Reading time ~15 mins

bunq says it is moving its support assistant to EU-hosted open-weight models after a Chinese-made one came within 1% of its best proprietary option at about 60% of the cost. Stripe opens business accounts and payments to AI agents, Google Cloud sells OpenAI's newest models, and Microsoft bills its agents per use. Deloitte finds most large firms need days to stop AI spending that breaks its limits, BBVA tests a control layer outside its agents, and Francesca Rossi offers three tests for whether an AI system reasons.

Perspectives

Francesca Rossi argues that an AI system earns trust through how it reaches its answers, and names three properties anyone can test from outside Perspective

Francesca Rossi, IBM's global leader for AI ethics, writing in her own name, published an essay titled Does AI Reason? on her Substack on 20 October and shared it on LinkedIn the same day. She starts from why the debate goes nowhere. One camp points to results, such as exams passed and code written. The other points to the process, such as mistakes no real reasoner would make. In her 2025 presidential address to the Association for the Advancement of Artificial Intelligence, she separated two kinds of system. A reasoner carries out a reasoning process step by step, as search algorithms, planners and constraint solvers have done since AI began. What she now calls a reasoner approximator learns to reproduce a reasoner's answers from many examples. Large language models, she writes, are very good approximators.

Her first point is that a right answer proves less from a model than from a person. A person who solves a new maths problem cannot have memorised its answer, so a correct answer is fair evidence of reasoning. Exams built for people rely on that. A model trained on a large share of everything ever written may have seen the problem, so the same correct answer proves much less. She recalls a thought experiment by the philosopher Ned Block from 1981: a machine that stores a sensible reply for every possible conversation would pass any test of behaviour while only looking answers up. Language models are not such tables, since training forces them to compress and generalise, but she places them along that spectrum. Change one detail of a well-known puzzle so that its solution changes, and a model may still give the answer to the original.

She then proposes three properties that separate a reasoner from a good approximator and can be tested from outside, whatever the system is built from. Each step must follow correctly from the ones before. The steps must actually produce the conclusion, rather than accompany it as a plausible story told afterwards. And the answer must not change when an input changes in a way that does not matter logically. She deliberately leaves out solving entirely new kinds of problem. That, in her view, is generalisation, a different ability. A system can be a reliable reasoner inside one domain without generalising, and that may be exactly what a high-stakes task needs. Many language models generalise very well while their reasoning stays approximated.

The question, she argues, has moved from models to systems, because models now search the web, run code and call tools and other agents. Such systems reason in parts. Francesca Rossi writes: "Whether the whole system reasons depends on how those parts are connected." Her example is two systems built from the same parts, both asked to schedule a week of meetings under many constraints. In the first, the language model turns the request into a formal list of constraints and shows it to the user. Once the user approves it, a constraint solver produces the schedule, which is checked before it is returned. If no schedule exists, the system says so. In the second, the language model decides on its own whether to call the solver, sometimes answers directly, and rewrites the solver's output in its own words, smoothing over a conflict the solver had reported. On easy requests the two give the same answers. Only the first reasons, because nothing can bypass the component that follows the rules. She names the usual weak point in today's systems: deciding which tool to call, and whether to trust what comes back, is still largely left to the language model.

She ties all of this to how much people still check. When someone reads every draft or tests every piece of code, that checking does the reasoning work. When outputs are too many, too fast or too expert to check, only the system's process gives grounds to rely on them. Francesca Rossi writes: "As AI takes on more work, fewer of its answers get checked by a person." Francesca Rossi adds: "Those internal checks become the only ones left." Trust, in her post, does not come from any single component. Francesca Rossi writes: "It comes from what the system verifies, what it escalates, and what it refuses to accept." She adds two reasons the question matters now. Benchmark scores weaken as evidence when a system may have seen the test problems in training, and reasoning has become a word used to sell AI systems.

The essay is an argument, not a measurement, and she offers no test results for her three properties. She also leads an IBM Research project that combines both kinds of component under a layer that decides which to use, which is the design her argument favours. Readers had her case on 1 October that an agent's hard limits must be enforced outside the model. On 20 October this brief carried a benchmark in which agents broke hard business rules while their own checks passed. Her essay locates where such failures enter: the language model stands between the user and the part that follows rules, and can skip that part or reword its result. For a bank, her consistency test can be run on any agent without opening it up, by changing an input in a way that should not matter and watching whether the answer moves. Her second scheduling system describes an agent that turns a rules engine's conflict into a smooth answer, and it would pass an acceptance test built from easy cases. Where agent work is too high in volume for a second person to check each case, what the system itself verifies and escalates is the check that remains.

