model risk, proportional to size, complexity and interconnectedness.
Broad model definition covering AI and ML
Theoretical, empirical, judgmental and statistical techniques, including AI and ML, that turn input data into results.
Full model lifecycle governance
Five components across the life of a model: design, review, deployment, monitoring and decommission.
Model risk rating and tiering
An inherent-risk rating from quantitative and qualitative factors, refreshed on trigger events.
Independent validation and review
Validation independent from development, confirming sound specification and fitness for purpose.
Defined roles and accountability
Owner, developer, independent reviewer, approver and user, with senior management holding enterprise accountability.
Firm-wide model inventory
All non-negligible models with ID, owner, version, risk rating, data sources, dependencies, approved uses, limitations, review dates and decommission status.
Ongoing monitoring and AI/ML controls
Performance monitoring plus transparency and explainability, controls for black-box or autonomous models, bias, privacy and drift monitoring under multi-disciplinary governance.
The five lifecycle components
E-23 governs a model from its initial rationale through formal retirement. Controls and documentation run across every stage, and VerifyWise touches each one.
Design
Rationale, data and development documented.
VerifyWise: Captured in the model inventory.
Review
Independent validation before use.
VerifyWise: Six-section independent validation reports.
Deployment
Approval and roles recorded.
VerifyWise: Approval and roles recorded on the model.
Monitoring
Ongoing performance monitoring.
VerifyWise: Metric thresholds with breach severities.
Decommission
Formal retirement and last review tracked.
VerifyWise: Inventory tracks retirement and next review.
How compliance is enforced
E-23 is a principles-based supervisory guideline, not a statute, so it creates no fines or monetary penalties. Institutions are expected to meet these standards, and gaps are addressed through OSFI's normal supervisory process.
Supervisory findings
Findings and requirements to remediate raised through OSFI's supervisory process.
Heightened scrutiny
Increased reporting and closer supervisory attention where gaps persist.
Intervention / ratings action
In serious cases, OSFI intervention or ratings actions rather than a fixed fine.
The guideline does not prohibit any modeling approach. The cost of non-alignment is supervisory, reputational and remedial, not a fixed fine.
E-23 expectations, mapped to VerifyWise capabilities
Each E-23 expectation maps to a capability already built into the VerifyWise model risk management module. This helps you meet the guideline, it does not replace your own judgment.
| E-23 expectation | VerifyWise MRM capability |
|---|---|
| Model risk rating and tiering by inherent risk (quantitative and qualitative factors), refreshed on trigger events | Manual tiering into three tiers by materiality drivers; a tier increase automatically opens a revalidation task, keeping ratings current on change. |
| Independent validation and review, separate from development, confirming sound specification and fitness for purpose | Six-section validation reports (purpose and scope, conceptual soundness, data review, outcomes analysis and more) with evidence links and findings logged by severity and lifecycle stage. |
| Ongoing monitoring of model performance, with enhanced drift monitoring for AI and ML | Metric thresholds such as PSI and AUC with warn, high and critical breach severities, and breach actions to notify or notify and flag for revalidation. |
| Review triggered by performance breaches, data changes, material change and on a periodic basis | Revalidation triggers (breach, material change, tier increase, scheduled) auto-open validation tasks, recorded in an append-only audit log. |
| An inventory of all models with non-negligible risk, tracking usage, review dates and decommissioning | Model inventory capturing owner, version, risk rating, review status and dates, plus lifecycle state through to decommission. |
| Enterprise-wide accountability and oversight of the model risk framework by senior management | Per-tier attestation roll-up (blocked or ok) covering tiering current, validation coverage, monitoring active and open findings for oversight visibility. |
| Full-lifecycle governance across design, review, deployment, monitoring and decommission | Tiering, validation, monitoring, revalidation and inventory span the lifecycle from initial rationale through formal retirement. |
| Continuous, current model performance data feeding monitoring without manual re-entry | Machine-to-machine metric ingestion tokens push live metrics into monitoring against configured thresholds. |
Where most model risk programs fall short
Many teams meet the expectations on paper but run them on spreadsheets and documents. Here is where those setups break, and what closes each gap.
