BankingNewsAI Daily Brief ·
OpenAI ships Broadcom-built Jalapeño inference chip, tightening control over AI compute costs.
Banking AI
Financial institutions & fintech technology
RBI’s draft AI governance rules for banks harden into “kill switch + human accountability” expectations
India’s RBI has issued draft guidance for banks that explicitly requires human oversight for AI decisions and operational controls such as kill switches, alongside a broader model risk management framework. This is a concrete supervisory signal that “agentic” and generative use cases will be treated like high-consequence models with mandated intervention and auditability.
Action
Stand up (or refresh) an AI model inventory that includes GenAI/agent workflows, map each to an accountable human owner, and implement tested rollback/kill-switch procedures before the regulator forces a rushed retrofit. Use the RBI direction as the template for group-wide controls in other jurisdictions likely to copy these mechanics.
Santander industrializes AI: tools rolled out to ~185,000 employees with quantified €1B target
Banco Santander has moved from pilots to broad deployment by making AI tools available to its entire workforce and publicly tying the program to a €1B benefit target by 2028. The notable change is operational scale (enterprise-wide access) paired with measurable outcomes and a large initiative pipeline (hundreds of projects).
Action
Set a bank-wide “AI adoption + value” operating cadence (usage telemetry, time saved, risk events, and P&L attribution) rather than treating GenAI as isolated CoE experiments. If you already have tooling, shift focus to workflow redesign and measurement—Santander is signaling that competitive advantage now comes from execution at scale, not model access.
General AI
Large language models & AI infrastructure
OpenAI ships its first custom inference chip with Broadcom (Jalapeño), signaling a step-change in cost/performance control
OpenAI and Broadcom unveiled Jalapeño, a custom processor optimized for LLM inference workloads. This is a strategic shift: the leading model vendor is now vertically integrating into silicon to reduce inference cost, improve throughput, and control supply—pressuring both cloud pricing and the economics of enterprise deployments.
Action
Re-forecast GenAI unit economics assuming a faster cost decline and more vendor-specific performance advantages; procurement leverage shifts if model providers can undercut hyperscaler pricing. Push architecture toward provider-agnostic interfaces where possible so you can arbitrage performance/cost across chips and clouds as this race accelerates.
Google adds “computer use” as a built-in tool in Gemini 3.5 Flash—agents can now operate UIs, not just APIs
Gemini 3.5 Flash now includes native computer-use capabilities, enabling agents to interact with on-screen workflows across applications. The practical change is that automation can target legacy and desktop/web processes without waiting for clean APIs, expanding the reachable surface area for agentic operations (and the associated control risk).
Action
Treat UI-operating agents as privileged automation: require sandboxing, least-privilege credentials, session recording, and deterministic rollback plans before letting them touch production ops. Identify 2–3 “API-poor” processes (ops reconciliations, casework, reporting) where UI agents can deliver step-function productivity—then wrap them in strong controls.
Mistral’s OCR 4 makes on-prem, structure-aware document intelligence credible for regulated enterprises
Mistral launched OCR 4, positioning document extraction as a full enterprise capability with stronger structure awareness and deployment flexibility. The key banking-relevant change is practical on-prem/inside-VPC viability for high-sensitivity documents—reducing the need to ship customer data to third-party SaaS for core document workflows.
Action
Accelerate replacement of brittle OCR + rules pipelines in onboarding, trade finance, and servicing with a model that can run in your controlled environment. Use this to reduce data-exfiltration objections that stall automation, while tightening QA thresholds and exception handling to avoid “silent” extraction errors in regulated processes.