// banking
AI in Banking & Financial Services
Agentic commerce rules are being written by card networks and banks, not supervisors — and the first named KYC deployment arrives with a throughput number and no accuracy figures.
standing summary updated September 25, 2026
Where AI in Banking & Financial Services actually stands
The active edge of this sector is agentic commerce, and it is now being contested from two directions at once — neither of them supervisory. On the network side, Anthropic has published a blueprint for retailers to build shopping agents on Claude with Visa and Mastercard attached to the launch, and Visa and Revolut have run what is reported as France's first passkey-authenticated agentic card payment, which put the consent question at the authentication layer. On the issuer side, six global banks including Bank of America and NatWest have now published principles covering scams, fraud and data privacy in agentic commerce. Liability for an agent's purchase is still unallocated, and both sets of moves read as attempts to shape where it lands before anyone is told. Inference
Supervision has still not caught up with any of it. Australia's ASIC and APRA have jointly told financial market entities to move from awareness of frontier-AI risk to decisive action, the Bank of England's governor has framed frontier models as a cyber security and financial stability question rather than a productivity one, and the FCA's finding that younger British investors trust AI tools more than the press, TV and radio or social media influencers still reads as the leading indicator it was. But no supervisor has spoken to agent-initiated payments specifically, and the one concrete governance framework anyone has named for payments AI — ISO/IEC 42001, published December 2023, which covers AI a firm buys as well as AI it builds — was named by a trade publication making an argument, not by a regulator writing a rule. The pattern across this sector is that AI governance is arriving through procurement clauses and industry principles rather than through supervision. Inference
Inside the institutions the shape is familiar — AI pointed at judgement work inside regulated processes rather than at the contact centre — but one long-standing gap has just closed. Deutsche Bank Private Bank has put agentic AI into source-of-wealth checks across its Singapore and Hong Kong booking centres, which makes KYC the first of the compliance processes banks have used AI longest to surface as a named live deployment. It joins treasury and liquidity forecasting, where Barclays and Deutsche Bank are among the banks using an Ant International model for cashflow forecasting and FX liquidity management; corporate credit assessment, where DBS is putting specialist agentic AI in front of roughly 1,500 bankers; and SME banking, where Starling has made an agentic assistant available to all of its business customers. Around them the capital keeps moving — Nasdaq has completed its acquisition of the AI-powered due diligence platform Dasseti, Goldman Sachs has been talking with banks, asset managers, insurers and private credit firms about participating in Nvidia's $500 billion AI financing initiative announced 10 August, and UBS is asking new recruits to demonstrate AI proficiency, still one bank's job spec until a peer copies it.
The gap remains evidence, and it narrowed this week by exactly one number. Deutsche Bank's estimated 30% uplift in client onboarding is the first figure this sector has attached to a named process in a named jurisdiction, and it is worth less than it looks: it is the bank's own estimate, it measures throughput rather than whether the checks are correct, and no false-positive rate, error rate or regulator file review has been published alongside it. Nothing has landed at all on fraud detection or credit models running in production. It was a thin week — three items, all single-sourced from trade outlets, all carrying an Unverified label — and what did not happen in it is the more useful summary: no regulator named a standard, no supervisor addressed agent-initiated payments, and no bank published the accuracy side of anything it has deployed. Inference
// Big moves
2026-W39Deutsche Bank Private Bank has deployed an agentic AI source-of-wealth solution across its Singapore and Hong Kong booking centres for KYC checks, estimating a 30% uplift in client onboarding — the first of the sector's long-standing AML-adjacent use cases to surface as a named live deployment with a number attached.
UnverifiedFinextrasource →
Six global banks including Bank of America and NatWest have published a set of principles addressing scams, fraud and data privacy in agentic commerce — issuers writing rules for agent-initiated payments before any supervisor has addressed them.
UnverifiedFinextrasource →
PYMNTS argues ISO/IEC 42001, published December 2023, could become the payments sector's AI governance benchmark because it covers AI a firm buys as well as AI it builds. No regulator has named it — the push, if it comes, would arrive through procurement rather than supervision.
UnverifiedPYMNTSsource →
// Daily stories
2026-09-26
NatWest to trial a generative AI tool that shows customers their spending via voice, text and visualsUnverifiedFinextra2026-09-25NatWest plans to trial a fully generative audio-visual AI tool for spending insights. Customers will be able to explore their finances by talking or typing to it.
why it matters
This is a trial, not a rollout. Retail banking leads should watch whether NatWest reports engagement or complaint data, since generated answers about a customer's own money carry more conduct risk than a static dashboard.
