BankingNewsAI Daily Brief ·
HSBC signs Google Cloud deal to operationalize AI agents across the bank.
Banking AI
Financial institutions & fintech technology
HSBC signs Google Cloud deal to operationalize “AI + agents” across the bank (not a pilot)
HSBC announced a multi-year partnership with Google Cloud to build and deploy AI capabilities globally, explicitly including agent-style tooling. This is a signal that large, regulated banks are now comfortable anchoring core AI delivery on hyperscaler stacks—with governance, security, and operating-model changes implied, not just model access.
Action
Mandate a cloud-and-model sourcing posture review: decide where you will standardize (e.g., Google/AWS/Azure) for agent deployment, and lock in guardrails (data residency, model logging, third-party risk) before business lines sign their own deals. Accelerate a production “agent runway” (identity, permissions, audit, human-in-the-loop) so you can ship comparable workflows without one-off exceptions.
Santander puts AI in the hands of all 185,000 staff—and is now publishing hard € business-value targets
Santander expanded AI access to its full workforce and tied it to quantified value creation, targeting over €200m in 2026 and >€1bn over 2026–2028. This moves the conversation from experimentation to accountable P&L impact, with implications for workforce enablement, controls, and measurement.
Action
Set explicit AI value KPIs by function (contact center, ops, compliance, software delivery) and require monthly scorecards tied to cost-out/revenue lift—not usage metrics. Fund enterprise enablement (training + prompt/agent standards + auditability) as a shared service so value capture isn’t trapped in a few early-adopter teams.
NAB deploys Databricks ‘Genie’ conversational analytics to cut 2–4 days per analytics use case
National Australia Bank is rolling out Databricks’ Genie/Genie Code so users can ask plain-English questions that translate into SQL and structured analytics work. The concrete claim—saving 2–4 days per use case—frames GenAI as cycle-time reduction in governed data environments, not just chat.
Action
Push your data org to quantify cycle-time wins (days-to-insight, backlog burn-down) from NLQ-to-SQL tooling inside your lakehouse, then scale only where query lineage, role-based access, and prompt logging are enforceable. Re-baseline analyst productivity expectations once the tool is stable—otherwise you pay for AI but keep legacy throughput assumptions.
General AI
Large language models & AI infrastructure
Midjourney jumped from image generation into a real medical scanning hardware prototype—signal that AI vendors may expand into regulated, physical-world products
Midjourney announced and published a technical dive on a new “Midjourney Scanner,” described in the discussion as radiation-free, magnet-free, fast, and low-cost, but requiring a water-immersion tank and having coarser resolution than CT/MRI. At least one attendee reported physically trying the demo scanner with their hand, implying this is a tangible prototype rather than a pure concept. For banks, the takeaway isn’t the modality—it’s that a major AI brand is moving into hardware + regulated domains, which changes how you should assess AI counterparties’ operational, liability, and regulatory maturity over time.
Action
Tighten third-party risk questions for AI vendors (roadmaps into regulated domains, safety/validation processes, auditability) and pressure your key AI providers on their governance posture at the next QBR—this is the direction credible AI companies are starting to move.
IBM watsonx.ai v2.4 doubles down on “governed AI” as a product category (not consulting)
IBM released watsonx.ai v2.4 positioned around governed development and enterprise controls. The direction is that governance features (policy, lineage, approvals, monitoring) are increasingly packaged as first-class platform capabilities rather than bespoke risk projects.
Action
Benchmark your AI platform against governance-as-product requirements (policy enforcement, audit logs, model registry, approval workflows) and decide whether to buy, build, or hybridize—then standardize. Stop letting each GenAI use case invent its own control stack; centralize controls so new models/agents can be swapped in without re-auditing everything.
Multi-model orchestration is emerging as a performance strategy: Sakana’s ‘Fugu’ system ships as a single API
Sakana AI released its Fugu system and “Fugu Ultra,” arguing it can reach frontier-level results by orchestrating multiple models/agents rather than relying on one monolithic model. For enterprises, this is a practical architecture shift: routing, ensembles, and tool-using agents become the lever for cost/performance and control.
Action
Have your AI platform team prototype an orchestration layer (policy-based routing across models, fallback behavior, cost caps, eval gates) instead of assuming a single ‘best model’ vendor will cover all workloads. Build procurement and risk processes that approve an orchestration pattern (with auditability) so you can swap models without redoing every use-case approval.