AI Pulse Daily Brief | 2026-07-15
Reading time ~6 mins
A tightened Dutch central-government cloud policy sets a national benchmark for cloud exit plans and foreign-jurisdiction exposure that a supervised bank will recognise from its own resilience reviews. A widely-shared summary of the UK regulator's new Mills Review maps four AI shifts reshaping financial services by 2030. Gary Marcus argues voluntary vendor safety testing is not real assurance for frontier models. And three banks show different operating-model bets: JPMorgan on dedicated on-premises AI hardware, HDFC on one governed in-house platform, Banco do Brasil on a fully-logged customer-facing copilot.
Perspectives
UK regulator's Mills Review is read as a blueprint for four AI shifts hitting banking by 2030. Independent
A widely-shared LinkedIn summary from fintech commentator Richard Turrin describes the UK Financial Conduct Authority's newly published "Mills Review" as the most comprehensive regulatory blueprint for AI in financial services so far. The summary reports four systemic shifts the review expects by 2030: AI becoming core to every bank function, customer journeys becoming agent-led (handled by AI systems acting on the customer's behalf), market power moving to whoever controls the AI layer between customers and products, and fraud scaling faster than defences. It cites survey figures that 26 percent of UK consumers already trust general AI assistants for financial advice, against 9 percent who use traditional advice. The primary FCA document was not independently retrieved for this brief, so the review's exact wording and status remain unverified.
UK financial-services rule-making often previews themes that later surface in EU and Dutch supervision, which is why a British regulator's framing of agent-led customer journeys and AI-layer concentration lands here before any Dutch equivalent exists. The same concentration and consumer-protection questions the review raises already sit inside the supervisory conversations a Dutch bank is having with DNB and AFM.
Richard Turrin (LinkedIn; original source not verified) (publication date unverified)
Gary Marcus argues voluntary vendor safety testing is not real assurance for frontier AI. Skeptic
AI researcher and critic Gary Marcus argues that frontier AI models (the most capable systems from leading labs) should be tested by an independent standards body before broad release, rather than vetted by the labs themselves. He points to a public proposal attributed to Google DeepMind chief executive Demis Hassabis, under which labs would first submit models voluntarily up to 30 days before release, with a later requirement to pass an assessment before US deployment. Marcus's view is that arrangements run jointly by governments and large technology companies lack the independence, transparency, and mandatory force he considers necessary. This is a reasoned opinion, not measured evidence that any such regime will be adopted.
The argument speaks to a live supplier-assurance question: when a bank relies on a frontier-model vendor, that vendor's own safety attestation is a claim, not independently testable control evidence. The distinction is why this take earns space over the day's louder product news, and it maps onto a gap that already sits inside a supervised bank's third-party model controls.
Netherlands & Sovereignty
Dutch government tightens central cloud rules, requiring exit plans and curbing non-EU providers for critical tasks. Authority
The Dutch government published a tightened central-government cloud policy (Rijkscloudbeleid) on 3 July 2026 that requires public bodies to assess geopolitical risk and single-supplier dependency before using public cloud services. For organisations expected to count as critical entities, it advises against providers subject to law outside the EU or European Economic Area for core tasks, and against putting email and document services in the public cloud. Bodies using important cloud services must prepare an exit plan, and existing services have four years to comply, with limited extensions for complex cases.
The policy binds only central government and does not reach private banks, but it sets a concrete national benchmark for the exact tests a supervised bank already applies to its AI and cloud sourcing: whether a workload has a credible path to another provider, and how much foreign-jurisdiction exposure it carries. Framed as substitutability and control rather than data location alone, it sits alongside the operational-resilience and cloud-concentration questions inside the bank's own resilience reviews.
Industry & competition
JPMorgan picks dedicated on-premises hardware to run agentic AI inside its own walls. Media
Retail Banker International reports that JPMorgan Chase selected SambaNova, a maker of dedicated AI inference hardware, to run secure AI workloads on its own premises rather than through a general cloud service. The bank plans to install SambaNova systems purpose-built for agentic AI (systems that can take actions on their own, not just answer questions). JPMorgan's infrastructure-platforms CIO said the bank will test the architecture's speed and security against demanding enterprise workloads, so this is a stated plan to evaluate the approach, not a proven production result.
The choice is a concrete reference point on a question every regulated bank faces: which AI workloads are sensitive enough to keep on locally controlled hardware, versus reaching a model over the cloud. It earns a place because a peer of this scale committing to on-premises inference marks where the boundary between cloud convenience and control is being drawn for the most sensitive agentic workloads.
HDFC Bank routes all its AI through one governed in-house platform. Media
Business Today reports that HDFC Bank launched Neev, an in-house generative-AI platform meant to serve customer service, lending, wealth management, and internal operations through a single governed layer for model access, controls, and deployment. The bank frames it as an alternative to individual departments building disconnected AI systems, and says it will run across its mobile, online, and messaging channels. No measured outcomes have been reported yet, so the claimed benefits are unproven.
The move is a live example of the central-platform-versus-fragmented-builds choice that sits in front of any bank scaling AI across business lines. Its value here is as a comparison point on operating-model design, with the benefits still to be demonstrated.
Banco do Brasil pilots an AI assistant that logs every customer conversation for oversight. Vendor
Software vendor NiCE announced that Banco do Brasil is piloting its AI assistant for relationship managers, which summarises customer interactions automatically, surfaces customer history, reads sentiment in real time, and creates traceable records of each engagement for governance and compliance. NiCE reports early productivity and consistency gains but discloses no figures, so these are unquantified, vendor-reported pilot claims.
The traceability design is the point of interest: standardised, auditable records of what an AI assistant did in a customer conversation are the kind of control a regulated bank needs before putting a copilot in front of client-facing staff. The pilot shows one control pattern for that, though its benefits remain vendor claims rather than measured results.