Canada's federal financial regulator has published new guidance on generative and agentic artificial intelligence, warning that the technology's autonomous reasoning and execution capabilities can outpace existing risk controls. The Office of the Superintendent of Financial Institutions (OSFI) released a bulletin in July 2026 outlining sound practices for insurers, banks, and trust companies to manage the operational risks these systems introduce.
Generative AI creates content, while agentic AI can reason and act on its own. Both can lift productivity, but they also speed up and scale up cyber and operational risks in ways that test existing risk management frameworks. The bulletin builds on OSFI's earlier work on frontier models and lines up with three existing guidelines: B-13 on technology and cyber risk, E-21 on operational risk and resilience, and B-10 on third-party risk.
Governance gaps grow as AI systems gain autonomy
AI adoption can move faster than the oversight frameworks meant to control it. Systems can operate with little human supervision while leaning heavily on third-party models, data, and application programming interfaces. When senior managers do not fully grasp the technology, OSFI warns, they can lean too hard on vendor assessments. The regulator suggests strengthening management literacy, tying AI strategy to risk appetite, folding AI risk into enterprise risk management, and setting clear limits on how much autonomy a system is given. OSFI's call for stronger management literacy and clear autonomy limits mirrors the discussions in AI for Insurance, where leaders are working to balance innovation with operational safety.
When AI hallucinates, downstream damage follows
Generative AI can produce hallucinated outputs-inaccurate or misleading results-and agentic AI can act on them, triggering downstream actions and data leakage. OSFI advises treating AI outputs as inputs to a decision rather than the decision itself, and keeping a human accountable for material calls. Accuracy problems are not just technical; they directly affect underwriting decisions, claims assessments, and customer communications.
Tightening software and access controls
Generative AI can write code quickly but may reproduce insecure patterns, and an agent can push flawed code into production without human review. OSFI suggests validating AI-generated code before deployment and enforcing approval checkpoints for high-risk actions. On access, the regulator flags the danger of over-privileged AI agents and shared credentials. It recommends unique non-human identities, least-privilege and scoped permissions, and short-lived, just-in-time credentials to prevent multi-step risks such as data exfiltration through tools or APIs.
Security and third-party dependencies
AI cuts both ways in security, OSFI says. It fuels automated phishing and malicious code, but institutions can also use it to detect and contain threats faster. The bulletin urges insurers to map their AI dependencies, test outage scenarios, keep manual fallbacks, and require third parties to disclose how they use AI. A failure in one AI service can ripple across institutions, so resilience planning must extend beyond the organization.
Why this matters for insurance professionals
For insurers, the bulletin is a practical checklist. Boards and senior managers need to understand AI well enough to set risk appetite and autonomy limits. Underwriting and claims teams should treat AI outputs as inputs, not decisions, and keep a human accountable for material calls. IT and security teams must enforce least-privilege access for AI agents, validate AI-generated code before deployment, and test outage scenarios with manual fallbacks. Third-party due diligence must now include disclosure of supplier AI use, because a failure in one AI service can cascade across the institution.
Your membership also unlocks:
