Reading time ~15 mins

The EU cites a standard for the quality-management duty of companies that build high-risk AI, and Reuters reports that the ECB and the Bank of England are pressing banks on AI in their recovery plans. The Dutch Tax Administration reverses its move to Microsoft's cloud. Starling trains its own transaction model, and a commodity trader prepares its bank guarantee requests with an AI agent. Wharton and McKinsey find AI returns thin on evidence and sensitive to usage costs, as BCG reports token prices up more than fourfold. A Moody's paper sets out how banks can validate AI agents, and MIT's expert panel rates finance highly vulnerable to every AI risk it assessed.

Regulatory

The EU cites a European standard on quality management for companies that build high-risk AI systems. Authority

On 14 October the European Commission published Implementing Decision (EU) 2026/2274 in the Official Journal, and it entered into force the same day. It cites EN 18286:2026, a standard on quality-management systems for providers of AI systems. The standard includes post-market monitoring, the tracking of a system once it is in use. Under Article 40 of the AI Act, a high-risk system that conforms to a cited standard is presumed to meet the requirements that standard covers. The decision lists this one standard. Secondary sources describe it as the first AI Act standard to reach this stage, although the decision itself does not say so. Article 17(4) is the part that reaches banks. It treats a financial institution's governance rules under banking law as meeting the AI Act's quality-management duty. Three parts are excepted: the risk-management system, post-market monitoring and serious-incident reporting. The same paragraph says cited standards are to be taken into account. A bank that builds its own high-risk system, such as a model that scores the creditworthiness of individuals, keeps its banking governance for most of that duty. For the three exceptions, EN 18286 is now the named benchmark, with the main high-risk duties due from December 2027.

European Commission: Implementing Decision (EU) 2026/2274

Reuters reports that the ECB and the Bank of England are pressing banks to plan for AI-driven market disruption. Media

Reuters reported on 13 October that the European Central Bank and the Bank of England are pressing banks to improve their recovery plans. It cited three people familiar with supervisors' thinking. According to those sources, supervisors fear that rapid advances in AI could disrupt financial markets sooner than previously expected. The report links the urgency to European banks' limited access to Mythos, Anthropic's most capable AI model. Asked for comment, the ECB pointed Reuters to its July letter to banks and to a June speech by Frank Elderson, a member of its Executive Board. Neither supervisor has published anything that confirms the dialogue, so this account rests on Reuters' sources alone. The July letter gave significant euro-area banks until 31 October to submit action plans against AI-enabled cyber threats, as this brief reported on 8 July. A recovery plan is a different document. It sets out how a bank would restore its finances after a severe shock. The sources did not say whether they meant that formal plan or resilience planning in general. If they meant the formal plan, supervisors' AI questions are moving from cyber defences to a bank's capital and liquidity. That shift comes in the same weeks that the cyber action plans fall due.

Reuters

Perspectives

A Moody's paper argues that banks must validate the whole system around an AI model, now that US model-risk guidance leaves agents to them Perspective

Jakub Szarmach, an AI governance practitioner, shared on 13 October a white paper from Moody's, From Principles to Practice: Governing Generative and Agentic AI Systems, written by Nils Grevenbrock, Zhengpu Zhao and Nihil Patel. Jakub Szarmach writes: "You cannot govern an AI agent by evaluating its model alone." The paper starts from a gap in US rules. On 17 April the Federal Reserve, the FDIC and the OCC issued revised model-risk guidance, SR 26-2, which replaced SR 11-7, for years the reference text on model risk for banks well beyond the US. The new guidance leaves generative and agentic AI outside its scope as novel and fast-moving. Each bank decides for itself how to govern them.

The paper treats the governed object as a whole system. One part is a third-party model the bank cannot see inside, which the paper calls the backbone. The other is the harness the bank builds around it: prompts, output formats, settings, tools, memory and the logic that links the steps. The bank answers for the whole, the paper argues, as an aircraft maker answers for an aircraft whose engine a supplier certifies. Jakub Szarmach writes: "Governance must cover the entire system, including prompts, tools, orchestration, memory, and execution paths." One consequence is that a register of approved models can miss generative AI that reached production through an ordinary software change. The paper suggests searching gateway logs, procurement records and code repositories, and asking engineering teams. A changed prompt or a new tool can shift outputs as much as a new model.

Szarmach singles out three of the paper's methods. The first measures how much a system's answers vary when it runs the same input many times. Citing published studies, the paper says that turning a hosted model's randomness setting to zero does not make it repeatable. Jakub Szarmach writes: "Average performance can hide serious failures in the distribution's tail." Alert thresholds and retry limits are then set from the measured spread, and a vendor's silent model update shows up as a shift in it. The second judges an agent by the path it took as well as its answer. Jakub Szarmach writes: "A correct result can conceal unacceptable actions." The paper's example is a run that reaches the expected final action while skipping a required check. The third, which the paper calls Anchored Continuous Evaluation, lets an AI model score the live output while its measured bias against expert labels is subtracted. It holds only while the labels stay independent, the labelled sample stays representative and the scoring model is unchanged.

The paper also grades changes. A new model version, tool permission or data source means measuring everything again, and a new scoring model means revalidating the measurement itself. It says the people who validate these systems need independence and the standing to halt a deployment, and that expertise in generative AI is often their weakest point. The paper states its limits. It leaves safety and fairness to separate governance for each use, and it reports no results from applying the framework. Its parts are established statistics and engineering practice, assembled into one validation cycle.

Readers have had the case for governing the system rather than the model from security agencies and research institutes. This paper adds the model-risk layer: how a validator measures such a system, and when the evidence must be rebuilt. SR 26-2 does not bind a European bank. The paper also maps its controls to the AI Act, including the duty of a bank that uses a high-risk system to monitor it under Article 26(5). A vendor-hosted model that changes without notice is a change the bank never approved, and only a measured baseline makes that change visible.

Moody's via LinkedIn (By Jakub Szarmach)

Two researchers argue that firms favour AI that replaces people because headcount savings are the easiest gain to count. Institute

Gabriel Unger and Arvind Narayanan, researchers linked to Princeton University, published an essay on Normal Tech on 14 October. They argue that AI research and investment lean toward replacing workers. Their explanation is measurement. Headcount savings are easy to model, and benchmarks that test a system working alone are easy to score. The value of people and AI working together is harder to measure, so it gets built less. They propose a trilemma: a system that is fully autonomous, reliable when the stakes are high and useful across many tasks is hard to build all at once. Their alternative is systems whose working steps people can see, that respond quickly and that people can interrupt and correct as they go. They would test such systems as teams, on real work. The essay is a research agenda. It reports no new data and no comparison showing that collaboration pays better than automation. Tonight's Wharton survey supports the measurement point. Hours saved is the most common way departments there judge AI investment, at 40%, and net-new revenue the least used major measure. On the authors' reasoning, a bank case that counts only hours saved pushes toward autonomy in credit and financial-crime work. Those are the high-stakes settings where they say autonomy is hardest to make reliable.

Normal Tech: Steering AI beyond automation

Tony Moroney argues that an AI agent's decisions not to act should be recorded as decisions. Advisory

Tony Moroney writes a newsletter on digital change in financial services. In its 93rd issue, posted on 10 October, he says the question for firms is no longer whether AI can create value. Tony Moroney writes: "It is: Can the enterprise expand its reliance on AI without losing the ability to observe, stop, attribute and challenge what happens?" Readers had much of his case on governing agents while they run. Two parts are new. First, he argues that an agent's choice not to act should be recorded as a decision. A deferral, a rejected recommendation, a threshold crossed or evidence judged insufficient would each be kept with its reason and the condition for looking at it again. Too many such rules would slow everything down, he warns, while too few let uncertainty turn into action. Second, he sets five conditions for human oversight to be real. The overseer needs enough knowledge to spot weak reasoning, enough practice to keep their judgement, time to investigate, authority to intervene and the confidence to say no. He cites surveys and analyses without naming them and reports no results of his own. In bank case work such as financial-crime alerts, the case an agent chose not to escalate is the one a supervisor or a complainant asks about. A log that records only actions holds no answer to that question.

