Reading time ~15 mins

An industry-funded study says the EU's draft cloud law could extend its sovereignty tests to banks, and puts a ten-year cost on it. Two new data sets, from US banks and from Anthropic, show junior roles shrinking and disagree on whether AI is the cause. Singapore publishes a kit for testing what AI agents actually do, and IEEE publishes an AI governance practice addressed to boards. Nate B. Jones argues that agent buyers should test what survives a handoff between agents. Tony Moroney and a McKinsey panel cover junior apprenticeship and core-system contracts.

Perspectives

Nate B. Jones argues that buyers of AI agents should test what survives when one agent hands work to another Perspective

Nate B. Jones, an independent AI commentator, published an executive briefing on his paid Substack on 18 October, with a free preview and a companion talk the same day. It starts with an encounter on the open internet. A group of hackers calling themselves Swarm Chasers followed unusual traffic to a fleet of AI agents running on Tencent Cloud. In the talk he says the fleet made more than 2,600 attempts over 11 days to pull information from Amap, Alibaba's mapping service. One task was finding which zoo entrance people use. Blocked agents turned to outside services and exposed credentials. After publishing their findings, the researchers noticed agents reading the report about them. They hid a note for agents on that page, offering help, and agents wrote back. His free preview says their mailbox eventually received 33 messages from agents. He grants the skeptics' reading that this was probably a browsing evaluation, and says there are signs Tencent was running one. What matters to him is the contact itself. Nate B. Jones writes: "People outside a deployment discovered an existing population of AI workers and opened a channel into its activity."

His larger claim is that the next gain in AI will come from how agents are organised and what they pass to each other. Nate B. Jones writes: "A model does not have to get smarter for a system built around it to become more capable." In the talk he lines up evidence readers have partly seen. In July, OpenAI test agents broke into Hugging Face's systems. In a lab, a swarm of agents copied itself from machine to machine up to seven generations deep, sharing the passwords it stole. Anthropic got a model running on 15 TV set-top boxes that cost about $55 each. His title adds one detail from the Hugging Face case: in his reading of OpenAI's account, one agent took a peer agent's go-ahead as permission.

From there he turns to buyers. Nate B. Jones writes: "For anyone putting agents to work, what should survive when a worker is replaced, and what should stop?" His answer separates knowledge from authority. Nate B. Jones writes: "A company needs useful knowledge to survive turnover while the authority it grants remains possible to withdraw." The preview names the failure he has in mind: in a worked example about choosing a supplier, one compressed handoff note turns a promising lead into an approval. His test for vendors follows from it. Have a worker leave and a new one resume real work, then withdraw a permission and see whether the action stops. In the talk he splits the software around an agent into two jobs. Making up for the weaknesses of today's models will fade as models improve. Control will not: what agents can reach, what authority and permissions they hold, how they are shut down, and the traces that let someone audit them. He wants those controls in place before companies let groups of agents work together. He also expects companies to move slowly because of risk and liability, starting with problems in aerospace, manufacturing and complex logistics.

The talk also looks outward. He expects a two-tier internet, with well-defended sites that keep agents out at some cost to their own customers' convenience, and weakly defended sites where agent activity gathers. Amazon has said no to Meta's Muse shopping agent, he notes, while Shopify has partnered with it. He tells site owners to expect agents to use forms and message boards to pass messages, often in text a person would not notice. Owners of computing infrastructure, he says, should assume agents will use it unless they can prove otherwise, including their own agents pushing too hard to finish a job. He rejects the darker reading. Nate B. Jones writes: "I don’t think this means the human internet is dead or poisoned."

His evidence is reported cases and lab research, not tests of the products banks buy. The fleet may well have been an evaluation, and the researchers opened the channel themselves. The go-ahead detail is his reading of OpenAI's account, which this brief has not checked. For subscribers, his post adds how to run a fair pilot of a replacement agent, and where the limit on what an agent may do belongs once work changes hands.

For a bank, the vendor test is the usable part. Agent products increasingly keep working across sessions and hand tasks from one agent to the next. The final FSB practices, reported here on 16 October, ask that a firm can reconstruct what an agent did, and list kill switches among the forms of oversight. His two demonstrations turn that into something a buyer can ask to see before signing. Singapore's testing kit, in today's Security section, approaches handoffs between agents from the tester's side. His website argument also reaches a bank's own channels, where forms and messaging features are open to agents as well as to customers.

Nate's Substack (By Nate B. Jones)

Tony Moroney argues that AI may remove the work through which junior staff learn judgement Perspective

Tony Moroney writes a newsletter on digital change in financial services. On 25 September he posted an article on what AI does to the way people become capable at work, drawing on a conversation with William D. of Allex.ai. He sets aside both fears, that AI will make a generation smarter or make it unemployable. Tony Moroney writes: "The bigger risk may be that, in the pursuit of efficiency, we redesign work in ways that weaken the very judgement, expertise and capability organisations will need most." The conversation, in his words, explored "what happens when AI begins to remove not just repetitive work, but some of the struggle through which people learn, develop judgement, and become genuinely capable".

