BankingNewsAI Daily Brief ·
US let Nvidia sell AI chips to an empty office in China
Banking AI
Financial institutions & fintech technology
BMO launches AI agent to help credit card customers get answers
For the head of card servicing at a regional bank
BMO puts numbers on it: 500,000 card calls a year, 12-14 minutes each, a 20-30% cut on eligible calls. It never says how many calls are eligible. Your business case should.
AI Agent Assist is a tool for BMO's call-centre staff, not its customers. When a client calls about a card transaction, the tool matches the transaction, pulls the relevant policy and prepares the case. An employee reviews it and decides. BMO gets more than 500,000 such calls a year from Canadian clients, and each takes 12-14 minutes. The target is a 20-30% cut on eligible calls, about three or four minutes off a typical call. BMO also aims to process about 40% of eligible refund requests automatically, so some credits land the same day. The simple model is fewer minutes per call. The flaw is the word "eligible," which BMO never sizes. And "automatic" refunds still pass a human. The savings come from searching less, not deciding less.
→ Action
Card servicing: Measure current handle time and refund-request share by call type, so AI vendor claims can be tested against a defined eligible pool.
Full story →Read article → from Finextra
Deutsche Bank says built-in AI controls cut production time to as little as three months
For the chief data officer at a regional bank
Deutsche Bank kept every control and cut up to nine months. If your AI projects take a year, the likely culprit is not model-risk review but each team meeting it alone, late, from scratch.
Deutsche Bank built one pre-approved road for AI projects. Teams get approved environments, data-access routes, reusable connectors and blueprints, with security, data protection and responsible-use checks at each stage rather than just before launch. Its AI Access Layer gives every application and agent one controlled route to approved models. Projects on that route went from about 12 months to as little as three. So were the controls the delay? Mostly not. The delay was every team building its own kitchen before cooking. The figure is the bank's own, and 'as little as' describes the best cases. The exception is projects that do not fit the route. For your bank, the question is where approvals happen, not how many.
→ Action
Model risk: Map where each AI approval enters your lifecycle and move data, security and responsible-use sign-offs to project intake.
Full story →Read article → from Db
Foster draft bill would have Fed test real-time supervisory software
For the head of government relations at a regional bank
The Fed would ask banks to help design the software that defines real-time reporting. Formats written for a demo tend to become the rule later. Staying out means filing in someone else's format.
Rep. Bill Foster's draft would have the Federal Reserve Board and other regulators build, test and publicly release open-source software. The software would check whether supervisors can see a bank's health in real time while making reporting cheaper. The Fed would ask banks, technology firms and researchers for technical help. Picture replacing hand-built report files with a shared pipe that carries data in standard formats. The simple model: faster data lets supervisors act before stress spreads, and shared tools cut small banks' manual work. The flaw: a regulator who sees stress the hour it starts will want an answer that hour. Trump opposes mandatory AI rules, but Foster is in line to chair the Financial Institutions Subcommittee. A demo needs no mandate.
→ Action
Regulatory reporting: Name an owner and budget to submit technical input on data formats if the Fed opens contributions under Foster's draft.
Full story →Read article → from PYMNTS
General AI
Large language models & AI infrastructure
METR report finds AI is speeding up Anthropic's own AI development
For the chief risk officer at a bank using Anthropic models
Your annual review of an AI vendor checks a model once a year; METR says Anthropic made two years of progress in the last one and finished three unreleased models in six months. Tie revalidation to releases, not the calendar.
METR spent 11 weeks inside Anthropic measuring how fast its models improve. Roughly twice as fast as a year ago: about two years of progress in one. METR says most of the speedup, as of October, came from AI helping Anthropic's researchers, not more computing power. Simple model: the vendor moves faster, so your sign-off expires faster. The weak spot: Anthropic redacted most of the evidence METR found most informative. The exception: three models finished training between July and October, and none is released. Why review once a year if the maker changes the product every few months? A yearly inspection fits a car you keep, not one swapped each quarter. The release, not the calendar, is the risk event.
→ Action
Model risk: Add a trigger to revalidate any Anthropic-based model at each vendor model release, not on the annual cycle.
Full story →Read article → from The Verge
Microsoft winds down Oracle AI cloud deal over outdated Nvidia chips
For the head of tech-infrastructure lending at a regional bank
A contract signed in late 2023 for scarce chips lapsed because the chips aged, while Oracle says older GPUs still renew at a premium. Your compute loans likely price the second story; underwrite the first.
Microsoft rented AI servers from Oracle under a deal announced in late 2023 to run an AI version of Bing. It chose not to renew and has been shutting that capacity down, because the servers used an earlier generation of Nvidia chips that no longer fit its workloads. The simple model: a data centre is a building with a tenant, and a long lease makes the loan safe. The flaw is that the tenant rents the chips, not the walls, and chips age like a phone, not like a warehouse. The exception: Oracle says older GPU contracts have been renewed or reallocated at a premium. Maybe. Microsoft is also reviewing over $80 billion in outside cloud commitments. A lease is only as long as the chip stays current.
→ Action
Tech and infrastructure lending: Re-test collateral values and renewal assumptions on data-centre and compute loans by GPU generation and lead-tenant contract end date.
Full story →Read article → from Crowdfundinsider
US let Nvidia sell chips to an empty office in China
For the sanctions compliance head at a trade-finance bank
Commerce approved New Matrix to handle Nvidia H200 chips; the Times found an empty shell, wires dangling where lights should be, and owners tied to a firm blacklisted since 2024. The license is not your end-user check.
The Commerce Department approves which Chinese firms may buy Nvidia's H200 AI chips. One approved buyer, New Matrix, is 17 months old. At its listed address the New York Times found a gray shell, no directory entry, and a property manager with no record. Its owners have ties to a Chinese private equity firm the US has blacklisted since 2024. The Times says it is a pattern. Why would the government clear a buyer nobody can find? One fact sits beside it: limits eased after Nvidia agreed to give the government 15 percent of all sales. A license works like a building pass. It proves someone signed off, not that anyone is inside. It may settle export legality, not who owns the buyer. That question never left your desk.
→ Action
Trade finance compliance: Require independent end-user and ownership checks on AI-chip trades even with a Commerce license; screen owners for ties to blacklisted firms.
Full story →Read article → from The Verge
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime