Sovereignty and compliance

OSFI Guideline E-23 explained: why it matters for Canadian banks, insurers and lenders, and how DataNature Atlas helps

OSFI Guideline E-23 sets expectations for enterprise-wide model risk management at every federally regulated financial institution in Canada, effective May 1, 2027. It covers traditional, AI and machine learning models, including vendor and third-party models, across their whole lifecycle. DataNature Atlas helps institutions meet it by keeping a governed inventory of data, models and agents, with lineage from every decision back to source, access controls, PII masking, an AI gateway and a complete audit trail, all inside Canada.

Published . Last updated

On this page

What is OSFI Guideline E-23?

E-23 is the Office of the Superintendent of Financial Institutions' guideline on model risk management. OSFI published the final version on September 11, 2025, and it takes effect on May 1, 2027. It replaces the 2017 version, which applied only to deposit-taking institutions.

It is principles-based. OSFI does not prescribe tools. It expects institutions to understand, measure and control the risk that a model is wrong, misused or misunderstood.

Who does E-23 apply to?

All federally regulated financial institutions (FRFIs) in banking, financial services and insurance:

  • Banks and foreign bank branches
  • Life, property and casualty insurers, and fraternal benefit societies
  • Trust and loan companies

Federally regulated pension plans are excluded. Quebec's AMF has a closely aligned guideline for institutions it supervises, and provincially regulated credit unions often follow OSFI practice as a benchmark.

What counts as a model?

Almost anything that turns data into an estimate or decision with meaningful risk: credit scoring, fraud and AML detection, pricing, capital and actuarial models, forecasting, and explicitly AI and machine learning, including generative AI and large language models. Vendor models count too.

Why does E-23 matter now?

  • AI is everywhere. Fraud scores, credit decisions, chatbots and agents are models. Many institutions do not yet have a complete list of them.
  • Third-party models have no grace period. If a vendor's product embeds a model, or a fourth-party model inside it, the institution must still govern it under Guideline B-10 on third-party risk.
  • Evidence takes time. Inventories, lineage, validation records and monitoring cannot be created the week before an examination. With the effective date in May 2027, the work has to happen now.
  • Customers and boards are watching. A wrong or biased model is a conduct, reputational and financial risk, not only a technical one.

What are OSFI's expected outcomes?

  1. Model risk is understood and managed across the enterprise. Senior management has a single view of model risk.
  2. Model risk is managed with a risk-based approach. Every model is identified, rated and governed in proportion to its risk.
  3. Governance covers the whole model lifecycle: design, review, deployment, monitoring and decommissioning.

What does that mean day to day?

E-23 expectationWhat you need in practice
Model identification and inventoryA central, current list of every model, owner, purpose, risk rating and status
Data quality and suitabilityProof of where training and input data came from and that it is fit for purpose
Explainability and bias reviewRecords of how a model reaches decisions and checks on its training data
Independent review and validationThe ability to reproduce training data and results at a point in time
Controlled deploymentApprovals, versioning and access controls on who can change or call a model
MonitoringDrift and performance tracking, with alerts to owners
Third-party modelsThe same visibility and controls for vendor, open and hosted models
Audit trailWho did what, when, under which policy, with which data

How does DataNature Atlas help?

Atlas is the governance console of DataNature, the sovereign AI lakehouse. It runs inside your own deployment, so the evidence OSFI asks about stays under your control in Canada.

One inventory for data, models and agents

Atlas catalogs tables, columns, owners and tags alongside model endpoints and AI agents. Model owners, purposes and risk tiers can be recorded as tags and attributes, giving model risk teams one searchable inventory instead of spreadsheets.

Lineage from decision back to source

Atlas traces lineage from source systems to tables, features, models, dashboards and agent answers. When a validator asks "what data trained this credit model?" or an examiner asks "why was this customer flagged?", the path is there.

Reproducible training data

DataNature stores data in open Delta and Iceberg tables with time travel. Combined with experiment tracking and model versioning in DataNature AI, teams can rebuild the exact data and model version behind any decision for independent review.

Controlled access and PII protection

Role-based access, single sign-on, row and column rules and automatic PII detection and masking let developers and validators work on the data they need without exposing customer information they do not.

An AI gateway for generative AI and third-party models

The Atlas AI gateway controls which models are approved, logs prompts and responses, and sets spend budgets for model endpoints and agents. Open-weight models can be served inside your deployment, which reduces the third- and fourth-party exposure that E-23 and B-10 ask you to manage.

Monitoring and alerts

Model and data monitoring with alerts to owners supports the ongoing monitoring outcome. Scores, reasons and feature values can be kept with each decision, so explanations are available when a customer or examiner asks.

A complete audit trail

Atlas logs who accessed what, under which policy and when, across data, compute, BI and AI. Audit logs and catalog metadata stay in your jurisdiction, because governance metadata is sensitive data too.

What Atlas does not do

Atlas is a platform, not a model risk framework. Your institution still sets its model risk policy, risk ratings, validation standards and approval committees. Atlas gives those processes reliable data, controls and evidence.

Why does sovereignty matter for model risk?

E-23 asks institutions to control their models and the data behind them. That is harder when training data, prompts, logs and governance metadata sit with a provider subject to foreign law. DataNature keeps data, metadata, the control plane and AI inside Canada, on R2i Edwin or your own data centre. See the US CLOUD Act and Canadian data.

How do you get ready before May 2027?

  1. Find every model, including vendor and embedded AI, and record it in one inventory.
  2. Rate each model by risk and set the governance each tier needs.
  3. Connect lineage from source data to each high-risk model and its decisions.
  4. Make training reproducible with versioned data and models.
  5. Bring generative AI under control with approved models, prompt logs and budgets.
  6. Monitor and record performance, drift and decisions continuously.

Dlytica Consulting can run an E-23 readiness assessment on your current data and model estate. Talk to us.

Sources

Which DataNature features produce E-23 evidence?

E-23 expectationEvidence in DataNature
Model inventoryModel registry with versions and stages; Voir twins and their scopes listed in the platform
Data quality and lineagedbt tests on Silver keys, lineage from Bronze source to Gold metric in Atlas
ReproducibilityIceberg snapshots, so a model can be re-run on the exact data it was trained or validated on
Access controlRow filters, column masks and grants per group, with an audit log of access
Third-party and vendor modelsAI gateway routing per endpoint, so each external model is a named, governed dependency
MonitoringModel serving and experiment tracking; decisions written back to Iceberg for performance review
Human oversightAgent actions that change systems require approval and are logged

All of this runs in the institution's chosen Canadian deployment, so the evidence itself stays in Canada.

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.