Citi linked lower technology vendor spending to AI-assisted insourcing, with 14,000 engineers using its agentic coding tool. Its disclosures also covered faster product launches and process redesign, offering evidence of how AI is changing work allocation and service delivery.

The Hong Kong Monetary Authority (HKMA) explored accountability for AI agents, systems that can plan and execute connected tasks, while JP Morgan addressed outside access and the visibility of AI benefits in earnings. Other developments extended AI into mortgage processing, financial crime investigations and customer interactions.

Read more on the week’s key developments:

1. Citi links AI-assisted insourcing to 30% lower technology vendor spending

Citi Chair and Chief Executive Jane Fraser linked a 30% year-on-year reduction in technology vendor spending to AI-assisted insourcing on 13 October. She said 14,000 engineers use the bank’s Devin agentic coding tool, making it more efficient to perform work internally. Fraser also reported product launches running 30–40% faster and said around 100 manual processes had been, or were being, redesigned using AI and automation.

Fraser said financial benefits from AI and automation significantly outweighed investment year to date, although she did not quantify either in that exchange. In its Services business, AI tools supporting client service employees operated in 82 countries, with client response times improving by 21%. The disclosures describe changes in sourcing and service delivery alongside coding productivity. The vendor spending reduction concerns one cost category and does not represent an equivalent decline in total technology expenditure.

2. HKMA explores business accountability and an industry registry for AI agents

The HKMA’s 15 October concept note sets out preliminary considerations for governing AI agents, including accountability for the business functions responsible for their objectives and an industry-level registry of material agents. It also addresses operating boundaries, monitoring and intervention as agents access data, use tools and execute connected tasks. The note accompanies updated guidance under the Accountability, Resilience and Trust framework.

The proposals draw on observations from the second cohort of the HKMA’s Generative AI Sandbox, a programme for testing financial-sector applications, where more than 60% of use cases explored agentic AI. The HKMA’s 2019 principles already held boards and senior management accountable for AI-driven decisions. The concept note examines responsibility across agent workflows and visibility over supporting models, tools and third-party services. Proposed safeguards include suspending agents’ access or activating emergency stops before actions materially affect customers. Further testing through Sandbox++ will inform supervisory expectations.

3. JP Morgan limits outside agents as AI gains remain difficult to isolate in earnings

JP Morgan is blocking outside AI agents where they access accounts through credential sharing and screen scraping, Chief Financial Officer Jeremy Barnum said on 13 October. Chairman and Chief Executive Jamie Dimon cited risks to clients, data and liability while confirming that the bank uses agents internally and intends to use them for customers. Barnum said the bank was working on enabling access securely.

Barnum also said AI benefits were not yet observable in quarterly earnings at an institution of JP Morgan’s size. Earlier benefit estimates included established machine-learning applications, particularly consumer fraud prevention, while current opportunities increasingly concern generative and agentic AI in software engineering and other functions. He described experiments attempting to rewrite small legacy tools within a couple of weeks, with mixed results. Barnum said he was reluctant to quantify the benefits and noted that the cost of model usage would offset some productivity gains.

4. Starling builds transaction model for fraud detection and financial assistance

Starling announced on 14 October that it was developing a transaction foundation model using more than 15 billion anonymised transactions, initially to support fraud detection and subsequently its financial assistant. Developed with NVIDIA support and trained on Civo’s UK cloud infrastructure, the model is intended to enter production in early 2027. Planned assistant applications include more personalised budgeting guidance.

Group Chief Information Officer Harriet Rees said Starling wanted to connect its AI tools through a common intelligence layer trained on customers’ needs. The model is intended to identify patterns in proprietary transaction data for reuse across applications. Rees also linked home-grown cloud infrastructure to European companies’ ability to diversify suppliers and reduce concentration risk. Starling will use NVIDIA’s transaction-model development blueprint and training software alongside Civo’s hosting infrastructure, combining bank-specific data and model development with a separate choice of computing provider.

5. Amalgamated begins deploying AI to assemble money laundering investigation evidence

New York-based Amalgamated Bank has begun deploying an AI agent to prepare evidence for anti-money laundering investigations, technology provider FIS announced on 15 October. Developed with support from Anthropic, the system assembles customer information and transaction records for investigators, links statements to source material and records recommendations and analyst decisions. Employees must explicitly approve case closure. FIS simultaneously made the agent commercially available to US banks, including those outside its existing client base.

