BNP Paribas added Google Cloud to its enterprise AI arrangements this week, while Deutsche Bank extended AI into private-bank onboarding and Bank of America reported productivity gains among its developers. Société Générale put a figure on the savings it expects from AI by 2029.

The announcements span different stages of adoption. Some banks can point to tools in use and measures of their effect; others are building the infrastructure, controls and skills for wider deployment. Banks testing purchases made by AI agents are confronting questions about customer authority and liability, while supervisors are examining how AI could change the risks in banks’ loan books.

Read more on the week’s key developments:

1. BNP Paribas plans Gemini integration for AI assistant available to 65,000 staff

BNP Paribas and Google Cloud announced a five-year partnership on 24 September covering cloud infrastructure, Gemini models and tools for developing AI agents. The bank plans to integrate Gemini into LLM@CIB, an internal assistant available to more than 65,000 employees in Corporate & Institutional Banking. Proposed agent uses include preparing corporate credit memoranda. Google’s technology is already used at BNP Paribas subsidiary Nickel, where an assistant helps 200 customer advisers find answers in the business’s procedures.

The Google agreement expands BNP Paribas’ access to Google Cloud infrastructure alongside IBM Cloud hosted in the bank’s own data centres. Its Mistral AI partnership also covers the development of on-premises solutions for sensitive processes such as know-your-customer checks. In 2025, BNP Paribas said it had not placed client data or production environments containing sensitive data in public cloud. The Google agreement says existing security and data-governance rules will determine which data and workloads can be processed there, with certain data categories still excluded. The bank has not specified where its planned corporate-credit agents will run.

2. Deutsche Bank introduces AI to private-bank source-of-wealth checks

Deutsche Bank announced on 22 September that an AI tool for source-of-wealth checks had gone live in its Singapore and Hong Kong private-bank booking centres at the beginning of the month. Advisers in Dubai can also use it for accounts booked in Singapore. Integrated into the bank’s digital know-your-customer platform, it draws on client records and approved external sources to prepare research and documentation, identify gaps and support assessments that staff review. A wider rollout is planned.

Deutsche Bank has already applied agents to another process that requires staff to assess evidence. Its third-party risk tool, introduced in December 2025, reviews supplier documents, cites the passages behind suggested outcomes and leaves decisions with human assessors. The private-bank rollout extends that approach from suppliers to prospective clients, whose records must support a defensible account of how they acquired their wealth. In both cases, the bank is using AI to assemble material for review within an existing control process.

3. Société Générale links AI programme to EUR 500–600 million savings target

Société Générale’s 2026–2029 roadmap projects EUR 500 million–600 million ($570 million–680 million) in savings from AI-related initiatives by 2029, including approximately EUR 350 million ($398 million) it says it has already identified. The bank targets EUR 1.9 billion ($2.16 billion) in gross savings across the group over the period. Its plan also calls for lower IT and procurement spending, simpler operations and a reduction in employee numbers through natural attrition.

The estimate draws on a substantial existing programme. Société Générale says it has around 200 AI use cases in production and has created a group entity, SocGen AI, to coordinate deployment across client interaction, onboarding, back-office work, IT, compliance and productivity. The savings target now rests on whether those applications can help the bank simplify processes and systems across the group.

4. Singapore’s financial firms commit to training more than 80,000 staff in AI

The Institute of Banking and Finance launched its AI Workforce Co-Lab on 24 September with 23 pioneer institutions. They have committed to train their entire Singapore workforce, comprising more than 80,000 employees, in critical AI skills by 2028; more than half have already received training. Participants include DBS, OCBC, UOB, BNP Paribas, Deutsche Bank and Singapore Exchange. The programme will also test changes to jobs and develop pathways for leaders, wealth managers and operations staff.

The work on specific roles takes the programme beyond general AI training. Firms can examine how employees’ responsibilities change when AI prepares client material or handles parts of an operational process. The Monetary Authority of Singapore has also supported industry work on Safeguards for Agentic Finance at Runtime, which proposes controls over actions taken by AI systems. The Co-Lab addresses the skills and oversight those controls will require inside financial institutions.

5. Reserve Bank of India links AI expansion to banking resilience

Reserve Bank of India deputy governor Rohit Jain called for stronger technology governance, cyber controls and oversight of critical dependencies in a 24 September address on resilient banking. He said AI governance should precede deployment at scale and urged banks to maintain visibility across their technology environments, including the external services on which they rely.