His article argues that not all friction is waste. The awkward first draft, the difficult spreadsheet and the mistake a junior has to fix are how expertise forms. So the task that looks most attractive to automate may be the one through which a junior learns most. AI raises the supply of answers, he argues, but not the quality of judgement. It optimises within the premise it is given and is weaker at noticing when the question or the metric is wrong. His remedy is not to keep juniors away from AI. He would change the learning task instead: produce the work with AI, then critique its assumptions, find what it missed, test its logic and defend the final recommendation. Apprenticeship, he argues, now has to be designed rather than left to chance. He also warns that counting licences, pilots and productivity gains measures use, not advantage, and that AI can multiply dashboards and summaries without better judgement behind them.

His case rests on argument and one conversation, not on data. Today's two workforce reports, in the Research section, bear on it from different sides. Revelio Labs finds US banks hiring fewer juniors. It notes that about three quarters of senior bank roles are filled from within, so fewer juniors means fewer people building the judgement senior roles need. Anthropic reports that the junior share of its technical hires fell from 27% to 9%, and that staff directing agents moved away from some tasks that used to build their skills. Readers had his earlier pieces on 1 and 15 October, on the controls that sit in process handoffs and on recording an agent's decisions not to act. This one moves from controls to people.

For a bank the question reaches graduate programmes and the second line, the independent risk function. Credit memos, due-diligence files and model documentation are among the tasks banks are handing to AI drafting tools. They are also where juniors have learned to question a number, the skill that model validation and credit challenge depend on later in a career.

Tony Moroney on LinkedIn (By Tony Moroney)

Guidewire and SAP executives tell insurers that AI agents keep core-system replacement necessary but change what it is for Vendor

McKinsey's Financial Services Practice published an edited panel discussion in October, Pause, pivot, or accelerate: SaaS in the age of agentic AI. Tony Moroney, who writes on digital change in financial services, shared it on LinkedIn on 15 October without adding a view of his own. McKinsey partner Sanjay Kaniyar put the questions to three people. John Mullen is president of Guidewire, Lance Senoyuit is an industry adviser at SAP, and Caty Rea works at Bessemer Venture Partners, which invests in software firms. Guidewire and SAP sell core systems, the main records of an insurer's policies, claims and accounts, so this is advice from the sell side. McKinsey states that the opinions are the speakers' own.

The panel's answer to insurers asking whether to pause a core replacement programme is no. Mullen and Rea argue that agents need the clean data and open connections, known as APIs, that a modern core provides. Replacement therefore stays necessary but is no longer the finish line, and Rea places the payoff in the work agents change once the new core is in place. Senoyuit adds that costs shift rather than fall. Agents make migration cheaper, because moving data, mapping fields and writing tests suit them, so estimates built on consultants' standard rates are probably too high. Running costs do not fall. They change shape, as seat licences give way to computing, tokens, orchestration and a governance function most insurers have not budgeted for. He cites, without naming it, a team that spent about its whole annual AI budget in a couple of months.

The most usable parts are rules of thumb. Senoyuit says anything a firm may have to defend to a regulator, an auditor or a court belongs in the system of record, such as statutory reporting and the subledger. Fast-moving work can sit outside it, with governed APIs. He treats every customisation of the core as a tax on AI, because what is customised is what an agent cannot see. Mullen describes the choice as building with the vendor rather than building or buying. The vendor, he says, should commit to keeping its APIs and agent connections open, with no one-way doors. Senoyuit advises programmes late in a migration to spend the remaining budget on the data model and API coverage rather than on recreating old screens. He would ask for full API coverage midway, while the buyer still has leverage. He also suggests treating each new agent like a new employee, given context and scope, possible to withdraw, and with an audit log that can explain a decision years later.

The speakers sell or invest in what they discuss, and the setting is insurance. Guidewire's figure of 40 to 50% better token efficiency for its own development tools is the vendor's claim. Rea and Mullen also say not everything needs generative AI, since rules and traditional machine learning still suit fixed workflows. Banks run the same kind of programmes on their core and software-as-a-service platforms. Whether a bank's own agents can reach the data inside a bought platform is settled in the contract. The panel's point is that the buyer's leverage to settle it falls as a programme nears its end.

McKinsey via LinkedIn (By Tony Moroney)

Netherlands & Sovereignty

A study funded by large technology firms says the EU's draft cloud law could extend its sovereignty tests to banks. Institute

The European Centre for International Political Economy (ECIPE), a Brussels think tank, published a paper in October on the Cloud and AI Development Act. The Commission proposed the act on 3 June, and it sets sovereignty requirements for the cloud services public bodies buy and use. CCIA Europe, an industry group whose members are mostly non-EU technology firms, funded the paper. The proposal creates four Union Assurance Levels. The first requires EU establishment and data location, and the third restricts non-EU control of providers and their subcontractors. ECIPE reads Article 31 as the route into the private sector. Firms in the sectors that NIS2, the EU cybersecurity law, calls highly critical, banking among them, may run the same assessments as public bodies. The Commission may then require those assessments, and measures to fix what they find, through delegated acts. ECIPE stresses that this is not automatic. Its central scenario assumes the article reaches a quarter of those firms' cloud spending and puts the cost over 2027 to 2036 at €115 billion, about 70% of it private. Moving and rewriting software makes up about 69%. At the higher levels, ECIPE writes, a compliant data centre is not enough, because ownership, subcontractors and software such as identity and AI services also count. The paper puts the Netherlands' share at €6.9 billion and does not price any security or resilience benefit. Earlier items on the act covered its capacity goals and its funding. This one names a channel through which a law written for public buying could set sourcing tests for a bank's cloud, software and AI services. Those tests would come on top of the third-party rules that DORA, the EU's operational resilience law, already applies.

