BankingNewsAI Daily Brief ·
BoE and FCA find 80% of UK financial firms now run AI agents
Banking AI
Financial institutions & fintech technology
BoE and FCA find 80% of UK financial firms run AI agents
For the chief risk officer at a us bank
80% of UK firms run AI agents; 2% give them broad autonomy, and at most 9% plan to let them move payments or change limits. An agent of yours doing either is the outlier examiners will ask about.
The Bank of England and FCA surveyed UK financial firms on AI. 86% use it, and 80% have agents deployed. Most firms deploy agents that only advise: they recommend an action, and a human decides and executes it. Just 2% grant broad autonomy. Even counting 12-month plans, no more than 9% expect agents to approve customer outcomes, initiate or block payments, or change limits. The model is imperfect. Why did high-materiality systems fall from 16% to 7% of the total? The total grew; the absolute count rose. The exception is vendors, who supply 53% of high-materiality systems, which 81% of firms oversee through vendor assurance and contracts. Ongoing monitoring runs at about half. Peers lock the door tightly and check the room half the time.
→ Action
Model risk: Tag every AI agent by whether it can move payments, approve customers or change limits, and add post-deployment monitoring to each.
Full story →Read article → from Bankofengland
AI cuts some Deutsche Bank corporate loan decisions to minutes
For the head of commercial lending at a regional bank
Deutsche Bank now decides some small-corporate loans in minutes and is building for its agents to deal with clients' agents. Your borrowers will ask why your credit memo still takes weeks, and who answers when their bot calls.
Deutsche Bank now uses AI to make some credit decisions for smaller corporates in minutes rather than weeks. For multinationals, months shrink to weeks. Smaller firms have simpler legal structures, so the paperwork a credit officer once walked through by hand is short enough for software to process. The bank runs this on ADA, an internal platform that tracks agent activity and lets it swap AI models. The fit is imperfect: complex groups still take weeks. The exception is the next step. Deutsche says it is preparing for its agents to deal directly with clients' agents, while major US banks are blocking third-party agents from customer accounts. It plans more than €2.5 billion over three years; JPMorgan budgets $20 billion for 2026 alone. Speed is cheap. Accountability costs.
→ Action
Commercial credit: Set a decision-time target for simple small-business loans and define approval authority for requests submitted by client AI agents.
Full story →Read article → from Globalbankingandfinance
Backbase launches AI agents to speed up banking software delivery
For the CIO of a Backbase client bank
Backbase says its agents double delivery speed and cut spec-writing effort 80%. If your next statement of work bills the old hours, the vendor keeps the savings and your controls still review what the agents wrote.
Delivery Factory gives each job on a Backbase implementation to an AI agent. One turns meeting notes and approved requirements into Jira stories. Another drafts architecture proposals. Others write build plans for the bank's version of Backbase. Engineers then run the plans and check the code in the bank's own pipelines. Backbase says teams across more than 10 implementations worked twice as fast, with 80% less effort on specifications and 40% less on coding and review. Those are the vendor's own numbers. Why would the bill stay the same? It can, if the statement of work bills hours, like a contractor who buys a nail gun and keeps charging per hammer stroke. One exception: the Upgrade Agent aims to make upgrades routine work rather than separate projects. The speed is the vendor's. The sign-off is still yours.
→ Action
Vendor management: Before the next Backbase statement of work, require hours priced with Delivery Factory against the conventional-team baseline, plus SDLC review of agent-written specs.
Full story →Read article → from FinTech Global
General AI
Large language models & AI infrastructure
Anthropic merges Claude chat and Cowork for enterprise users
For the CISO of a bank running Claude Enterprise
On Nov. 16 a setting your bank never switched on becomes the default, and ordinary chats can move into Anthropic-run sandboxes that run code and handle files. Decide before then: opt out, or redo the vendor approvals.
Claude Enterprise has had two modes: chat, and Cowork, where Claude runs code and works with files. Anthropic is merging them. Every conversation starts as a chat. When a task needs code or files, Claude carries it into an isolated cloud sandbox managed by Anthropic, then returns the finished work for review. The user no longer picks. The simple model is a chatbot that sometimes borrows a workshop. The weak spot is where the workshop sits. Your approvals may cover chat, not code execution and file handling on Anthropic's infrastructure. If the feature is off today, why act now? The setting, "Chat and Cowork unified," is off in existing organizations only until Nov. 16, when it turns on by default. Doing nothing is also a decision, and it takes effect Nov. 16.
→ Action
Third-party risk: Before Nov. 16, set the "Chat and Cowork unified" admin toggle or re-approve Anthropic for sandboxed code execution and file handling.
Full story →Read article → from PYMNTS
NVIDIA unveils OpenShell architecture for secure enterprise AI agents
For the head of AI platform engineering at a regional bank
The useful idea is the guard sitting on a separate chip the agent cannot touch. Demand that separation from every agent vendor, even if you never buy the Nvidia hardware sold with it.
OpenShell is Nvidia's blueprint for running AI agents inside a fenced yard. The agent runs on Vera CPUs inside OpenShell software. A separate chip, the BlueField-4 DPU, runs Nvidia Sentry, which watches from outside and stops an agent in milliseconds if it tries to leave its boundary. The simple model: the guard does not live in the house it guards, so an agent that rewrites its own software cannot switch the guard off. The imperfection: Sentry only enforces the policy someone wrote, and a bad policy gets enforced just as fast. The exception: the full design needs Nvidia silicon. BBVA and Santander are only exploring it. Still, the separation is the part worth copying. Ask every agent vendor where its guard sits.
→ Action
AI platform engineering: Require each agent vendor to show where its monitoring and kill switch run, and whether the agent's own runtime can disable them.
Full story →Read article → from Engineering
SpaceX secures $40 billion debt for AI chips, Musk warns neoclouds
For the chief credit officer at a regional bank
SpaceX raised $40 billion for GPUs by pointing to rockets and satellite internet, then built the debt to sit off its balance sheet. A neocloud loan secured by GPUs has neither.
SpaceX borrowed $40 billion to buy Nvidia chips due to arrive in the first half of 2027. Bloomberg reported the deal was built around a special purpose vehicle intended to keep the debt off SpaceX's balance sheet and protect its investment-grade rating. Musk says lenders agreed because Starlink and launch revenue stand behind the chips. The simple model: GPU debt gets repaid from something other than GPUs. The model is imperfect. SpaceX's own AI unit, SI, made $7.56 billion last quarter, up from $813 million a year earlier, so AI revenue carries weight too. The exception is the neocloud, which rents out chips and repays from rent alone. Musk says that credit has almost shut for 200-plus such firms. A used-car lot gets less for each car when every dealer in town sells the same week.
→ Action
Credit risk: List loans secured by GPUs or made to neocloud borrowers, and re-mark that collateral at forced-sale values.
Full story →Read article → from Techstartups
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime