AI Pulse Daily Brief | 2026-10-23
Reading time ~14 mins
ING is rolling out four AI agents to its private bankers and says one saves 45 minutes per business-client onboarding. Danske Bank lets corporate clients' AI agents in as US banks block them, Crédit Agricole builds a group AI platform with Mistral, and CIBC caps a top model on cost. Barclays names developers among 5,000 job cuts. Google Cloud splits who processes Claude data, and three studies show where single-pass checks on AI mislead.
Netherlands & Sovereignty
A Dutch survey finds more organisations keen on AI, while less of their AI use follows policy or goals. Media
Consultancy.nl reported on 13 October on the 2026 AI Trend Study by Berenschot, a Dutch consultancy, run with Nyenrode Business University and other partners. More than 200 people from Dutch organisations took part. The share who consider AI use desirable rose from 65% in 2025 to 75%. Use guided by policy and objectives fell from 27% to 19%, while ad hoc use rose from 32% to 38%. Over the same year, the share saying their organisation has a leading AI vision rose from 74% to 83%. Berenschot sums up the result as strategy first, then AI. The article gives no sampling method or response rate, so part of the shift may reflect a different mix of respondents. On 21 October this brief carried Oliver Wyman's finding that staff route work around AI rules that feel too strict. Berenschot puts a Dutch number on the same gap: more organisations report an AI vision, while a shrinking share of their actual use follows one. Measured by enthusiasm, Dutch AI adoption is rising; measured by use under policy, it is falling.
Industry & competition
ING is rolling out four in-house AI agents to its private bankers and says one saves 45 minutes per business-client onboarding. Media
WealthBriefing, a wealth-management trade outlet, interviewed Anneka Treon, ING's global head of private banking, wealth management and investments. ING has built a Commercial Assistant Suite with four agents. One handles Know Your Customer checks; the others are an investment expert, a sales coach and a meeting-preparation assistant. Treon said the KYC agent saves relationship managers an average of 45 minutes when onboarding a company or other legal entity. Some agents are live in Belgium and others in the Netherlands. ING plans to open the suite to relationship managers in both countries by the end of 2026, with full access during 2027. The investment expert stays an internal tool while ING works through governance and client-protection questions, and the sales coach is still a pilot. Treon said her aim is to make relationship managers more effective, not to cut costs. On conversational investment guidance for clients, she said ING is not there yet. The 45 minutes is ING's own average for one task, and the page carries no publication date. On 6 July this brief reported ING naming KYC among its five agentic AI tracks. The interview adds a measured saving inside a Dutch private bank and a stated boundary on advice. The agents prepare the banker, and investment guidance does not yet reach the client. A study in today's Research section measures why that boundary is hard to move.
WealthBriefing (publication date unverified)
Danske Bank lets business customers connect their own AI agents to its banking systems, while large US banks block such agents. Corporate
Danske Bank said in a press release, numbered for 21 October, that business customers can now bank through AI agents. It calls this a first in the Nordic countries. The service is an MCP server, a standard connector that lets an AI agent use another company's systems. It links the customer's AI tool to Danske's Premium APIs, the interfaces corporate clients already use to connect their own software to the bank. Agents enter through the same authenticated route as any other integration, with onboarding, certificates, credentials and subscriptions. Danske's examples are balance overviews, transaction analysis, tailored financial views and cash-flow forecasts, and the release says customers can ultimately act with a prompt. An agent can also find Danske's interfaces, read their documentation and test them on mock data. Flaminia Franca, head of transaction banking, said the aim is to let customers choose how and where they deal with the bank. Danske reports no take-up yet. On 16 October this brief reported that Schwab blocks agents it has not authorised. US banks now report outside agents acting in customer accounts (see On the radar). The US restrictions target agents that log in with a customer's username and password. Danske admits agents only through the entrance it already secures for corporate software. A corporate treasurer's agent now meets two models at large banks. It is refused when it borrows the customer's login, or admitted through the bank's corporate interfaces under the bank's controls.
Danske Bank (publication date unverified)
Crédit Agricole has given its new group AI company a partnership with Mistral to build one AI platform for customers and staff. Corporate
Crédit Agricole announced on 20 October that its new group AI company has signed a partnership with Mistral, the French AI model maker. The company, Crédit Agricole Artificial Intelligence, pools the group's AI assets. Mistral will help build an AI platform meant to serve all the group's customers and employees. The first shared projects are a conversational assistant for retail customers, agents for key business processes, and cybersecurity, which the bank calls a major focus. The AI company plans to adapt Mistral's models with its own data. Its infrastructure arm, CA-GIP, places each AI workload by how sensitive and critical it is: in the group's private cloud, with European partners or with other providers. Staff can already build and share agents on a separate tool. Nearly nine in ten trained employees of Crédit Agricole S.A. use its internal chat assistant regularly. The release reports no measured results. On 2 October this brief carried ECB President Christine Lagarde naming lost access to US frontier AI models as a risk to Europe's financial system. A large cooperative banking group has answered with a routing rule set at group level rather than a single provider choice. Data sensitivity decides where each workload runs. Model choice, workload placement and agent building now sit in one group company, with a European model maker at its core.
