Radar · 09.09.2026 · banking · 7 sources

OSFI Draws the Line: Canada's Agentic AI Rules Are Now Concrete

OSFI bulletin plus Guideline E-23 give Canada the first explicit AI agent regulatory framework. Compliance clock is ticking.

OSFI’s July 2026 technology risk bulletin is no longer a draft signal. It is the first explicit Canadian regulatory guidance for agentic AI in financial services, and it lands hard. Unique non-human identities for AI agents. Least-privilege access. Human oversight for high-impact decisions. AI outputs treated as inputs, not outcomes. Third-party AI dependency mapping. Manual fallbacks. Guideline E-23, effective May 2027, extends model risk governance to self-learning and autonomous models. The FIFAI II AGILE framework explicitly names agentic AI as a tracked risk.

The compliance timeline is compressed. Banks need to map their AI governance against this framework now.

The core provider agentic platform wave has crystallized in a single quarter. Fiserv’s agentOS went GA in August with OpenAI and AWS. FIS launched Financial Crimes AI with Anthropic, BMO as first partner. Jack Henry rolled agentic security with Google Cloud across 7,400 institutions. Mambu shipped its Intelligent Core on September 2. MCP has become the de facto agent-to-system connector standard. As CCG Catalyst put it: your provider’s AI roadmap is quietly becoming your AI strategy, on their timeline, unless you decide otherwise. The real question is what banks build above the vendor agent layer.

Forbes crystallized the deposit war this week: the decision that costs banks the most happens when a customer’s AI evaluates financial products and no one from the bank is in the room. Santander and Mastercard proved agent payments in March. MX and Primitive’s Growth Agent automates deposit switching. Curinos reports AI deposit targeting delivers 2x response rates and 5.5x higher balances without aggressive pricing. The constraint is consumer trust: only 38% are comfortable with AI completing purchases.

Bain published the canonical AI-native bank framework: agentic customer experience, autonomous workflows targeting 80-90%, trust by design, AI-native data stack, AI-built operations. The key architectural insight is the deterministic-probabilistic boundary. That is THE architecture decision. Banks spend too much time choosing models when the focus should be clean data.

Canada’s Real-Time Rail entered industry assurance testing, on track for Q4 2026. Fifteen-plus new members including Adyen Canada. Real-time payment data is the highest-frequency behavioral signal layer available, and it converges with open banking Phase 1 and digital identity into a single Q4 inflection point.

Sources

  1. 01osfi-bsif.gc.cahttps://www.osfi-bsif.gc.ca/en/risks/technology-cyber-risk-management/technology-risk-bulletin/generative-agentic-artificial-intelligence-implications-technology-cyber-security-operational
  2. 02ccgcatalyst.comhttps://ccgcatalyst.com/thought-leadership/research-snapshot/sector-spotlight-ai-agents-and-connectors-for-banks-and-credit-unions
  3. 03forbes.comhttps://forbes.com/councils/forbesfinancecouncil/2026/08/13/banks-arent-competing-for-customers-anymore-theyre-competing-for-ai-agents
  4. 04bain.comhttps://www.bain.com/insights/what-it-takes-to-build-the-ai-native-modern-bank/
  5. 05payments.cahttps://www.payments.ca/canadas-real-time-rail-quarterly-update-jude-pinto-2026-q3
  6. 06nocash.rohttps://nocash.ro/the-battle-for-the-bank-customer
  7. 07openbankingexpo.comhttps://www.openbankingexpo.com/news/dgsi-ofnc-publish-draft-technical-specifications-on-open-finance-data-portability-in-canada/
Read next