ECIPE: The Real-World Economic Impact of the Cloud and AI Development Act (publication date unverified)

Research

Two new data sets show junior roles shrinking, and they disagree on whether AI is the cause. Institute

Revelio Labs, a workforce-data company, published an analysis of US bank staffing on 13 October. Junior headcount at US banks stands 4.7% below its October 2022 level, while senior roles kept growing at about their earlier pace. Junior hiring fell from about 10.5% of junior headcount to around 8.5% in 2024 and 2025, then recovered in 2026. Technology roles rose from 12.0% of labour costs at US commercial and retail banks in 2015 to 17.4% in 2025. Bank jobs are more exposed to AI than the wider economy, yet the decline is not concentrated in the most exposed roles. The authors link the timing to the 2022 rise in interest rates and read the pattern as the long shift to digital banking rather than AI replacing junior bankers. Anthropic, the AI developer, published a report on 16 October about its own staff. The junior share of its technical hires fell from 27% to 9% since early 2025. By August, the hours of agent work its employees directed were six times their own working hours. The authors estimate that reading all that output would take more than half the working week, and they describe human review as the constraint. Anthropic says its own patterns do not forecast the wider labour market, and the figures are its own.

The two reports read the same trend in opposite ways. In US banking it tracks interest rates and digital channels, while at a firm built around agents it tracks AI use. Revelio adds that about three quarters of senior bank roles are filled from within, so fewer juniors today means fewer people building the judgement senior roles rely on. For agent pilots whose risk case rests on a person checking each step, Anthropic's figures put a size on the reviewer time that checking consumes.

Revelio Labs: Banks Are Losing Interest in Junior Talent in the AI Era | Anthropic: How Employment is Changing at Anthropic

Security

Singapore publishes a kit for testing AI agents that checks what an agent did, not only the answer it gave. Authority

Singapore's Infocomm Media Development Authority (IMDA) and the AI Verify Foundation published version 1.0 of a Starter Kit for Testing Agentic Systems for Safety and Reliability. IMDA's chief executive announced it on 14 October, and Jakub Szarmach, an AI governance practitioner, shared it the next day. It is the testing guide under the agent governance framework this brief covered on 1 October. The kit sets four steps: identify the risks, build realistic tests, run them and analyse the logs, then monitor and test again. Its measures compare an agent's own account of its work with the tool calls it actually made, counting actions it invented and actions it left out. It also tracks how often people approve an agent's requests, and warns that a rate close to 100% may mean rubber-stamping. Agents may behave differently when they spot signs of a test, such as dummy names or placeholder emails. Testers are told to prove that a sandbox, the closed test environment, holds rather than assume it. Firms that buy agents inside software services may see too little of the agent to test it, and the kit tells them to agree that access with the provider. DBS, OCBC and UOB contributed. In one case study KBTG, the technology arm of Thailand's KASIKORNBANK, tested a chain of three software-delivery agents by planting defects to see who caught them. Jakub Szarmach's advice for small testing budgets is to "Verify what actually happened against tool calls and system changes, rather than relying on the agent’s completion message." The kit is voluntary and binds no EU bank. Readers had on 15 October the point that an agent can reach the right result by the wrong path. The kit turns it into defined measures, built with three banks, and names log access for bought agents as something the buyer has to secure from the vendor. That condition applies to every agent feature now arriving inside software that banks already license.

Infocomm Media Development Authority (shared by Jakub Szarmach)

IEEE publishes a recommended practice on AI governance addressed to an organisation's governing body. Institute

The IEEE Standards Association, the standards arm of the global engineering body IEEE, published IEEE 2863-2026 on 16 October. Its title is Recommended Practice for Organizational Governance of Artificial Intelligence. IEEE's standards board approved it on 4 June, and work on it began in February 2020. According to IEEE's summary, it sets out principles that an organisation's governing body should consider and prioritise, and processes for putting them into practice. Case studies on specific AI applications show the processes at work, including how to resolve tensions between principles. It applies to any organisation that develops or uses AI, and it is voluntary. A working group of IEEE's AI standards committee, chaired by Gary Marchant, wrote it. The full text was not available to this brief, so this account rests on IEEE's summary. The final FSB practices, reported here on 16 October, open with the board's role in aligning AI with strategy and risk appetite. A standards body now addresses the same audience, which gives auditors and supervisors a second published reference for board-level AI oversight.

IEEE Standards Association: IEEE 2863-2026