CIBC cut access to a top AI model from thousands of staff to hundreds after costs reached up to C$3,000 per employee a month. Media
Seoul Economic Daily reported on 22 October, citing Bloomberg, that finance chiefs at global banks are clamping down on AI bills that run above plan. Canadian Imperial Bank of Commerce found that staff use of Anthropic's top model could cost up to 3,000 Canadian dollars per employee a month. It cut access from several thousand people to a few hundred. Its own AI workspace, still in testing, is limited to 250 employees at 4 to 250 Canadian dollars per person a month. At one Asian bank, some teams used their whole token allowance in the first week; tokens are the units AI providers bill by. That bank pooled the allowances and built an agent that shortens prompts before they reach a model. J.P. Morgan Private Bank wrote in July that a model from Z.AI, a Chinese developer, matched the best model of March at more than 80% lower cost. United Overseas Bank says its vendor-neutral framework is open to Chinese suppliers. The article also says Canada's banking regulator told its staff to draft minimum standards, security included, for banks' use of frontier models. The figures are secondhand and attributed. Yesterday this brief carried bunq's test of a Chinese-made model at about 60% of the cost of its best option. Cost now limits who inside a bank gets the strongest model. The cheaper route on offer often leads to a Chinese developer, the dependence Bruegel warned about on 21 October.
TD Bank says its multi-agent tool cuts a first-draft commercial credit memo from up to three days to under 30 minutes. Corporate
TD Bank U.S. announced Credit Assist on 22 October at the Evident AI Symposium. The tool drafts six key sections of a commercial credit memo, with citations. More than 300 credit professionals, who write over 10,000 memos a year, review the drafts. A single credit review can involve 20 to 30 deal documents of more than 50 types. An orchestration layer hands the sections to specialised agents that draft them in parallel. TD says first-draft preparation falls from up to three working days to under 30 minutes. It tested the tool on 220 selected deals with input from more than 70 domain experts, and 95% of evaluated outputs passed expert review. Credit staff keep responsibility for underwriting and risk assessment. The figures are TD's own, and the test deals were chosen for the purpose. On 26 August this brief reported DBS putting AI agents into corporate credit memo drafting for about 1,500 staff. TD adds an evaluation design: a deal count, an expert panel and a pass rate. Yesterday's University of Pennsylvania study found a team of agents doing worse than a single agent on the same budget. TD does not say what simpler design its tool was measured against. Its 95% figure measures drafts against expert judgment, the evidence a model-risk review of a credit-memo agent starts from.
Barclays is cutting about 5,000 jobs this year and says about 60% fall in support functions such as developers. Media
Reuters reported on 22 October that Barclays is cutting around 5,000 jobs this year, about 5% of its workforce. It is using AI and other technology to lower fixed costs. Chief financial officer Anna Cross told reporters that about 60% of the cuts fall in support functions such as developers. About 1,000 roles had already gone, and the rest will follow by year-end. Barclays booked a further 201 million pounds of structural cost actions in the third quarter. Reuters called it one of the clearest signs yet of a big bank using AI to reduce headcount in such roles. Barclays does not say how many of the cuts come from AI alone. On 2 October this brief reported Barclays' plan to give half its developers Anthropic's coding tool by the end of 2026. Within three weeks, Barclays has named developers both in its coding-agent rollout and among its job cuts. For any bank rolling out coding agents, the effect on developer roles is now a question for this year's workforce plan. A large European peer has put figures on it.
Innovation
Google Cloud now sells Claude in two forms with the same models, price and bill, but a different data processor. Vendor
Google Cloud announced on 23 October general availability of an Anthropic-managed option for Claude, Anthropic's AI models, next to its existing Google-managed option. Both give the same models on release day, the same pricing and the same Google Cloud bill. In the Google-managed option, Google is the data processor under the customer's existing agreements. It offers controls such as customer-held encryption keys, network perimeters and data residency in the US, EU and Asia-Pacific. New Claude features arrive once Google has integrated them. In the Anthropic-managed option, Anthropic runs the service on Google's infrastructure as data processor, under Anthropic's own commercial terms. Customers get new Claude features at once, such as Managed Agents, in beta, and code execution. They still sign in through Google and keep Google's audit logs. The announcement lists customer-held keys, network perimeters and EU residency only for the Google-managed option. On 22 October this brief carried OpenAI's models arriving on Google Cloud. A team that switches to the Anthropic-managed option for its features changes who processes the bank's data, and under which contract. The console and the invoice stay the same. Under DORA, the EU's Digital Operational Resilience Act, that can amount to a new ICT third-party arrangement.
