OSFI's AI Governance Framework for Canadian Financial Institutions: What E-23, the July 2026 Bulletin, and the Fall Risk Outlook Mean for Your FRFI

Canadian financial institutions — banks, trust companies, insurers, mortgage companies, and federally regulated credit unions — are entering the most concentrated period of AI-related regulatory change in the sector's history. Three OSFI publications issued between September 2025 and October 2026 together define what governance a federally regulated financial institution (FRFI) must have in place for AI: the revised Guideline E-23 on model risk management, the July 2026 bulletin on generative and agentic AI, and the October 2026 Semi-Annual Risk Outlook update. Getting ahead of these requirements is not a compliance exercise — it is the foundation for capturing AI's documented productivity gains without incurring the supervisory, reputational, and operational risks that have already materialized at institutions that moved without adequate controls.

The Sector That Has the Most to Gain — and the Most to Govern

The financial services sector is, by exposure, the most AI-intensive industry in Canada. The Dais Institute's *Banking on AI* report (March 2026), commissioned by the Future Skills Centre, found that 98% of financial sector workers hold jobs highly exposed to generative AI — compared with 56% of the overall Canadian workforce. Of those workers, 73% are in occupations where AI is likely to replace tasks outright (business, finance and administration, and sales and service roles), while 24% — primarily senior management — are in roles where AI primarily augments rather than replaces.

On the adoption side, KPMG Canada's 2025 generative AI survey found that over 90% of Canadian financial services leaders see generative AI as critical to competitive advantage, with 86% actively investing despite economic uncertainty. Insurers are generally further ahead in core operations than banks, but both sectors have moved well beyond pilots. Across all industries, Statistics Canada's Q2 2026 analysis puts national business AI adoption at 19.2% — but financial services runs well ahead of that average.

The opportunity is real. So is the regulatory obligation to govern it properly.

Layer One: Guideline E-23 — Model Risk Management Binding Rule

The most consequential of the three regulatory publications is OSFI's revised Guideline E-23, published September 11, 2025, and effective May 1, 2027. This is a binding guideline, not a bulletin or advisory. Every FRFI — including banks, federally regulated insurers, trust companies, and foreign branches — must comply.

The revised E-23 replaces the 2017 version, which applied only to deposit-taking institutions. Its expanded scope and explicit treatment of AI and machine learning are the material changes.

What E-23 covers: The guideline creates a single enterprise governance regime for all models that carry non-negligible risk. It explicitly covers machine learning models, generative AI, and agentic AI — not as special categories, but as technologies subject to the same model risk management principles as any quantitative or statistical model. A model is defined broadly: any tool that applies statistical, quantitative, or AI/ML techniques to produce output used in decision-making. If your AI tool informs an underwriting decision, a credit assessment, a fraud flag, or an operational process, it is within scope.

The five pillars E-23 establishes:

That last pillar is critical for AI governance in 2026. Many financial institutions are using foundation models and cloud-based AI services — Claude Enterprise, Microsoft Copilot, AWS-hosted AI services — that they did not build internally. Under E-23, using a third-party model does not transfer the FRFI's model risk obligations to the vendor. As Torys LLP noted in its commentary on E-23, many AI and machine learning vendors do not yet have governance, validation, and reporting capabilities consistent with MRM standards — which means FRFIs using them must assess and manage the residual risk themselves.

Timeline pressure: With an effective date of May 1, 2027, institutions have approximately eighteen months from the September 2025 publication to build or validate their MRM frameworks. For FRFIs that have deployed AI tools organically — teams using Copilot, fraud models built on ML, underwriting assistants — the first task is completing a model inventory. Many institutions will find their AI footprint larger and less documented than expected.

Layer Two: The July 2026 AI Bulletin — Sound Practices for Generative and Agentic AI

OSFI's *Generative and Agentic Artificial Intelligence: Implications for Technology, Cyber Security, and Operational Resilience* bulletin, issued July 2026, is non-binding guidance — it does not create new legal duties. But it is explicitly grounded in three existing binding guidelines: B-13 (Technology and Cyber Risk Management), E-21 (Operational Risk and Resilience), and B-10 (Third-Party Risk Management). Supervisors will look to the bulletin when assessing whether an institution's existing controls adequately address the specific risks of generative and agentic AI.

McCarthy Tétrault's analysis of the bulletin identifies four key risks OSFI has targeted:

1. Hallucination and output reliability. AI-generated outputs must be treated as inputs to decision-making, not as decisions. Staff who use AI-generated analysis, summaries, or recommendations in their work need clear protocols for verifying those outputs before acting on them. This is not optional good practice — it is what B-13's accuracy and reliability expectations require.

2. Autonomous code deployment. AI-generated code — from tools like GitHub Copilot or Claude Code — must be validated by a human reviewer before deployment. The risk is both functional (incorrect code) and security-related (AI-generated code can introduce vulnerabilities that are harder to spot in code review).

3. Data leakage through third-party AI. When employees use AI tools that process customer data, sensitive financial information, or proprietary models, that data flows to the vendor's infrastructure. Under B-10 and PIPEDA, the FRFI retains accountability for that data. The bulletin recommends that institutions inventory which AI tools are in use and map the data flows through each — a practice entirely consistent with what a sound B-10 third-party risk assessment requires.

4. Agentic AI and access rights. AI agents — systems that can take actions autonomously rather than just producing text or analysis — require specific access governance. OSFI recommends that agent identities be scoped with least-privilege access, meaning an AI agent should be able to do only the specific actions its function requires, not everything an employee account would permit. This is a concrete control gap at most institutions today.