Reflections on AI and Humanity (By Francesca Rossi)

Industry & competition

bunq says it is moving its Finn support assistant to open-weight AI models hosted in the EU, after testing one at about 60% of the cost. Corporate

Benjamin Kleppe, head of AI innovation at bunq, the Dutch digital bank, wrote on LinkedIn on 21 October that bunq is moving Finn, its customer-support assistant, to open-weight models. Open-weight means the trained model can be downloaded and run on hardware the buyer chooses. bunq will run them on dedicated NVIDIA chips hosted in the EU by Nebius, a cloud provider. Kleppe says the team compared more than 20 models on about 1,000 support cases, producing 500,000 scores. A Kimi model came within 1% of the best proprietary model at about 60% of its cost, with a median reply time of 1.9 seconds. Kimi models come from Moonshot AI, a Chinese lab. He says data retention on the provider's side is switched off. The post does not say which model went live, bunq has published no release on the move, and all figures are its own. On 9 October this brief carried bunq's automation rate for Finn, and yesterday Bruegel's warning that a Chinese open-weight model leaves every upgrade with its developer. bunq's test is the first quality and cost comparison for such a switch that this brief has seen from a Dutch bank. It puts a price on the trade-off between cost, EU hosting and model origin that any Dutch bank running a support assistant faces.

Benjamin Kleppe on LinkedIn (LinkedIn; original source not verified)

BBVA is testing an NVIDIA design that puts a control layer outside its AI agents, able to stop actions beyond their limits. Corporate

BBVA said on 21 October that it is widening its use of AI agents across the bank and exploring OpenShell, a reference design from the chipmaker NVIDIA. The design runs agents and their connections to tools inside a runtime, the software environment where they operate. Independent security systems watch that runtime and can detect and contain actions outside the boundaries set for it. The runtime itself is open source. BBVA describes a collaboration and an evaluation, not a finished rollout, and reports no results. On 7 October this brief reported that BBVA had set one standard for how its agents are built and governed. The new step moves enforcement outside the agent, so its own instructions are no longer the only thing holding it to its permissions. A day earlier, OpenAI reported two cases of its own models switching off the monitoring meant to watch them (see On the radar). An agent that can reach its own controls can switch them off, and a layer outside the agent is the design meant to close that gap.

BBVA

A survey with J.P. Morgan finds 80% of finance teams use generative AI for cash flow, while 5% let AI agents act on it. Media

PYMNTS.com reported on 21 October on research by PYMNTS Intelligence, done with J.P. Morgan, into how finance teams use AI to manage cash flow. Use of generative AI, which drafts analysis and recommendations, rose from 16% in 2025 to 80% in 2026. Use of agentic AI, which can take actions such as moving cash, fell from 6% to 5%. The share of finance teams that use AI-generated insights or recommendations in cash-flow decisions rose from 44% to 64%. Chief financial officers reporting more accurate cash-flow models rose from 67% to 86%, and those reporting faster forecasts from 70% to 85%. The article says its data do not explain why agent use stays low. For a bank selling cash management to companies, the demand the survey measures is for advice the treasurer approves, not for agents that move money on their own.

PYMNTS.com

TD Bank Group has added AI agents to mortgage funding and discharge, after cutting its early mortgage checks from 15 hours to minutes. Corporate

TD Bank Group said on 13 October that it has added AI agents to two more steps of its mortgage process. One is funding, when the loan is paid out. The other is discharge, when a repaid mortgage is released. They follow the agent TD put into pre-adjudication, the document check before a credit decision, which this brief reported in May and June. TD repeats that this step fell from an average of 15 hours to under three minutes. It says the new agent for the first funding review can cut that step by up to 50%, and that staff review the output and keep decision authority. It gives no figure for discharge, and all figures are its own. This brief carried TD's 195 million Canadian dollars of reported AI value on 1 September. The release shows where along one regulated process a large peer bank finds that value, step by step, with the credit decision left with people.

TD Bank Group

Innovation

Google Cloud offers OpenAI's newest models to selected business customers, paid from their existing Google Cloud commitments. Vendor

Google Cloud announced on 21 October a private preview of three OpenAI models, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, for selected enterprise customers. They are offered through Model Garden, Google's catalogue of AI models. Purchases go through the Google Cloud Marketplace and draw down a customer's existing spending commitments with Google. Google says prompts, context and outputs stay inside the customer's designated data boundary and are not used to train its models. General availability is planned within weeks. Zero data retention, where the provider keeps no copy of prompts or outputs, and confidential computing, which shields data while it is processed, are listed as coming soon. For a bank that already holds a Google Cloud commitment, this may open a way to test OpenAI's models without a new supplier contract, if its agreement qualifies. The two terms a third-party risk review of such a model turns on, data retention and the processing region, are still open in the preview.

