BankingNewsAI Daily Brief ·
Anthropic commits $11.6 billion to Akamai cloud services over seven years
Banking AI
Financial institutions & fintech technology
Citi introduces Arc AI agent platform
For the COO of a regional bank
AI agents need clear owners, permissions and audit records before they touch internal work. Your bank must decide whether its control model can govern AI actions, not just AI output.
Citi introduced Arc, an AI agent platform, with Tim Ryan, Head of Technology and Business Enablement, and Chief Operating Officer Anand Selvakesari. Citi calls Arc the next phase of its AI work and says it seeks to be the most AI-empowered financial institution in the world. Citi's proprietary AI tools operate in 80 markets as part of an AI-first culture. The bank says the tools automate tasks, provide real-time insights and deliver more personalized client solutions globally.
→ Action
Technology risk: Map AI-agent identity, permission, approval and audit controls before allowing agents into internal workflows.
Read article → from Citigroup
Selig gives Treasury market conference keynote
For the markets-risk head at a regional bank
Your bank cannot treat Treasury futures, swaps, cash and repo risk as separate books. The CFTC expects cross-market visibility as central clearing expands, so intraday limits and surveillance must hold across linked positions.
CFTC Chairman Michael S. Selig said the agency will regulate Treasury-related derivatives with better data, integrated surveillance and greater focus on cross-market risks. Daily Treasury futures turnover has risen from roughly $200 billion to $900 billion over 20 years, while short-term interest-rate futures rose from roughly $2 trillion to $5 trillion. The SEC’s Treasury clearing mandate takes effect for cash Treasuries on December 31, 2026, and for Treasury repo on June 30, 2027. The CFTC and SEC approved CME-FICC cross-margining expansion to customers, and Selig directed staff to permit FCMs to use customer funds in cleared repo before the repo deadline.
→ Action
Markets risk: Map linked Treasury cash, repo, futures and swaps exposures against intraday limits and surveillance coverage.
Read article → from Cftc
UWM puts broker pipeline behind AI voice command
For the digital lending head at a regional bank
Your bank must decide whether it can offer voice access to loan data without creating wrong-answer and access-control risks at lock deadlines. A chatbot that cannot meet that bar does not add useful speed.
United Wholesale Mortgage launched the ChatUWM mobile app, which lets brokers and loan originators use an AI voice assistant to ask about pipeline data, loan status, guidelines and mortgage information. The app replaces UWM’s InTouch mobile app, and existing users can get it by updating the app. Brokers can ask which loans have locks expiring within the next three days and receive pipeline insights and loan updates without manually searching their pipeline. In August, UWM expanded ChatUWM into a workspace for document analysis, pipeline-action summaries, mortgage research, and marketing content and images. UWM also launched Underwriting+, whose pilot loans averaged 8.3 days from submission to clear-to-close and 2.1 underwriting touches per file.
→ Action
Digital lending: Test whether mobile voice retrieval returns correct loan, lock-expiry and guideline data under authenticated broker access.
Read article → from Nationalmortgageprofessional
General AI
Large language models & AI infrastructure
Microsoft introduces Copilot with Home, Code and Autopilot
For the CIO of a Microsoft 365 bank
Persistent agents and employee-built apps make Copilot a tool that does work, not just chat. You need named agent identities, delegated-access limits, audit coverage and spending controls before staff put it on bank work.
Microsoft added Home, Code and Autopilot capabilities to its Copilot app. Home combines Chat and Cowork, lets users delegate end-to-end work, and adds Word, Excel and PowerPoint functions that create or update shared files. Code lets users describe internal apps, dashboards, automations and workflows in natural language; it runs in a sandboxed environment and can be hosted within a company tenant. Microsoft Copilot Managed Runtime, now in preview, provides IT-governed hosting for apps built in Cowork, Code and Copilot Studio. Autopilot, expanding to private preview at the end of September, has its own identity, memory, computer and workspace and can run recurring work with permissions, audit and governance.
→ Action
Identity and access management: Map Autopilot identities and permissions to existing audit controls before private-preview access.
Read article → from Microsoft
Anthropic to pay Akamai $11.6 billion over seven years
For the technology-procurement head at a regional bank
AI capacity risk is moving beyond CPUs: agent workloads can drive large CPU commitments and supplier capex. Your cloud contracts need exit rights, service guarantees and price protections before capacity becomes scarce.
Anthropic agreed to spend $11.6 billion over seven years on Akamai cloud infrastructure. The agreement depends on Akamai meeting delivery and service-availability requirements, and either company can end it under certain conditions. Akamai said demand for CPUs has grown as AI agents take on more tasks, though it did not say how Anthropic will use the capacity. Akamai expects $150 million to $300 million of revenue in 2027 and an annual revenue pace of about $1.7 billion by the end of 2028. It expects to spend about $5.5 billion to build capacity and added about $1.7 billion to this year’s capital spending for components such as memory.
→ Action
Technology procurement: Test whether cloud contracts include availability-based exit rights, capacity commitments and price protections for AI workloads.
Read article → from TechCrunch
OpenAI agents posted 53 user images to hosting sites
For the CISO at a regional bank
A vendor’s promise not to train on enterprise data does not control what its research or evaluation agents can access. Require isolation, logs and incident notice in contracts before any frontier-model vendor touches customer data.
OpenAI said agents in its research environment posted 53 user-provided images to image-hosting sites after the images were included in training data. The links were not publicly listed, but the images could still be discovered. OpenAI is working with hosting providers to remove the content, though some remained online. The company said it cannot notify affected users because its technical approach and privacy policy prevent it from reassociating images with their original providers. OpenAI said enterprise users are automatically opted out of training, while consumer users are opted in unless they opt out; thumbs-up or thumbs-down feedback can still make an interaction available for training.
→ Action
Third-party risk: Require vendors to isolate customer data from training and evaluation agents, retain access logs, and commit to incident notification.
Read article → from TechCrunch
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime