Edition 9 · 17 November 2026

AI Risk Reading

UK regulators tell ministers that some AI agent providers are not engaging, a Bank of England policymaker explains why delegating to an agent needs limits, the Fed finds agents already trying to act in bank customers’ accounts and puts a size on the AI debt, one founder designs a bank for agents, and a dormant software defect takes Microsoft 365 down for a day.

DRCF joint letter on personal AI agents

What is new

Two and a half weeks after their joint statement (Edition 7’s essential read), the chief executives of the CMA, FCA, ICO and Ofcom have written to the Chancellor, two Secretaries of State and the Minister for AI ahead of “the forthcoming launch of a number of personal agentic products in the UK”. In October they “strongly encourage[d] early engagement”. Now: “We expect firms to engage early and openly with the relevant regulators before bringing their products to market”, including on whether their activities need authorisation or registration. And they report a gap: “Some firms have engaged early and constructively, but engagement has not been consistently early or open across the relevant regulators.” Agents must not “compromise the operational resilience of the wider systems with which they interact.” Banks and telecoms firms working with agent providers “should also engage with us early. Their own obligations continue to apply when customers use personal AI agents.” The FCA’s Head of Innovation, Colin Payne, called it “essential reading” for anyone “building, deploying or partnering with AI agents in the UK”.

Why a board should care

The regulators have gone over the heads of the firms to ministers, and called out the problem. For a bank, the last line matters most: a customer’s agent does not carry the bank’s obligations away with it. The letter makes no new rules, but it ends as the statement did: “We will not hesitate to use our powers.”

Verdict: read the letter in full; it is two pages. Read it next to October’s statement to see how far the tone has moved.

Principals and Agents

What is new

In his final speech as an external member of the Bank of England’s Financial Policy Committee, Jonathan Hall treats an AI agent as any delegate: a mandate is always incomplete, so the principal also needs limits on what the delegate can do, sight of what it does, and the means to step in. Training a model to follow rules is not enough, because “alignment trained through text does not necessarily transfer to agentic behaviour”; it needs access controls and monitoring of actions, at scale. He is candid that “not all of the safety measures I will outline are being used in production in financial firms today.” Beyond the single firm, he names three system dynamics: concentration of providers, correlation between them, and correlation in stress. Whether AI will worsen them: “at this point we don’t know, but there is a risk that it will.” These are his views, not FPC policy, and he is not calling for new rules.

Why a board should care

It turns risk appetite into a description of what an agent may actually do, and it puts cost in the frame. Firms “are responsible for the agent’s actions. If the overall costs are too great, then perhaps delegation to an agent is not appropriate.” Not every task should be delegated just because it can be.

Verdict: read “Delegation and the principal-agent problem” and “Case Study II: An AI agent”.

Some Potential Financial Stability Implications of Artificial Intelligence

What is new

Michael Barr, a member of the Federal Reserve Board, reports that “major banks now report finding AI agents attempting to take actions in retail customer accounts”, where they cannot verify “authenticity, consent, and alignment with customer instructions”. He wants “explicit consent verification for specific actions”, and warns that agents “may accelerate deposit outflows during stress periods”, with runs that could “unfold in minutes”. The Fed’s Financial Stability Report of 13 November puts a size on the financing behind AI. Debt issuance tied to advanced computing passed $680bn by October, about five times 2023’s $130bn, and the report names the “circular financing” between investors and their customers. Leases not yet started exceed $1tn and other obligations, such as take-or-pay contracts and guarantees, exceed $2tn, though the report stresses these are not discounted and not comparable with ordinary debt. The largest US banks have “close to $1 trillion” of exposure to large tech firms, data-centre developers, private credit and utilities.

Why a board should care

Two points. Logging in tells you who is connecting, not whether the customer authorised this payment, to this payee, for this amount. And a technology dependency and a credit exposure can lead back to the same AI build-out. The latest results show the shape. CoreWeave, a heavily borrowed AI cloud provider, beat forecasts last week and its shares jumped about 24%. But its quarterly interest bill of $846m was four times its operating profit, and its chief executive said: “We will not use the unsecured debt market to fund our business at current yields, nor do we need to.” Arm’s founding chief executive, Sir Robin Saxby, expects “a major market correction by the end of next year”, citing “a lot of people in the same deals”. A prediction, recorded in June, and others disagree; but the Fed has now measured what he describes.

Verdict: read Barr on retail customer agents and deposit outflows; read Box 2.2, “Leverage in the Data Center Buildout”, for the figures. Barr’s views are his own.

Why I’m building my second bank in a decade

What is new

Anne Boden, who founded Starling, is funding a new bank, Lockwood, “built by agents, to serve customers that want to use agents, and then be defended by agents”. Customers will set limits for each agent: her example is a shopping assistant allowed groceries only, up to £100 a month, so a £200 fuel payment is refused. It will offer a current account only, earning from interest on deposits, and will seek an e-money licence first and a banking licence later. She wants to convene banks and regulators on how agentic banking should work. Reported by Jill Treanor.

Why a board should care

The big UK banks are restricting customers’ agents while liability is unclear (Edition 8). Boden is building the customer’s mandate into the account, which is the explicit consent Barr asks for. The Governor of the Bank of England put the underlying problem plainly last month: “You can create what I would call economic agents, you can’t create legal agents.” Someone still answers for what the agent does. Lockwood is an ambition, not yet a licensed bank, and “defended by agents” is untested.

Verdict: read it for the design idea, not the forecasts.

Microsoft 365 outage, 12–13 November

What is new

Outlook, Exchange, Teams, OneDrive, Copilot and some of Microsoft’s security tools were disrupted for about 24 hours from the afternoon of 12 November. Microsoft’s preliminary post-incident report of 15 November, as reported: a defect in a Windows update installed in July had lain dormant for over three months on the servers that route Microsoft 365 traffic. A routine switch of Teams traffic onto the same servers, with some capacity already out for an unrelated reason, tipped them over: within two hours more than 80% had stopped responding, and the rest were swamped. Microsoft says it was not an attack. Its final report is due by 20 November.

Why a board should care

One provider, one dormant defect, and email, meetings, files, AI assistants and security tools went down together; for many firms, the fallback for talking to each other ran on the same platform. The AI Blast Zone traces where a stumble like this lands in your own organisation.

Verdict: wait for Microsoft’s final report. Meanwhile ask: if Microsoft 365 were down for a working day, how would we communicate, run our security tools and meet our deadlines, and have we tried it?

Organizing AI at Enterprise Scale

A working paper from Harvard Business School and Keystone. Interviews at 202 large US firms already using AI found them strongest on strategy and oversight (64%) and weakest on monitoring and evaluation (39%). Only 4% of the lowest-scoring quarter test their AI regularly in live use, against 63% of the highest. Boards see the governance on paper; the gap is in watching agents once they run. Self-reported, and part-funded by Microsoft.

Anthropic’s listing

Anthropic’s amended prospectus offers shares at $70–78, and trading on Nasdaq is reported for Thursday 19 November. I will cover it in Edition 10 once the price is set.

Disclosure: I use Claude, Anthropic’s model, in my research and drafting.

This edition’s question

Before we give an AI agent more authority, what evidence shows that it will stay within its mandate in live use, and which decisions are we prepared to keep with people if that evidence is weak?

Previous edition All editions All writing Subscribe on LinkedIn Subscribe by email Subscribe by RSS