BankingNewsAI Daily Brief ·
FSB opens consultation to tighten global financial-stability controls for agentic AI systems.
Banking AI
Financial institutions & fintech technology
FSB just put agentic AI controls on the global financial-stability agenda (consultation open)
The Financial Stability Board released a consultation on “sound practices” for responsible AI adoption across financial institutions, explicitly calling out agentic AI as a risk amplifier (autonomy, tool-use, speed/scale, and potential for correlated failures). This is the clearest signal yet that supervisors will converge on expectations for AI governance, testing, third‑party controls, and incident readiness—not just model risk in isolation.
Action
Stand up a board-visible “agentic AI control framework” now: pre-deployment kill switches, permissioning for tool/action execution, continuous monitoring, and stress tests for cross-model/cross-vendor correlated behavior; use the FSB practices as your baseline for upcoming exams and procurement standards.
Visa + OpenAI move agentic commerce from demos to payments rails (registry + scoring + ChatGPT integration)
Visa announced new “intelligent, programmable commerce” capabilities—agent scoring, an agentic registry, and a Large Transaction Model—alongside an agentic commerce partnership that connects Visa’s network into OpenAI/ChatGPT experiences. Separately, Santander’s Getnet launched AI-powered capabilities for merchants to accept agent-initiated payments, showing the acquiring side is preparing for machine-originated transactions too.
Action
Treat agent-initiated payments as a new channel: update fraud/risk policies (who is the ‘customer’—human vs agent), revise authentication and dispute flows, and require merchant/PSP partners to support agent identity, spend controls, and transaction-level explainability before volume arrives.
ING is using agentic AI to accelerate mortgage decisions—pressure is shifting from ‘AI pilots’ to cycle-time wins
ING deployed agentic AI in the Netherlands to speed up mortgage applications, indicating real workflow execution (not just copilots) in a regulated credit decisioning process. This raises the bar on operational controls: audit trails, model governance, and how automated actions are constrained and reviewed when the agent can progress a case end-to-end.
Action
Benchmark your mortgage/consumer lending cycle-time vs peers and identify one agentic “thin slice” (document ingestion → data validation → case routing) that can be deployed with full auditability; prioritize controls for action-permissioning and adverse-action/decision explainability before expanding scope.
General AI
Large language models & AI infrastructure
Google’s DiffusionGemma points to a step-change in inference economics (extreme throughput, open model)
Google released DiffusionGemma, an open model that generates text via diffusion rather than token-by-token autoregression, with reports of very high throughput (order-of-magnitude faster in some settings). If the approach generalizes, it could materially reduce latency/cost for on-device or private deployments—especially for high-volume agent workloads.
Action
Revisit your unit economics for AI at scale (contact center, fraud ops, developer tooling): pilot diffusion-style models in controlled workloads and renegotiate capacity plans with the assumption that the cost curve for inference may bend downward quickly for certain classes of tasks.