What happened
OSFI's July 2026 Technology Risk Bulletin says generative and agentic AI amplify the speed, scale and automation of cyber and operational risk. It lists sound practices tied to Guidelines B-13, E-21 and B-10: define ownership and limits on autonomy, keep human oversight with audit trails for material decisions, restrict AI inputs to approved data, ban sensitive data in public or unapproved tools, treat AI output as input to a decision rather than the decision, and control AI-generated code before it reaches production. (Office of the Superintendent of Financial Institutions (OSFI), July 2026)
Note on figures: This bulletin is supervisory guidance, not a new binding guideline.
What it means for Canada
OSFI expects banks and insurers to set limits on AI autonomy, keep humans accountable for material decisions and control AI-written code. Their AI vendors will be asked to show the same.
- Sectors
- Banks · Insurers · Trust and loan companies · Fintech and AI vendors to them
- Tools involved
- Generative AI assistants · AI agents · AI coding tools
- Canadian rules
- OSFI B-13 (technology and cyber risk)
OSFI E-21 (operational resilience)
OSFI B-10 (third-party risk)
OSFI E-23 (model risk, in force May 2027)
Takeaways for Canadian businesses
- BusinessWrite down what each AI agent may do on its own and where it needs human approval; OSFI calls this defining limits on autonomy.
- BusinessLog prompts, outputs and agent actions so material decisions can be audited.
- BusinessBlock sensitive data from public or unapproved AI tools.
- BusinessPut AI-generated code through the same review and change control as human code before production.
For the Canadian AI sector
AI-agent vendors selling to Canadian financial institutions should expect questions on least-privilege access, logging, kill switches, data provenance and who approves high-impact actions.
Written by AgentsNorth from the source below. General information, not legal or security advice.