BankingNewsAI Daily Brief ·
India’s RBI mandates enterprise-wide model risk management and AI kill switches.
Banking AI
Financial institutions & fintech technology
India’s RBI moves from “AI principles” to enforceable model governance: enterprise-wide MRM plus a kill switch
The Reserve Bank of India has issued draft enterprise-wide Model Risk Management (MRM) principles that explicitly cover AI use in banks, including requirements for human oversight and an operational “kill switch” to disable models quickly. This is a step-change from policy-level ‘responsible AI’ talk to control requirements that will drive how banks design, validate, monitor, and retire models across the institution.
Action
Stand up (or harden) an enterprise MRM program that treats genAI/agentic workflows as models: inventory, tiering, validation, monitoring, and documented fallback processes—then design technical controls for rapid model shutdown and reversion to manual/legacy paths. Use the RBI draft as a template to benchmark your own governance before other regulators copy the pattern.
Revolut trained a fraud “foundation model” on 40B events and claims a step-change in recall
Revolut says it trained PRAGMA, a banking-focused foundation model, on 40 billion events from 25 million users using Nvidia infrastructure, reporting a 64.7% improvement in fraud recall. This is notable because it’s not generic genAI—it's a large, institution-scale model built on proprietary behavioral/transactional telemetry for real-time risk decisions.
Action
Prioritize building a unified event/identity graph (card, ACH, login, device, merchant, disputes) that can support foundation-model style training/inference, or you’ll be structurally unable to match next-gen fraud performance. Pressure-test your fraud stack roadmap against competitors moving from rules/GBMs to foundation-model approaches with measurable lift.
Jack Henry + Google Cloud push “agentic defense” into the community bank core vendor layer
Jack Henry and Google Cloud expanded their partnership to deliver AI-driven security capabilities for banks and credit unions, leveraging Google Cloud’s agentic defense solutions to build a proprietary AI security platform. This signals AI-based cyber detection/response is moving from optional tooling to being embedded by core providers for smaller institutions that can’t staff modern SOC functions.
Action
Demand clear operating-model answers from your core and cloud providers: which controls are automated, what data is shared for detection, how actions are authorized, and how incidents are audited. If you’re mid-core renewal, treat AI security capabilities (and liability/SLAs) as a first-order negotiation item, not an add-on.
General AI
Large language models & AI infrastructure
Claude Tag turns Slack into an always-on “team member,” raising the stakes on enterprise context capture and control
Anthropic launched Claude Tag, a persistent Slack-native AI teammate designed to monitor channels, learn context, and take ongoing action rather than respond only in 1:1 chats. The shift is from “chatbot in a window” to “ambient agent in the workflow,” which materially increases both productivity upside and governance/data-retention risk.
Action
Treat Slack-embedded agents as a new privileged system: define which channels/data classes they can access, how outputs are logged for audit/eDiscovery, and what actions require approvals. If you’re rolling out internal agents, copy this pattern deliberately—persistent presence plus permissions, not ad-hoc chat usage.
Liquid AI’s 230M-parameter model claims strong extraction performance, pointing to cheaper on-prem doc intelligence
Liquid AI released LFM2.5-230M, a small model it says outperforms models up to 4x its size on data extraction and can run “anywhere.” If real, this supports a growing trend toward compact, task-optimized models that can be deployed close to sensitive data without heavy GPU spend.
Action
Re-evaluate your document automation stack (statements, trade docs, onboarding, disputes) for places where smaller models can replace larger LLM calls to cut cost and reduce data movement. Push vendors to prove extraction accuracy with your real documents and to support on-prem/VPC deployment, not just hosted APIs.