The inventory lives in a spreadsheet
A shared workbook goes stale between reviews, has no owner per row and cannot show a supervisor when a model was last touched.
A single model inventory with owners and metadata per model, kept current through machine-to-machine ingestion tokens rather than manual edits.
Validation reports sit in scattered documents
Word files and email threads make it hard to prove independence, find evidence or show that every required section was covered.
Six-section validation reports with evidence links and findings logged by severity and stage, produced independently of the model's developers.
Monitoring is manual or does not happen
Drift and performance decay are caught late, if at all, and there is no record of what threshold was breached or what was done about it.
Metric thresholds such as PSI and AUC with warn, high and critical severities and defined breach actions, so decay surfaces and is actioned on a schedule.
Revalidation depends on someone remembering
A material change or a breach should reopen validation, but without a trigger it waits for the next annual cycle.
Breach, material change, tier increase and scheduled triggers open validation tasks automatically, each written to an append-only log.
Tiering is inconsistent across teams
Different groups rate the same class of model differently, so validation effort does not track real materiality.
One tiering scheme by materiality drivers that sets validation depth, monitoring cadence and attestation expectations per tier.
Program health is a manual quarterly scramble
Pulling together coverage, overdue validations and open findings for the board or a supervisor takes days of spreadsheet work.
A per-tier attestation roll-up (blocked or ok) that reads current coverage, monitoring status and open findings on demand.
How VerifyWise supports OSFI E-23
The model risk management module covers tiering, validation, monitoring, revalidation, attestation and inventory in one audit-logged workflow.
Model tiering
Manually tier every model into one of three tiers by materiality drivers, giving you the risk rating E-23 expects and driving proportional governance.
Independent validation
Six-section validation reports (purpose and scope, conceptual soundness, data review, outcomes analysis and more) with evidence links and findings tracked by severity and lifecycle stage.
Ongoing monitoring
Metric thresholds such as PSI and AUC with warn, high and critical breach severities, and configurable breach actions (notify, or notify and flag for revalidation).
Revalidation triggers
Breach, material change, tier increase or scheduled triggers automatically open a validation task, with an append-only audit log.
Attestation roll-up
Per-tier attestation (blocked or ok) covering tiering current, validation coverage, monitoring active and open findings, so oversight can see readiness at a glance.
Inventory and metric ingestion
A model inventory plus machine-to-machine metric ingestion tokens so monitoring stays current without manual entry.
Every tiering decision, validation, breach and revalidation is timestamped in an append-only log, the reproducible trail E-23 documentation expectations call for.
E-23 puts AI and ML models directly in scope
The 2027 model definition explicitly names AI and ML methods, so machine learning, deep learning and generative models sit inside the lifecycle, tiering, validation, inventory and governance expectations.
Transparency and explainability
Black-box and autonomous models need controls for explainability and feature governance, which E-23 folds into validation.
Drift, bias and privacy monitoring
AI and ML models typically warrant higher ratings, with enhanced monitoring for drift, bias and privacy risks after deployment.
Multi-disciplinary governance
Reviewers must be fluent in ML failure modes, and vendor or foundation models stay the institution's responsibility to validate and control.
What E-23 ready looks like
Per-tier attestation rolls four checks into a single status you can show leadership and supervisors before a gap is found for you.
Tiering current
Every model tiered and dates fresh
Validation coverage
Required validations complete per tier
Monitoring active
Thresholds live and breaches actioned
Open findings tracked
Findings logged by severity and stage
When all four are green per tier, attestation rolls up to OK. Any gap shows as blocked before a supervisor finds it.
Frequently asked questions
Common questions about OSFI Guideline E-23 and model risk management.
Get E-23 ready before May 2027
Walk through your model inventory, tiering and validation gaps with the VerifyWise team.