Tony Moroney via LinkedIn (shared by Tony Moroney)

Netherlands & Sovereignty

The Dutch Tax Administration drops its move to Microsoft's cloud and will run mail on its own servers from 2027. Media

ZDNet reported on 13 October that the Dutch Tax and Customs Administration has abandoned its planned move to Microsoft 365 in the public cloud. About 5,000 employees had already moved when the Advisory Council on ICT Assessment, the government's review body for large IT projects, recommended reversing course in June. The government paused the rollout in July. State Secretary for Finance Eelco Eerenberg told Parliament that mail and calendar will run on the administration's own servers in 2027. European open-source tools for personal storage and collaboration follow later in 2027 and 2028. The administration keeps its existing Microsoft software for now and has not chosen a replacement. The decision covers one agency that handles tax data under strict rules, not the whole government. It is still the clearest Dutch case of an organisation reversing a move to US cloud software after staff had migrated. DORA, the EU's digital operational resilience law, requires banks to keep exit plans for critical technology providers. Those plans now have a Dutch public example of what an exit involves and how long it takes. Microsoft's AI assistant features depend on its cloud, so moving mail back in-house likely removes them for those staff, a cost the report does not discuss.

ZDNet

Industry & competition

Starling trains its own AI model on 15 billion customer transactions, on a UK-owned cloud. Corporate

Civo, a UK cloud provider, said on 14 October that Starling, the UK digital bank, is building a transaction foundation model with support from NVIDIA. A foundation model is a large AI model trained on broad data and then adapted to many tasks. Starling's will learn from more than 15 billion anonymised customer transactions. It will be trained and hosted on Civo's UK sovereign cloud, which keeps the data and the computing under UK ownership. Starling plans to put it into production in early 2027, first to improve fraud detection and then to support its customer assistant. It builds on an NVIDIA blueprint for transaction models. Starling links the choice to the UK government's July 2026 plan for AI in financial services. The plan warns about reliance on a few global AI and cloud providers. Its group chief information officer, Harriet Rees, co-authored that plan. No results exist yet, and the announcement comes from the cloud provider. On 13 October this brief reported that Plaid runs a model trained on transaction histories in live US credit products. Plaid sells its model to lenders, while Starling builds its own from its own customers' data. Those are now the two routes open to a bank with years of transaction records, and NVIDIA's blueprint lowers the cost of the second.

Civo

A commodity trader now prepares its bank guarantee requests with an AI agent and reports processing times cut by more than half. Corporate

Centalion, a commodity trading group, said on 8 October that it runs an AI agent in its live treasury work. It made the announcement with Komgo, a trade-finance platform that links companies and banks. The agent prepares requests for guarantees and standby letters of credit, the bank undertakings that back a trader's obligations to its counterparties, and for changes to them. It combines an AI model with Centalion's own business rules and reaches Komgo's platform through a standard connector for AI agents. A treasury employee still approves each transaction. On its first use, a facility shared by several banks, Centalion reports issuance processing time down about 55% and amendment time down about 80%. Extended across its trade finance, it expects to free more than 60 treasury working days a year. All figures are Centalion's own. The steps the agent speeds up are the requests that arrive at the banks on the other side. A trade-finance desk serving such a client now receives requests prepared by software, at the client's pace, whether or not the bank uses agents itself.

Centalion Group

Research

Wharton and McKinsey both find AI's business case thinner on evidence and more exposed to usage costs than headline returns suggest. Institute

Two reports published this week measure the same weak point from different sides. Wharton Human-AI Research and GBK Collective, a research consultancy, surveyed 1,416 employees of large US companies in July for their fourth annual report. 80% of managers reported some positive return from AI, and 28% strongly positive returns. Only 7% of organisations met all three of the report's success markers: formal proof of return, changed workflows and alignment across the organisation. 64% of managers called rising costs for tokens, licences and access to AI models a major barrier to using more AI. Tokens are the units in which AI model use is billed. Only 40% were very confident they could track total AI spend once costs bundled into other software are counted. 67% of departments use AI agents in some form, but only 17% let them act without a person approving each step. All figures are self-reported.

