BankingNewsAI Daily Brief ·
The UK FCA shifts toward regulating general-purpose AI models in financial advice.
Banking AI
Financial institutions & fintech technology
UK FCA signals a shift toward regulating general-purpose AI models used in financial advice
The FCA’s newly published “Mills Review” and related commentary urge the UK to consider whether general-purpose AI (e.g., ChatGPT/Claude/Gemini) should fall within financial regulation as consumers increasingly rely on these tools for money decisions. The thrust is that existing firm-level rules may be insufficient when risk concentrates in a handful of upstream model providers and agentic interfaces start acting like de facto advisers.
Action
Treat upstream model governance as a regulated third‑party risk: tighten vendor due diligence, require evidence of model controls (testing, monitoring, incident response), and document how you prevent customers from construing LLM outputs as advice. Accelerate work on disclosure/guardrails in any customer-facing copilots before the FCA formalizes expectations.
Visa’s ‘agentic commerce’ moves from demo to live bank transactions in Europe (CaixaBank, BBVA)
CaixaBank and BBVA report completing transactions initiated by an AI agent on behalf of a cardholder under Visa’s Agentic Ready program. This is a concrete step toward AI agents being allowed to select merchants, initiate checkout, and trigger payments within card rails—shifting fraud, disputes, consent, and authentication assumptions.
Action
Define policy now for “agent-initiated” authorization (customer consent capture, step-up auth, transaction limits, dispute handling) and align with your card network and fraud teams on liability and controls. Build monitoring specifically for agentic behavioral patterns (high-velocity micro-purchases, new-merchant bursts, automated refunds/chargebacks).
Taktile raises $110M to automate credit and risk decisions—pushes ‘agent-first’ underwriting into banks
Decisioning platform Taktile raised $110M to expand automation of bank decision workflows (credit, fraud, onboarding, limits) using AI-driven policy and agentic orchestration. The financing signals continued institutional appetite for AI that directly changes approval/decline outcomes—not just productivity tooling.
Action
Pressure-test your model risk and auditability posture for agentic decisioning: require reason codes, policy traceability, and reproducible outcomes across model versions. If you’re buying decision automation, negotiate for regulator-ready evidence packs (data lineage, monitoring, bias testing, override governance) as a standard deliverable.
General AI
Large language models & AI infrastructure
Anthropic raised and simplified Claude API rate limits; Fable is slated to return to subscriptions once capacity allows
Anthropic increased Claude API rate limits and simplified tiering, and said the Fable model should return to subscription plans when capacity permits. This is an immediate availability/throughput change for any enterprise workloads using Claude, and it signals current constraints are operational capacity rather than a product pullback.
Action
Press your key vendors (contact center, document automation, developer tooling) to confirm their Claude usage, new rate-limit headroom, and fallback behavior so your SLAs and model-routing controls don’t degrade silently under load.
Microsoft creates a $2.5B ‘Frontier Company’ to embed 6,000 engineers inside customers for AI deployment
Microsoft stood up a new operating unit with $2.5B of internal funding and ~6,000 engineers focused on getting AI into production inside enterprises (the hard part is integration, governance, and change management—not model access). This formalizes “forward-deployed engineering” as a primary go-to-market motion for enterprise AI.
Action
Assume your peers will buy time-to-value by importing armies of vendor engineers; respond by funding your own AI platform/product operating model (standard patterns, reusable controls, and a deployment factory). If you engage Microsoft, ring-fence security and data-access boundaries up front—this delivery model only works if governance is pre-negotiated.
Tencent releases Hy3 under Apache 2.0, making a large ‘enterprise-grade’ model commercially usable without restrictive licensing
Tencent shipped Hy3 (Hunyuan 3) and shifted to an Apache 2.0 license, reducing legal friction for commercial adoption and redistribution. The release underscores that serious, globally usable non-US model options are improving, with reliability/enterprise positioning emphasized alongside capability.
Action
Expand your model portfolio strategy beyond the usual US providers: benchmark Hy3 (and peers) for multilingual performance, cost, and controllability, and validate procurement/legal paths for open-license deployment. Use the competitive pressure to renegotiate pricing and terms with incumbent model vendors.