2026-09-25
BNP Paribas signs a five-year Google Cloud deal covering Gemini Enterprise and Gemini modelsUnverifiedFinextra2026-09-25Finextra reports BNP Paribas has signed a five-year agreement expanding its access to Google Cloud infrastructure and AI capabilities, including Gemini Enterprise and Gemini models. No financial terms or deployment scope were reported.
why it matters
Another large bank standardising on one hyperscaler's model stack for a five-year term. Concentration risk and exit cost belong in the business case now, not after the first agent ships.
2026-09-24
Deutsche Bank puts agentic AI into source-of-wealth checks in Singapore and Hong KongUnverifiedFinextra2026-09-23Deutsche Bank Private Bank has deployed an agentic AI Source of Wealth solution across its Singapore and Hong Kong booking centres to streamline know-your-customer checks. Finextra reports the bank estimates a 30% uplift in client onboarding.
why it matters
Private-bank onboarding is one of the few AI use cases carrying a hard before-and-after number. Worth watching whether that 30% estimate holds up under a regulator's file review before it becomes a business case elsewhere.
2026-09-23
Six global banks including Bank of America and NatWest publish agentic commerce principlesUnverifiedFinextra2026-09-22A group of six global banks, including Bank of America and NatWest, has put together a set of principles designed to address concerns about scams, fraud and data privacy in agentic commerce, Finextra reports.
why it matters
Liability for an agent's purchase is still unallocated. Issuers moving first with principles rather than waiting on regulators is worth watching — if these harden into scheme requirements, they become integration constraints, not guidance.
2026-09-22
PYMNTS argues ISO/IEC 42001 could become the payments sector's AI governance benchmark, bought AI includedUnverifiedPYMNTS2026-09-21PYMNTS argues payments firms need a way to identify where AI is used and who is responsible for it, and to monitor those systems over time. It points to ISO/IEC 42001, published in December 2023, which sets requirements covering AI a company buys as well as AI it builds.
why it matters
If procurement starts asking for 42001 conformance, the burden lands on vendors as much as on banks. Worth watching whether it appears as a contract requirement before any regulator names it.
2026-09-12
Navy Federal says fraud attempts fell about 25%, but scams need a different signal setUnverifiedBanking Dive2026-09-11Fraud attempts against Navy Federal members fell roughly 25% from 2024 to 2025, the credit union said. Its VP of fraud and physical security, Carrie Foran Sepulveda, said scams are now the bigger focus and detecting them takes "a totally different set of signals, data, tools."
why it matters
Better fraud defenses can redirect criminal effort rather than reduce it. If your loss numbers are improving, worth checking whether exposure has moved into scam categories your existing models were never trained to see.
2026-09-11
Accenture's banking lead says AI spend keeps rising on competitive fear, with returns unmeasuredUnverifiedBanking Dive2026-09-10Accenture found bank leaders continue to increase AI spending because they see it as essential to remaining competitive, global banking lead Mike Abbott told Banking Dive, but returns "haven't exactly been measurable just yet." Banks' AI spending exceeded $40bn last year, by one estimate.
why it matters
AI budgets are still being defended on competitiveness rather than a measured return. Worth deciding now which two or three use cases you would be willing to have measured, before the board picks for you.
2026-09-09
Visa and Revolut run France's first passkey-authenticated agentic card paymentUnverifiedFinextra2026-09-09Revolut and Visa have carried out what Finextra reports is the first passkey-authenticated agentic card payment in France. No further detail on volume, rollout or participating merchants was given.
why it matters
Authentication is the unsolved piece of agentic payments — who consents when software transacts. One live European transaction is not a standard, but issuers should watch whether passkeys become the default consent mechanism.
2026-09-08
UBS asks new recruits for AI literacy, among the first banks to make it a hiring requirementUnverifiedFinextra2026-09-07Finextra reports that UBS is asking prospective employees to demonstrate AI proficiency, making it one of the first financial institutions to set AI literacy as a hiring requirement.
why it matters
A single bank's hiring bar is a thin signal, but job specs move fast once one peer copies another. Worth watching whether other tier-one banks add AI proficiency to requirements this hiring cycle.
2026-09-05
Anthropic launches shopping-agent blueprint for retailers with Visa and MastercardUnverifiedFinextra2026-09-04Finextra reports that Anthropic has launched a blueprint for retailers to build their own shopping agents on Claude, and names Visa and Mastercard as part of the launch.
why it matters
If the card networks are attached to agent-initiated purchases, authorisation, chargeback and fraud rules follow. Worth asking now which of your rails would even recognise an agent as the buyer.