McKinsey's AI unit, QuantumBlack, published a task-level model on 14 October. It compares what a task costs today with the full cost of doing it with AI, including the human review and rework still needed. It puts the yearly global value at $2.52 trillion, with financial services the largest pool at $530 billion. A 30% rise in AI usage costs cuts that to $1.67 trillion, a drop of 34%, and pushes some automation cases into loss. The model leaves out one-off transformation costs, and McKinsey calls it directional. Read together, the survey shows that most managers are not fully confident of their total AI spend, and the model shows how much of the value depends on it. A bank case that assumes today's model prices will keep falling rests on the number only 40% of managers feel sure of.

Wharton Human-AI Research and GBK Collective: The Adaptive Advantage | McKinsey: The AI value curve

BCG reports that token prices for leading US AI models rose more than fourfold, and proposes booking token spend by what it pays for. Advisory

Boston Consulting Group published an article on 2 October on how AI is reshaping technology budgets, which Tony Moroney shared on 12 October. BCG writes that average token costs for the top US AI models rose from below $5 per million tokens in early 2025 to $20 in 2026. It calls that a snapshot of a market whose pricing keeps changing, and it also cites server prices peaking at 80% above 2025 levels. Its scenario model finds that technology costs settle higher whenever business units adopt AI faster than the IT function captures AI savings in its own work. In a modelled industrial company, IT cost stays at 2.6% of revenue when both move quickly. It rises to 2.8% when the business runs ahead. BCG names financial services among the sectors with the most room to capture such savings. It proposes booking tokens by purpose. Tokens that build reusable capability would count as capital spending, those that run internal work as operating spending, and those inside a product a customer buys as cost of goods. The scenarios rest on BCG's own stated assumptions, and its advice on managing AI spend repeats points readers had in September and October. The accounting split is new. Where token spend is booked changes reported costs and margins, and banks are judged closely on their cost-to-income ratio.

Boston Consulting Group via LinkedIn (shared by Tony Moroney)

MIT's expert panel rates finance and insurance highly vulnerable to all 24 kinds of AI risk it assessed. Institute

On 14 August this brief reported that a panel of 272 AI experts ranked finance and insurance among the five sectors most exposed to AI risk. Peter Slattery, an MIT researcher and one of the study's authors, posted the sector's own results on 14 October. The study used three Delphi rounds, a method that refines expert judgements over repeated rounds, to rate 14 sectors against 24 subdomains of AI risk. Peter Slattery writes: "The experts rated the Finance and Insurance sector as being highly vulnerable across all 24 AI risk subdomains." Seventeen received a median rating of extreme and the other seven of high. Fraud and scams led, with 92% of the experts rating the sector extremely vulnerable. Discrimination followed at 87%, then governance failure and loss of privacy at 81% each. Around 80% gave the same rating to AI-enabled cyberattacks, AI systems that fail or behave unreliably, and security flaws in AI itself. The panel did not agree on power concentration or environmental harm. These are expert ratings, not counts of incidents. Peter Slattery writes: "These findings suggest that the sector faces a broad range of AI risks, extending well beyond fraud and cybersecurity." Supervisors' AI attention this year has centred on cyber threats and fraud. The panel puts discrimination, which in a bank shows up first in credit decisions, almost level with fraud.

MIT AI Risk Initiative via LinkedIn (shared by Peter Slattery)

On the radar

• Korea's Financial Services Commission said on 14 October that 16 financial companies have had recent data breaches exposing about 100,000 customer records, that investigators are checking whether the cases are linked and whether AI was used, and that it plans to require more frequent vulnerability testing. Financial Services Commission