Because the agent works alongside existing monitoring and case-management systems, banks can change evidence preparation without replacing alert detection. It compares transactions with customers’ historical activity and expected behaviour before assembling evidence for review. Amalgamated projects that initial alert review time could fall by more than 90%, based on analysis of its workflows against the agent’s anticipated capabilities.

6. TD extends mortgage AI from underwriting preparation to funding and discharge

TD extended agentic AI into mortgage funding and discharge on 13 October, adding capabilities that coordinate document verification, connect information across systems and flag discrepancies. For funding, the technology consolidates client and property details into a case summary, allowing initial review in up to half the time, according to the bank. For discharge, it assembles information for statements that employees check before issuance.

The expansion follows TD’s May deployment for mortgage and home-equity lending pre-adjudication. That application classifies documents, extracts information, calculates income and checks selected policy requirements before generating summaries for underwriters. TD reported early results reducing this preparation from an average of 15 hours to less than three minutes. The additional capabilities extend document preparation and checking into later lending stages, with employees retaining responsibility for judgement and review.

7. FSB strengthens AI action records and supplier-risk practices

The Financial Stability Board (FSB) strengthened provisions for reconstructing AI agents’ actions and managing AI embedded in suppliers’ products when it finalised its responsible-adoption practices on 15 October. Changes following consultation also expanded considerations for independent validation, real-time monitoring and continuity arrangements for critical third-party services. The 12 practices remain recommendations for financial institutions to apply proportionately.

The consultation-response summary identifies documentation as a basis for accountability, traceability and reconstruction of actions, particularly in agentic applications. Supplier provisions extend to subcontracted activities within the AI supply chain and encourage more granular assurances, transparency and exit strategies for critical services. These additions address dependencies beyond the models a bank directly selects. The final framework also clarifies the respective roles of boards and senior management and explains how smaller institutions can draw on existing roles and external expertise when adapting their governance arrangements.

8. Deutsche Bank embeds internal precedents and team standards in AI workflows

Deutsche Bank has deployed AI that incorporates previous work, proprietary information and team-specific standards into document preparation across investment banking and capital markets in Europe and the Americas. Provider Farsight announced the deployment on 13 October. The platform runs within the bank’s technology environment, where bankers generate and refine materials through a conversational interface while sensitive data and workflow activity remain inside its infrastructure.

“Our teams have very specific standards for what goes in front of a client,” said George Tierney, Managing Director and Global Chief Operating Officer for M&A at Deutsche Bank, adding that these vary between groups. The platform incorporates those requirements and precedents alongside banker feedback as outputs are refined. Drawing on previous work and house formats makes institutional knowledge part of document production. The announcement did not disclose user coverage, productivity results or financial returns.

9. NatWest cuts speech-recognition errors across tested UK accents

NatWest reported on 14 October that its adapted speech-recognition system misheard 20% fewer words across the accents tested. The work supports Spending Insights, a conversational service being tested with 75% of Royal Bank of Scotland customers ahead of wider group rollout. NVIDIA Nemotron converts speech into text before NatWest’s proprietary small language models interpret requests. Ninety employees recorded around 3,000 banking questions during initial testing.

NatWest researchers and NVIDIA specialists identified particular difficulties with merchant names, dates and financial terminology. In one test, a Scottish employee’s phrase “last quarter” was transcribed as “last water”, changing the meaning before interpretation began. NatWest subsequently used anonymised contact-centre transcripts to adapt recognition to real banking language. Those transcripts remained within its environment, with no customer data shared with NVIDIA, and adaptations using the bank’s data are ringfenced to NatWest.

10. Mashreq tests AI agents for integration design and troubleshooting

Mashreq is piloting IBM’s Bob software engineering platform across integration design, development, code review and troubleshooting, IBM announced on 12 October. IBM reported early pilot results showing reductions of up to 70% in time spent on selected design and troubleshooting activities. The bank also uses a secure Context Forge gateway handling about 4,000 monthly AI-agent requests for enterprise context, including interface specifications and architecture standards.

The pilot accompanies modernisation of Mashreq’s integration platform, which connects digital channels, payments and enterprise systems and processes an average of 650 transactions per second. Access to bank-specific interface specifications and architecture standards helps engineering teams assess whether proposed changes fit existing systems. Separately, Mashreq is exploring an on-premise deployment of Bob for mission-critical applications that may require locally hosted models to address data residency and regulatory requirements.