2026-09-03
Nasdaq announced it has completed its acquisition of Dasseti, an AI-powered due diligence and monitoring platform for investment consultants, institutional investors and asset managers.
why it matters
An exchange operator is buying into the manager-selection workflow, not just the trade. If your diligence questionnaires now run through a platform Nasdaq owns, watch where the data and the pricing go next.
2026-09-01
Bank of England's Bailey warns frontier AI models pose a financial stability riskUnverifiedFinextra2026-08-31Bank of England governor Andrew Bailey has warned that frontier AI models carry cyber security risks that could affect global financial stability, per a single trade report. No supervisory measures or timelines are described.
why it matters
A governor framing frontier models as a stability and cyber question, not a productivity one, is usually how supervisory expectations begin. Worth watching whether this hardens into written guidance.
2026-08-28
ASIC and APRA tell Australian financial firms to move from frontier-AI awareness to actionCorroboratedFinextra2026-08-27The Australian Securities and Investments Commission and the Australian Prudential Regulation Authority are jointly urging financial market entities to move from gaining awareness of risks linked to frontier AI to taking decisive action.
why it matters
Two regulators framing frontier AI as an action item rather than an awareness one is a cue to test whether current model-risk governance would survive a supervisory ask. Worth watching whether other supervisors adopt the same framing.
2026-08-27
FCA research finds younger British investors trust AI tools over the press, TV and influencersCorroboratedFinextra2026-08-27Financial Conduct Authority research found that younger Brits trust AI tools more than TV and radio, the press, or social media influencers when it comes to investment advice, Finextra reports.
why it matters
If AI is becoming a first-stop advice channel for younger customers, suitability and disclosure controls built for human channels miss where the guidance actually lands. Worth watching what the FCA does with its own finding.
2026-08-25
Starling Bank launches an agentic 'Starling Assistant' for all business customersCorroboratedFinextra2026-08-24Starling Bank says it has launched an agentic AI-powered 'Starling Assistant' to all of its business customers. The announcement gives no detail on the assistant's capabilities, limits or uptake.
why it matters
A challenger putting an agentic assistant in front of every business customer raises the baseline for SME banking UX. Worth watching what it is actually permitted to do on a customer's behalf.
2026-08-22
DBS to deploy agentic AI for about 1,500 bankers on corporate credit assessmentsUnverifiedFinextra2026-08-21Finextra reports that Singapore's DBS will roll out agentic AI services to roughly 1,500 bankers globally, deployed as specialist AI agents to support corporate credit assessments.
why it matters
Credit assessment is judgement-heavy and regulated, so this is a live test of agents inside a controlled workflow. Worth watching what DBS says about review and audit controls before scoping a similar pilot.
2026-08-21
Barclays and Deutsche Bank among banks using Ant International's AI cashflow forecasting modelUnverifiedFinextra2026-08-21Finextra reports that banking giants including Barclays and Deutsche Bank are already using a new Ant International AI model to improve their cashflow forecasting and FX liquidity management capabilities.
why it matters
Treasury forecasting is a rare AI use case that lands on the balance sheet rather than the contact centre. Worth watching whether these named adopters publish anything measurable on accuracy or liquidity buffers.
2026-08-15
Goldman Sachs canvasses banks, insurers and private credit for Nvidia's $500B AI financingCorroboratedPYMNTS2026-08-14Goldman Sachs has been talking with banks, asset managers, insurers and private credit firms about participating in Nvidia's $500 billion AI financing initiative announced Aug. 10, Reuters reported. Goldman's investment bank can help place the debt, according to the report.
why it matters
AI infrastructure is being packaged as a debt product for mainstream institutional balance sheets. Worth deciding your exposure limits and concentration rules before these allocations arrive as ordinary origination.
2026-08-13
Microsoft strikes deal to put S&P Global data and analytics inside 365 CopilotUnverifiedFinextra2026-08-13Microsoft has struck a deal to integrate S&P Global's AI-ready data, insights and analytics into its 365 Copilot workflows and agentic experiences, per a single-line Finextra report that gives no scope, pricing or timing.
why it matters
Licensed market data moving into the assistant layer changes where analyst work happens — and what a Copilot seat is worth. Worth asking your vendor team which entitlements this actually covers before it lands in workflows.
2026-08-12
HSBC Asset Management has funded London-based AI startup Model ML. Deal terms were not reported.
why it matters
Watch whether banks keep shifting from licensing AI vendors to holding equity in them. One undisclosed deal is a data point, not yet a trend.
// Track AI in Banking & Financial Services
One email a week on what actually changed — in banking & financial services and across the frontier. Verified, labelled, no hype.