---
title: What does OSFI Guideline E-23 mean for AI agents at Canadian banks?
url: https://opencontroller.lyzr.com/blog/osfi-e-23-model-risk-and-ai-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What does OSFI Guideline E-23 mean for AI agents at Canadian banks?

**OSFI's final Guideline E-23 on model risk management, published in September 2025 and effective 1 May 2027, explicitly brings AI and machine learning methods into its definition of a model. For federally regulated financial institutions, that means AI agents that use models to produce results need to be in the model inventory, risk-rated, governed across their lifecycle, monitored, and formally decommissioned.**

E-23 is not an AI law. It is a prudential guideline from Canada's Office of the Superintendent of Financial Institutions (OSFI). But its scope is wide enough that most production AI agents at a bank or insurer will fall inside it, and the 2027 date leaves roughly 18 months to build the evidence.

## Who does E-23 apply to and when?

E-23 applies to federally regulated financial institutions (FRFIs): banks, foreign bank branches, life insurance and fraternal companies, property and casualty insurers, and trust and loan companies. Application is risk-based and proportional to size, strategy, risk profile and complexity. The guideline takes effect on **1 May 2027**.

## Does an AI agent count as a "model" under E-23?

E-23 defines a model as an application of theoretical, empirical or judgmental assumptions or statistical techniques, "including AI/ML methods", which processes input data to generate results. It describes three components: data input, processing, and a result output. It adopts the OECD's definition of AI.

An AI agent built on a large language model fits that description. The LLM processes inputs (prompts, retrieved documents, tool results) and generates results (answers, decisions, actions). E-23 also calls out extra considerations for AI/ML such as explainability, bias and autonomous behaviour, which are exactly the properties that distinguish agents from traditional models.

The practical question is not "is it a model?" but "what is the unit of inventory?" Most banks will need to record both the underlying model (for example, a vendor LLM) and the agent that uses it, because the agent's prompts, tools and permissions change its risk.

## Which E-23 principles create work for agent teams?

E-23 sets out principles across governance, model identification, and the lifecycle. The table maps the ones that matter most for agents to the evidence you would need.

| E-23 principle | What it says (summary) | Agent evidence to produce |
|---|---|---|
| 2.1 Model identification | Identify and track all models in use or recently decommissioned | An enterprise inventory of agents, their models, versions, owners, and retired agents |
| 2.2 Risk rating | A rating approach covering key dimensions of model risk | Risk tier per agent based on autonomy, data access, tools and customer impact |
| 2.3 Intensity | MRM effort commensurate with risk | Review depth and monitoring frequency tied to each agent's tier |
| 3.4 Review | Independent assessment of conceptual soundness and performance | Pre-production evaluation results, red-team results, sign-off by an independent reviewer |
| 3.5 Deployment | Quality and change control in deployment | Version history, change approvals, record of what changed between versions |
| 3.6 Monitoring and decommission | Defined standards for monitoring and decommission | Production quality scores, drift alerts, thresholds, and a retirement record |
| Third-party models | Framework covers externally sourced models and data, aligned with Guideline B-10 | Inventory of vendor models and vendor-built agents with third-party risk links |

## What are the hard parts for AI agents?

Three things make agents harder than classic credit or market models.

1. **Discovery.** Agents are built on many platforms, often by business teams. An inventory built from questionnaires misses agents nobody declared. Automated discovery across cloud accounts and SaaS platforms closes the gap. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
2. **Change frequency.** A prompt edit or a new tool can change behaviour more than a model retrain. Decide which changes trigger re-review; [when an agent change needs re-approval](/blog/when-does-an-ai-agent-change-need-re-approval) covers this.
3. **Ongoing monitoring.** Back-testing does not fit open-ended language output. Banks are using LLM judges on sampled production traffic, groundedness checks and task-adherence scores as monitoring metrics, with thresholds that trigger review.

## A 2027 readiness checklist

1. Confirm your MRM policy's model definition explicitly includes LLM-based agents.
2. Decide the inventory unit: agent, agent version, and underlying model.
3. Populate the inventory from automated discovery, not only self-declaration.
4. Define agent risk-rating criteria and rate every production agent. Our guide to [risk-tiering AI agents](/blog/how-to-risk-tier-ai-agents) gives a starting rubric.
5. Set pre-deployment evaluation standards per tier, including adversarial testing for higher tiers.
6. Put version control and change approval on prompts, tools and model choices.
7. Define production monitoring metrics and thresholds per tier.
8. Write a decommission procedure and keep retired agents in the inventory.
9. Link vendor models and vendor-built agents to third-party risk records under B-10.

For broader context, read [AI agent controls for banks' model risk](/blog/ai-agent-controls-for-banks-model-risk) and [how to map agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller produces evidence for E-23 work rather than certifying anything. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connects to platforms such as Salesforce, SAP and Databricks, feeding a governed registry of agents, versions, tools and MCP servers. Multi-turn and red-team evals support pre-deployment review; LLM judges score production traffic and live groundedness and task-adherence guardrails supply monitoring signals. Policy as code ties controls to risk tiers. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**When does OSFI E-23 take effect?**
On 1 May 2027. OSFI published the final guideline in September 2025.

**Does E-23 cover vendor AI models?**
Yes. The framework must cover models and data from external sources, and OSFI ties this to its third-party risk Guideline B-10.

**Is a chatbot without tools in scope?**
If it uses an AI/ML method to process inputs and generate results used by the institution, it fits the definition. Proportionality decides how much effort it needs.

## Key facts

OSFI Guideline E-23 Model Risk Management (2027) was published September 2025 and takes effect 1 May 2027. It applies to federally regulated financial institutions (banks, foreign bank branches, life, P&C, trust and loan companies). It defines a model as an application of theoretical, empirical, judgmental assumptions or statistical techniques, including AI/ML methods, which processes input data to generate results, and adopts the OECD AI definition. Principles include model identification (2.1), risk rating (2.2), intensity (2.3), review (3.4), deployment (3.5), monitoring and decommission (3.6); third-party models align with Guideline B-10. Opencontroller provides evidence for E-23 via discovery, registry, evals and production LLM judges. Licensing starts from $250K per deployment per year.

## Related

- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [When does a change to an AI agent need re-approval?](https://opencontroller.lyzr.com/blog/when-does-an-ai-agent-change-need-re-approval.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