OpenAI has released a dedicated service for short choice and scoring decisions, billed only on the data that goes in. Vendor
OpenAI made its Decisions API generally available on 22 October. The service returns a probability, a choice or a score for text, images or structured data. It runs on GPT-6 Luna, OpenAI's small model. OpenAI says it answers such bounded decisions up to ten times faster than the same model through its general interface. The price is 10 US cents per million input tokens, 1 cent for input it has cached, and nothing for output. Processing can stay in the US or the EU, and eligible customers can have zero data retention, meaning OpenAI keeps no copy of the data. Box, Datadog, Ramp, Sierra and Speak are named as early users. The speed claim is OpenAI's own. On 12 October this brief noted that the small models of OpenAI and Anthropic both list at 10 US cents per million input tokens. Both charge 50 cents per million output tokens. Banks run sorting steps such as complaint triage, payment-exception routing and document tagging at high volume. For that work, a price set by input length alone, with EU processing, now competes with running the same step on a general chat model.
Research
Changing only a client's risk-tolerance answer moved AI-recommended stock allocations by 10 to 16 points across seven models. Institute
Soovadeep Bakshi, Apurba Bose, Aishwarya Bose and Sandip Nandi posted an audit to arXiv on 16 October. They start from a known problem: people do not answer a risk question the same way twice. In a national US panel, 41% of households chose a different answer two years later. The authors asked seven AI models from five providers for investment allocations on 800 real household profiles, changing only the risk answer. The recommended share of stocks moved by 10.5 to 16.3 percentage points on average. Repeating the same prompt moved it by only 0.6 to 4.2 points. In a scripted test, one clarifying question cut the average swing from 33.5 points to 1.9 when the client corrected the answer. When the client confirmed the change, the allocation still moved. The study measures how much advice varies, not whether any advice was unsuitable or caused harm. In advice, the risk questionnaire is the input a suitability assessment rests on. The audit shows a model treating a one-off answer as settled unless the process makes it ask again. It also puts a size on the swing that follows. ING keeps its investment agent away from clients for now, as today's Industry section reports.
arXiv: Same Person, Different Advice
BNY researchers found AI models reading cheques often reported matching amounts where the two amount fields actually differed. Corporate
Joong Ho Choi, Aditya Paul and Nidhi Singh of BNY posted a preprint to arXiv on 19 October. A cheque states its amount twice, in figures and in words, and the two must match. A mismatch is a red flag for fraud or error. The authors tested two AI models that read images, GPT-4.1 and Claude 4.5 Sonnet, under ten conditions. On cheques where the recorded amounts disagreed, the models often returned two amounts that agreed. With the figures hidden, this happened in 73.7% of cases for GPT-4.1 and 49% for Claude in one condition. The models stayed confident when they did it. A remedy using compressed example images raised Claude's accuracy on 243 cheques, but the gain on a stricter set of 124 was not significant. The authors do not prove how the models fill the gap or how often it happens in production. Comparing two extracted fields is the check document systems use to catch a misread or altered document. The study shows a vision model can make that check pass by producing the agreement itself. Paired fields of the same kind appear in mortgage, identity and payment documents, where the effect has not yet been tested.
arXiv: CARLAR Influence in Multimodal Cheque Extraction
A re-analysis of a public banking-agent test finds the same case ending in different account changes in 14% to 26% of runs. Vendor
Scalac, a firm that builds AI agents for clients, published an analysis on 19 October of public results from a benchmark by Sierra. In it, AI agents serve simulated customers of a fictional bank. It covered six models on 97 cases, each run four times. Identical cases led to different changes to the account in 14% to 26% of cases, depending on the model. In one rewards scenario, 21 of 23 valid runs changed bank records after the simulated customer made a claim the agent never verified. For every model, the share of cases passed in all four runs was lower than the share passed in a single run. For GLM-5.2, a Chinese model, it fell from 37.1% to 13.4%. The bank is fictional, the customer is itself an AI, which adds variation, and Scalac sells agent development services. A score from one run can show an agent passing a case it fails on the next. The 21 of 23 result is a different failure: an agent changing records on a customer's word alone, the opening that social engineering relies on.
On the radar
- Axios reported on 22 October that US banks have detected outside AI agents acting in customer accounts without clear instruction for each step, citing bank sources and no confirmed losses. Axios
- The US Federal Trade Commission demanded records from OpenAI and Anthropic on recent AI safety incidents, safeguards, marketing and privacy, Axios reported on 15 October, with no complaint or finding so far. Axios
- Malicious versions of LiteLLM, a widely used open-source tool that connects applications to many AI models, were live on public download sites for about two hours on 21 October and carried malware built to steal credentials. LiteLLM