BankingNewsAI Daily Brief ·
Citi launches Arc, a bank-wide AI agent platform with shared controls.
Banking AI
Financial institutions & fintech technology
Citi moves from “AI pilots” to a bank-wide agent platform (Arc) with shared controls
Citigroup launched “Arc,” an internal platform to build, deploy, and scale AI agents across the firm, explicitly targeting automation of research, synthesis, and execution work. The significance is less the agents themselves than the platform move: Citi is standardizing how agents are built and governed across business lines instead of letting one-off tools proliferate.
Action
Stand up (or accelerate) a centralized “agent factory” with reusable components (identity, entitlements, audit logs, tool approval, evaluation) before business units deploy incompatible agents at scale. Set a near-term target for 2–3 repeatable agent use cases (e.g., KYC refresh, RM prep, policy/Q&A) to force platform hardening and governance decisions quickly.
General AI
Large language models & AI infrastructure
DeepSeek reportedly targets a $45B valuation—signal that low-cost frontier labs can scale into credible enterprise suppliers
TechCrunch reports DeepSeek could reach a ~$45B valuation in its first investment round, after gaining attention for training competitive models with far less compute than US peers. The strategic shift is that banks may face a broader set of powerful model vendors (including China-linked) competing on price/performance—complicating third-party risk, sovereignty, and procurement decisions.
Action
Add a formal “model vendor country-of-control/hosting” assessment to procurement and MRM, including data residency, legal access risk, and supply-chain dependencies. Use increased vendor competition to push for better terms (indemnities, audit rights, logging, eval access) from all model providers—not just the challengers.
US national-security testing expands: major labs agree to share frontier models with CAISI pre-release
Major AI companies (reported: Google DeepMind, Microsoft, xAI) agreed to share frontier models with the US government’s CAISI for national security testing ahead of public release. This is an early shape of a ‘pre-deployment safety review’ regime that could become a de facto standard—and a template regulators may expect critical infrastructure firms to mirror internally.
Action
Adopt a CAISI-like internal gate for any agent that can take actions (payments, customer changes, code execution): red-teaming, abuse testing, and documented signoffs before production. Align your third-party due diligence questionnaires to ask whether vendors participate in pre-release government testing and whether you can access results or equivalent assurance.