Layer Three: The October 2026 Risk Outlook — Frontier AI as a Systemic Risk

OSFI's Semi-Annual Risk Outlook update, released October 8, 2026, elevated frontier AI to the headline risk in OSFI's current supervisory perspective. The update's core finding: frontier AI models have materially shortened the time between vulnerability discovery and exploitation, enabled more sophisticated and coordinated cyberattacks, and increased third-party concentration risk through the small number of major AI and cloud providers that the financial system depends on.

OSFI Superintendent Peter Routledge framed it directly: resilience is now a competitive advantage in an era of advancing AI capabilities. Institutions that have invested in governance, resilience testing, and third-party oversight are better positioned to operate through the cyber and operational incidents that OSFI expects to increase in frequency and sophistication.

For practical purposes, the Fall 2026 update reinforces two things for FRFIs. First, the threat environment for AI-enabled attacks is worse than it was when B-13 was developed in 2023 — institutions need to assess whether their technology and cyber risk controls remain adequate against AI-augmented threat actors. Second, the concentration of AI infrastructure at a small number of hyperscale providers (AWS, Microsoft Azure, Google Cloud) creates systemic third-party risk that OSFI is actively monitoring and expects institutions to manage under B-10.

Four Actions for FRFIs in the Next 90 Days

With E-23 effective in May 2027 and the AI bulletin setting supervisory expectations now, four near-term actions directly reduce both regulatory exposure and operational risk.

1. Complete a model inventory. Identify every AI and ML tool that produces outputs used in any business decision — including vendor tools, embedded AI in enterprise software, and employee-facing tools like Copilot or Claude. For each model, document: what it does, who uses its outputs, what data flows through it, and whether it has been validated. This is the prerequisite for E-23 compliance and for the B-10 vendor assessment the bulletin requires.

2. Assess your third-party AI vendors against E-23 and B-10. For each vendor-provided model with non-negligible risk, determine whether the vendor can support the validation, documentation, and monitoring requirements E-23 places on the FRFI. Where a vendor cannot, document the residual risk and the compensating controls in place. This is what OSFI will review under B-10 — not whether the vendor is certified, but whether the institution has done its due diligence.

3. Establish an AI output review protocol. The July 2026 bulletin's requirement to treat AI outputs as inputs, not decisions, needs to be operationalized at the workflow level. For any process where AI-generated content influences a customer-facing outcome, a credit decision, or a compliance filing, define the review step, who performs it, and what documentation is required.

4. Scope agent access rights. If your institution is deploying or planning to deploy agentic AI — AI that takes actions rather than just producing outputs — review the access rights assigned to those agents now. An agent that can initiate payments, change account data, or execute trades needs scoped, least-privilege credentials managed with the same rigour as privileged human user accounts.

Where This Fits in a Broader AI Governance Program

The E-23/B-13/E-21/B-10 framework is specifically shaped for FRFIs, but its underlying logic — inventory, validate, monitor, govern third parties — is the same structure that non-regulated organizations need for sound AI governance. For Canadian mid-market companies outside the regulated financial sector, the OSFI framework is a useful model even where it is not a binding obligation.

For FRFIs, the practical reality is that the regulatory program and the business opportunity are the same work. An institution that completes an accurate model inventory, validates its AI models, and governs its AI vendor relationships has also done the work required to trust its AI outputs, capture productivity gains, and avoid the reputational and operational incidents that poorly governed AI systems generate.

Sources

• OSFI. *Guideline E-23: Model Risk Management.* osfi-bsif.gc.ca (September 11, 2025, effective May 1, 2027)
• OSFI. *Generative and Agentic Artificial Intelligence: Implications for Technology, Cyber Security, and Operational Resilience.* osfi-bsif.gc.ca (July 2026)
• OSFI. *Annual Risk Outlook — Semi-Annual Update, Fiscal Year 2026-2027.* osfi-bsif.gc.ca (October 8, 2026)
• The Dais. *Banking on AI: Generative AI Adoption in Canada's Financial Sector.* dais.ca (March 2026)
• KPMG Canada. *Generative AI Adoption in Canadian Financial Services.* kpmg.com (February 2026)
• Statistics Canada. *Analysis on Artificial Intelligence Use by Businesses in Canada, Second Quarter of 2026.* statcan.gc.ca
• McCarthy Tétrault. *Managing the AI Risks in the Financial Sector: OSFI's Bulletin on Generative and Agentic Artificial Intelligence.* mccarthy.ca
• Torys LLP. *OSFI Updates and Expands Scope of Guideline E-23.* torys.com (October 2025)
• Wealth Professional. *OSFI Warns Frontier AI Can Amplify Cyber and Third-Party Risks.* wealthprofessional.ca

The OSFI framework — E-23, the July 2026 AI bulletin, and the B-10/B-13/E-21 network of guidelines — creates a governance obligation that federally regulated financial institutions need to act on now, not in April 2027. Cloud Forces' AI Advisory service helps Canadian financial institutions and mid-market organizations conduct an AI Readiness Assessment that maps your model inventory, identifies your regulatory gaps against E-23 and the July 2026 sound practices, and builds the governance documentation that demonstrates accountability to OSFI examiners. As an AWS Consulting Partner since 2019, we work with both the infrastructure layer and the policy layer. Book a consultation to start with a clear picture of where your AI governance stands.