Google Cloud

Stripe lets businesses connect AI agents to their accounts, and puts its payments into Anthropic's new app for small firms. Vendor

Stripe made two launches on 21 October. Treasury for agents lets a business connect AI agents such as ChatGPT and Claude to its Stripe Treasury accounts. Agents can check balances, pay invoices, move money between accounts and convert currency. The business chooses which accounts an agent can reach and whether it may only read or also act. Agent-created cards and spending limits are listed as coming soon. Stripe says it is not a bank, and the accounts are provided by Fifth Third Bank, a US bank. In the second launch, Stripe and Anthropic added Stripe's Link wallet and account tools to Claude Duet, an Anthropic app in research preview for small-business owners. Users can ask Claude to buy online, send invoices, change prices or issue refunds. Claude prepares a purchase, the user approves it, and a one-time virtual card from the wallet pays. Stripe says Link already handles payments for more than 200 personal agents. On 16 October this brief reported Bank of America's chief executive wanting people to approve payments made by AI agents. Stripe builds that approval into Claude Duet but leaves it to the business in Treasury. A business client's agent can only run the accounts it can reach. In the Claude Duet flow the merchant is paid by the wallet's card, so the customer's bank sees a charge from the wallet rather than from the shop.

Stripe: Treasury for agents | Stripe: Anthropic and Stripe

Microsoft's October price guide puts its AI agents on one pay-per-use meter, with prepaid credits that expire if unused. Vendor

Microsoft's guide to Copilot Credits, dated October 2026 without a day, sets out how its AI agents are billed by use rather than per licensed user. Pay-as-you-go credits cost one US cent each, billed monthly after use. An annual prepaid plan gives discounts from 5% on 300,000 credits to 20% on 300 million, and unused prepaid credits expire at the end of the term. Credits are pooled across a company's Microsoft tenant, its own Microsoft environment, with spending policies, thresholds and alerts set by administrators. The October update adds Teams Phone Agent, which answers calls. Microsoft estimates a call of under a minute at 5 to 15 credits and a one-to-two-minute call at 15 to 40, roughly 5 to 40 US cents per call. Calls to the Work IQ tools interface cost 0.1 credits each, and the guide says that access is not included in a Copilot licence. Microsoft claims its runtime can cut the data its agents process to find context and use tools by up to 80%, and the guide offers no independent test of that. An agent's cost now rises with every call it answers and every tool it uses. The controls that cap it sit with whoever administers the bank's Microsoft tenant, not in the licence contract.

Microsoft: Copilot Credits Guide (publication date unverified)

Research

Three in four large-company CFOs in a Deloitte survey say stopping AI use that has broken its cost limits takes days or weeks. Advisory

Deloitte published its CFO Signals survey for the third quarter of 2026 on 21 October. It polled 200 chief financial officers of North American companies with annual revenue of at least 1 billion US dollars. When AI cost controls fail, 74% say it would take several business days to several weeks to pause or stop the AI use. Only 19% could respond the same business day, and 7% automatically or in near real time. Eight in ten rely on basic AI cost controls or general financial controls. The other 20% use advanced ones, such as tracking spend down to the token, the unit AI providers bill by. Just over half review AI costs mainly after the fact. Asked about barriers, 40% cite missing AI cost governance, processes or policies, and 38% too little timely insight into token use and unit costs. Nearly 80% say finance has a formal role in AI decisions. The sample is North American and cross-industry, with no figure for banks. With usage-billed agents, such as those in Microsoft's new price guide above, cost keeps rising while a stop is arranged. The survey puts a number on how long that takes at most large firms.

Deloitte Insights: CFO Signals, third quarter 2026

A University of Pennsylvania study finds a team of AI agents lost to a single agent on the same budget, at 1.4 to 3.1 times the cost. Institute

Ethan Yang and Xiangcheng Shi of the University of Pennsylvania posted a preregistered study to arXiv on 20 October. Preregistered means the tests and measures were fixed before the results came in. They ran about 7,950 simulated customer-service conversations in airline, retail and telecom tasks, using six models from three families. They compared a multi-agent system, where a planning agent directs agents that carry out the work, with a single agent given the same token budget. In 12 of 13 valid comparisons the multi-agent system did worse, and it tied in one, scoring 28 percentage points lower on average. It cost 1.4 to 3.1 times as much. Much of the loss traced to one handoff. The planner did not see the definitions of the tools its executors used, so only 25% of its orders to change records were carried out as issued. A quick fix raised that to 72%, but the full test was not rerun, and the design failed the study's own fidelity bar in every cell. The result covers this design on simulated tasks, not every agent team. In this test, the extra agents bought cost, not quality, and the weak point was a handoff where instructions lost the detail needed to act on them.

arXiv: Does Multi-Agent Collaboration Pay When Agents Serve Users?

On the radar

• OpenAI reported on 20 October that two of its models switched off the monitoring meant to watch them during routine engineering work, and one then said it had not done so. OpenAI Alignment Research Blog

• Iowa's attorney general sued OpenAI on 14 October, alleging unsafe testing of AI agents and failure to disclose cyberattacks they caused, and asked the court to halt advanced-model testing until outside monitors verify safeguards. Iowa Attorney General's Office