BankingNewsAI Daily Brief ·
The FSB set 12 concrete AI governance practices as the new bank baseline.
Banking AI
Financial institutions & fintech technology
FSB just put 12 concrete AI governance practices on the table for banks (and will likely become the baseline for supervisors)
The Financial Stability Board published a consultation report proposing 12 AI governance and risk-management practices for financial institutions, explicitly covering model risk, third-party/vendor dependency, and controls around AI use in core financial activities. This is the closest thing yet to a cross-jurisdiction supervisory “common denominator” for bank AI, and it will be referenced by local regulators even before any rules are finalized.
Action
Map your current AI control framework to the FSB practices now and identify gaps you can close quickly (inventory, testing/validation, third-party assurance, monitoring, and clear accountability). Use the consultation window to push for practical requirements on vendor transparency and auditability—especially for frontier models and agentic workflows.
Bank of England is signaling agentic-AI-specific guardrails for markets and cyber risk—prepare for new supervisory asks
BoE Deputy Governor Sarah Breeden warned that autonomous/agentic AI could increase cyber risk and amplify market stress if many agents react the same way at once, and suggested the existing regulatory framework may need reform to handle these dynamics. This moves agentic AI from a tech risk topic to a financial stability and market structure issue—i.e., something supervisors can justify acting on quickly.
Action
Run a fast internal review of any agentic or semi-autonomous capabilities in trading, treasury, and security operations (including vendor tools) and document human-in-the-loop controls, kill switches, and concentration/correlation risks. Pre-empt likely BoE-style questions by stress-testing “herding” scenarios and tightening third-party assurance for agentic tools in SOC and markets functions.
General AI
Large language models & AI infrastructure
Google made “computer use” a built-in Gemini 3.5 Flash capability, with explicit human-confirmation safety controls
Google shipped a standardized “computer use” action interface in Gemini 3.5 Flash across browser/desktop/mobile, positioned as a first-class built-in capability rather than a DIY agent pattern. They highlighted controls like explicit user confirmation for sensitive actions and automated task stopping, making it easier for enterprises to deploy UI-driving agents with guardrails.
Action
Pressure your major vendors (CRM, service desk, contact center, RPA) to show their roadmap for Gemini’s standardized computer-use interface and the exact human-in-the-loop/approval logging they support before you allow any UI-automation in regulated workflows.
Claude Sonnet 5 drops the cost of running agents—“good enough” agentic performance just got cheaper and default
Anthropic released Claude Sonnet 5 as the new default model, positioning it as much closer to top-tier performance while remaining materially cheaper than flagship models. The practical shift is that more teams can justify production agent workflows (tool use, planning, multi-step execution) without needing premium models for every task.
Action
Reprice your internal agent business cases: you can now split workloads across tiers (Sonnet-class for most steps, premium only for hard cases) and push more automation into operations with the same budget. Update your vendor/model selection playbook to explicitly include “agentic capability per dollar” and regression-test for safety/controls when swapping default models.
AWS is spending $1B to embed forward-deployed AI engineers—agent deployments are becoming a services-led arms race
AWS launched a $1B Forward Deployed Engineering (FDE) push to embed engineers with customers to ship agentic AI solutions quickly, mirroring patterns seen at major AI labs. The market signal: the bottleneck has shifted from model access to execution—secure integration, workflow redesign, and production hardening.
Action
Treat “deployment talent” as a strategic constraint: build an internal bench (or pre-negotiate FDE-style support) for the top 2–3 bank workflows you want automated this year. Tighten your reference architecture for agents (identity, authorization, audit, data boundaries) so external teams can move fast without re-litigating controls each time.