BankingNewsAI Daily Brief ·
FSB launched a consultation to set global responsible AI practices for finance.
Banking AI
Financial institutions & fintech technology
FSB opens consultation on “sound practices” for responsible AI in finance—global baseline is forming
The Financial Stability Board has launched a consultation on sound practices for the responsible adoption of AI in financial services. Unlike national guidance, this is aimed at cross-border consistency and supervisory expectations for systemically relevant firms, signaling a coming convergence on governance, model risk management, and operational resilience for AI.
Action
Mobilize a response team (Risk, Compliance, Model Risk, Tech) to map your current AI controls to likely FSB themes and identify gaps you can close quickly (inventory, third-party model oversight, incident response, monitoring). Treat this as a chance to shape the standard—and to preempt future exam findings across jurisdictions.
Shield starts surveilling Microsoft 365 Copilot interactions—closing a fast-growing compliance blind spot
Shield expanded its financial-services communications surveillance coverage to include Microsoft 365 Copilot interactions and image-based content. As Copilot becomes embedded in employee workflows, the chat prompts, outputs, and generated artifacts increasingly look like “electronic communications” that firms are expected to supervise and retain.
Action
Treat Copilot logs as regulated communications: update retention, supervision, and eDiscovery requirements and ensure your tooling can capture prompts/outputs with appropriate identity and context. If you’re rolling out Copilot broadly, require surveillance coverage as a gate for expansion beyond controlled pilot groups.
General AI
Large language models & AI infrastructure
Microsoft ships “Copilot Cowork” to general availability—agentic collaboration is moving into core M365
Microsoft announced general availability of Copilot Cowork, pushing multi-agent/assistant collaboration deeper into Microsoft 365. This matters because M365 is the default productivity layer in most large enterprises; once agentic features land there, adoption can scale from “innovation pilots” to everyday operating model quickly.
Action
Standardize governance for “agents in productivity tools” now: define what data Copilot can access, what actions it can take, and how you audit outcomes. If you wait until after broad enablement, you’ll be chasing shadow usage across email, documents, and Teams rather than shaping it.
OpenAI’s new “Deployment Simulation” formalizes pre-launch behavior testing using real conversation data
OpenAI introduced Deployment Simulation, a method to predict how models will behave in the wild before release by simulating deployment with real conversational patterns. The key shift is from static evals to scenario-based, distribution-aware testing—closer to how models actually fail in production and how they interact with users over time.
Action
Adopt simulation-style evaluation internally for any high-risk AI workflow (customer comms, credit decisions support, fraud/AML triage): test with realistic user journeys, adversarial prompts, and operational edge cases. Use the results to set go/no-go thresholds and monitoring metrics that match real usage, not lab benchmarks.