E-23 implementation: six pillars to turn requirements into practice

OSFI Guideline E-23 takes effect next year. For federally regulated financial institutions in Canada, the requirements are no longer news. The real question is how to put them into practice.

In our recent webinar, The E-23 countdown: turning MRM and AI governance requirements into an implementation plan, Atif, responsible for Asia Pacific at Yields, and Efrem, Head of Pre-Sales at Yields, looked at E-23 through an implementation lens. Both have a background in model risk management and modelling within banking, and both follow MRM regulation across Canada, the US, the UK, Europe and Singapore.

This article summarises their key points. Download the full recording here.

From principles to pillars

E-23 is built around three outcomes. Model risk management must be enterprise-wide, risk-based and applied across the full model lifecycle. Under these outcomes sit principles. Some are high-level. Others are quite prescriptive, such as the minimum information a model inventory must capture.

E-23 also makes clear that model risk is not only the model owner's responsibility. Users, approvers and implementers all play a part.

To make this actionable, Atif and Efrem grouped the principles into six pillars. Each one is a separate building block of an MRM framework, and the structure reflects what we see in the MRM policies of the global banks Yields works with. Underneath all six sits your model risk framework, ideally linked to your risk appetite statement.

Pillar 1: Model inventory

Every organisation knows it needs a model inventory. Building one that holds up under E-23 is harder than it sounds.

Models are increasingly developed and deployed in analytical environments such as Databricks, Snowflake and Azure. Your inventory should connect to those platforms, so records are created automatically wherever possible. It also needs to discover tools across the organisation and test them against your definition of a model. In practice, that is often a short decision tree followed by a final human decision.

Three more points stood out:

• A flexible data model. As the scope of models grows, so do the attributes you need to capture. An inventory is never "set and forget".
• Event awareness. If a model is switched off in Databricks, the inventory should know. Someone then needs to confirm the right governance was followed.
• Dependencies. Efrem stressed that capturing dependencies is critical for AI agents. Every tool, LLM or skill an agent uses carries its own risk. Recent agentic AI guidance from the Monetary Authority of Singapore points towards narrowly scoped agents chained together, which makes the dependency chain even more important.

Pillar 2: Lifecycle and workflows

A model moves from ideation and development through validation, approval, implementation, use and periodic review. Model risk exists at every stage.

The operational challenge is the number of stakeholders involved: model owners, validators, approvers, implementers and users. A modern solution should orchestrate work across these teams, with role-based access and evidence captured at each step. When an implementation team confirms a model is live, for example, the approver and model owner should be notified automatically. Validation findings should be raised, answered, tracked and remediated in the same place.

Validation capacity is a growing concern. AI use cases and agents are increasing the volume of work faster than MRM teams can grow. Many leading organisations are exploring an AI-assisted first pass, where an LLM reviews documentation, data and code against internal policy and E-23, and drafts a starting point for the analyst. As Efrem noted, when agents start running MRM activities, deterministic processes and a full audit trail become even more important.

Pillar 3: A risk-based approach

Not all models are equal. The depth, scope and frequency of validation, and the level of approval, should be proportional to a model's importance, whether you call it a tier or inherent risk.

For technology, this means two things. Workflows must be driven by the tier. And the tier itself must be dynamic. If a portfolio grows significantly, the model's tier should be reviewed automatically, and a higher tier should trigger more scrutiny and a more senior approver.

For AI, this matters even more. An agent's risk can change overnight when it gains access to new tools or APIs. Without a way to track those changes, you underestimate your risk. As Atif put it, the spirit of traditional MRM still applies. The speed and frequency at which you apply it has to change.

Pillar 4: Monitoring and integration

This pillar covers what you actually test, from methodology and documentation to code, data and implementation. For AI and ML models, it also includes questions of bias, ethics and transparency.

Monitoring deserves special attention. Whether it runs every six months or several times a day, its purpose is to confirm that the model keeps doing what it was designed to do. The key capability is feeding monitoring results back into your MRM system, flagging breaches and alerting the right people so they can act.

Demo: automated model monitoring with alerts

‍

‍

Pillar 5: AI, ML and third-party models

E-23 explicitly brings AI and ML models into scope. They are grouped here with third-party models for a reason. Both can behave like black boxes: AI models because of their complexity, vendor models because the methodology is proprietary. In both cases, the question is the same. How do you get comfortable that the model does what it is supposed to do?

The volume and velocity of AI use cases add a further challenge. Organisations need an automated way to identify these tools, test them against the model definition and record a clear in-or-out decision.

Governance also works on two layers. Runtime guardrails operate in the first line. The second line then periodically reviews the telemetry. If an agent was blocked in 50% of cases, that warrants a closer look. That data needs to flow from first-line systems into the MRM platform.

Demo: governing AI agents with dynamic risk scoring

‍

‍

Efrem highlighted two takeaways from this demo. First, almost anyone can now build an agent, on platforms such as Google's agent registry or LangChain. Without syncing your inventory to those systems, you cannot govern what you cannot see. Second, automated re-scoring is essential. In the demo, an agent that started as low risk became high risk once one of its skills was expanded. That shift should immediately raise the level of governance scrutiny.

Pillar 6: Oversight and reporting

Eventually, you need to evidence all of this. That means extracting, aggregating and reporting your MRM data, ideally linked to your risk appetite statement, with access tailored to each level of seniority. Regulators increasingly send specific questions that need fast, accurate answers.

Atif also raised a point few organisations consider until they face it: multi-governance. A Canadian bank with US operations may have the same model subject to E-23 and US supervisory expectations. Internally, credit risk, operational risk and AI governance may each set their own requirements, and AI use cases often cross these boundaries. One approval may need a more senior sign-off in one jurisdiction than in another.

The answer is a single model record linked to every relevant framework, with each framework's requirements and evidence tracked separately. As Efrem added, this approach also drives consistency and removes duplicate data entry across teams and jurisdictions.

Start with the framework. Then choose the technology.

E-23 does not change the spirit of model risk management. It asks you to apply it across the enterprise, in proportion to risk, and at the pace of AI. The six pillars give you a structure to test your current setup and to ask the right questions of any solution, whether you build it in-house or buy it.

Download the full webinar and slides here

Want to see how Yields supports your E-23 implementation? Request a demo and we will walk you through the platform. Or simply reach out to exchange ideas on E-23 and AI governance. We are always happy to talk.

‍

‍