BankingNewsAI Daily Brief · Tuesday, March 3, 2026
Banking AI
Financial institutions & fintech technology
US Treasury just gave banks a practical AI risk playbook—treat it like a supervisory baseline now
The US Department of the Treasury released two new AI risk management resources aimed at financial institutions, explicitly covering how to govern and control AI as it moves from pilots into production. This is one of the clearer signals yet of what US policymakers expect banks and fintechs to be able to evidence: accountability, model risk controls, third‑party oversight, and operational resilience for AI systems.
Action
Map your current AI/GenAI inventory, vendor stack, and model governance to the Treasury resources and close gaps before examiners use them as a de facto benchmark. Require procurement and MRM to adopt explicit AI-specific third‑party and change-management controls (including agentic workflows), not just “traditional model” checklists.
CaixaBank put an AI agent inside its app to drive in-app purchases—banks are turning agents into revenue surfaces
CaixaBank launched an AI agent that helps customers make purchases within its mobile app, signaling a shift from “support chatbots” to embedded agents that influence conversion and product selection in regulated channels. This raises the bar on explainability, suitability/fairness controls, and recordkeeping because the agent is effectively part of the sales journey.
Action
Treat in-app agent guidance as a regulated sales interaction: log prompts/outputs, define approved product scripts/guardrails, and test for steering bias and mis-selling risk. Tie the agent to measurable KPIs (conversion, churn, complaints) with a kill-switch and tight change control.
General AI
Large language models & AI infrastructure
Reliability is now a differentiator: Claude had a global outage, then surged in adoption during the Pentagon news cycle
Anthropic reported a worldwide Claude outage lasting roughly 2 hours and 45 minutes, underscoring that even top-tier foundation models can have platform-level instability. In parallel, Claude’s consumer traction spiked amid the Pentagon dispute, highlighting how fast sentiment and availability can shift usage between leading models.
Action
Design for multi-model failover for any customer- or ops-critical workflows (routing, caching, graceful degradation), and make uptime/error budgets contractual where possible. Add real-time model/provider status monitoring into incident management the way you do for payments and core systems.
Nvidia is reportedly building a dedicated inference chip—compute economics may tilt again toward cheaper, faster serving
Reporting indicates Nvidia is working on a dedicated AI inference processor that could debut soon, suggesting a push to optimize for serving (latency, cost per token, efficiency) rather than just training. If true, this would matter more to enterprises than another training monster: inference dominates ongoing run-rate costs for scaled AI deployments.
Action
Reforecast AI unit economics assuming step-changes in inference price/performance and revisit build-vs-buy decisions for on-prem/sovereign deployments. Task engineering to benchmark workloads (fraud, contact center, coding assistants) for latency/cost sensitivity so you can exploit new inference hardware quickly.