The RBI has already issued directions on IT governance and outsourcing and commissioned a framework for responsible and ethical AI use in finance. Jain’s address brought those strands together around banks’ ability to keep services running as technology arrangements grow more complex. AI applications may draw on bank systems, shared infrastructure and outside providers; effective oversight requires banks to understand where those connections create dependencies and how a disruption could affect customers.

6. Six banks seek clearer liability rules for purchases made by AI agents

ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest published principles on 22 September for AI agents that select and purchase goods on consumers’ behalf. Their paper calls for customers to be able to manage the authority they give agents, for auditable records of instructions and payments, and for dispute processes in which liability reflects where an error or risk was introduced. The banks’ voluntary principles leave the allocation of liability to be worked out.

Payment and commerce providers are already building different parts of the transaction. Stripe is developing ways for merchants to present products and complete checkout through agents; Visa has published a protocol for recognising trusted agents. The banks’ paper addresses what happens when those connections produce a disputed purchase. Evidence of the customer’s instruction, the agent’s actions and the merchant’s response will have to be available across organisations that do not share a single system.

7. TD commits up to CAD 25 million to AI work with Cohere

TD Bank Group announced on 22 September that it would commit up to CAD 25 million ($18 million) over three years to AI development and adoption through a collaboration between its Layer 6 research team and Cohere. The partners will examine applications including knowledge management, with Cohere specialists working alongside Layer 6. TD described the work as a way to identify and develop uses across the bank.

TD has previously used Cohere as a supplier to test large language models. Its Layer 6 team gives the bank internal research capacity to assess how external models perform on its own tasks. The dedicated team and funding deepen that arrangement, although the announcement identifies areas for joint work rather than a new application in production. TD has yet to say which uses will emerge from the collaboration or how it will measure their effect.

8. Bank of America reports coding gains across its developer workforce

Bank of America co-president Jim DeMare said at a 23 September conference that the bank was seeing coding-productivity gains of 15%–20% among roughly 19,000–20,000 developers. He also reported fewer internal help-desk calls after introducing Erica-style self-service capabilities for employees. CEO Brian Moynihan had said at a conference earlier in September that the bank expected to double its AI expense budget next year.

Coding assistance sits within a broader programme. In July, Bank of America reported about 200,000 active users of general-purpose AI tools, alongside specialist applications for wealth professionals, corporate and investment bankers, and payments staff. DeMare’s figure locates a measurable gain within that wider deployment, at a scale that could affect how the bank develops software. Its other applications serve different tasks and will need their own measures of impact.

9. Financial Stability Institute calls for scrutiny of AI’s effect on borrowers

Fernando Restoy, chair of the Financial Stability Institute, said in an 18 September speech that supervisors should assess how AI changes the industries banks finance, as well as the risks arising from banks’ use of AI. He called for forward-looking reviews of credit exposures and business models, alongside attention to AI-related cyber threats and interruptions at technology providers. His remarks set out a supervisory argument; they did not announce new requirements.

Earlier BIS research found that business development companies had lent about $115 billion to software firms, while uncertainty over AI’s effect on those borrowers’ revenues had yet to be reflected in lending or equity pricing. The figure describes private-credit portfolios rather than measured exposures at banks. It illustrates the question Restoy wants supervisors to ask of banks’ loan books: whether assessments of future borrower cash flows account for technological change. He also argued that a bank’s ability to adapt its own business model warrants scrutiny alongside its present financial strength.

10. Danske Bank and Mastercard complete an AI-agent purchase in Denmark

Danske Bank and Mastercard announced on 21 September that an AI agent had booked and paid for a coffee-tasting experience on a consumer’s behalf. The purchase used a Danske Bank Mastercard on Mastercard’s Priceless.com platform, with PayOS handling the technical execution. Mastercard said the process required the consumer’s consent and confirmed payment through its Payment Passkeys.

Danske Bank is also piloting a service that lets selected corporate customers connect AI agents to their financial data through the bank’s premium APIs. That pilot concerns permissioned access to corporate data; the Mastercard purchase tested consent and authentication for a consumer payment with Danske as the card issuer. Both place external AI agents at an interface with a banking service, but they involve different customers and controls. The purchase took place on Mastercard’s own platform, leaving its use across other agents and merchants untested.