Leadership1 publisher2 min readPublished
The Bank of England sorted firms' AI into fraud defence and low-risk help for years
Sarah Breeden, the Bank of England's deputy governor for financial stability, has given a public account of a long and benign period of AI use inside London firms. She is now watching agentic systems.
The Board Room · Leadership desk
What happened
- Sarah Breeden, the Bank of England's deputy governor for financial stability, discussed AI regulation with Wharton finance professor Itay Goldstein on the Future of Finance podcast series.
- Breeden said the Bank has been examining AI with financial services firms in London for five, six, seven years, through a public-private partnership set up to talk through how they were using it.
- She said the use cases fell into two categories: cyber and fraud work, where firms needed the best models, and relatively low-risk tasks such as helping people do research.
- Goldstein said the two had already discussed regulating AI in finance together on a panel at the European Central Bank Forum in Sintra, Portugal, in June.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision The artifact a firm will be able to produce is internal: a written record of which decisions an agent may take without a person, and who approved each boundary. A vendor's assurance about the model is not something a supervisor has endorsed.
- constraint The published account states the questions and leaves out the answers. A firm setting agentic policy this quarter absorbs the interpretation risk, and guidance written later may not fit a design already in production.
- precedent A standing dialogue with firms has been the Bank's instrument on AI for years, so the first expectations on agents are likelier to take shape there than in a rulebook. For large firms, a seat in that forum is a governance matter.
Breeden's summary of the years before agents is a benign one. "[T]here hadn't been anything that was flashing particularly red about how it was being used by the firms," she said [7]. In the defensive uses, cyber and fraud, she said firms needed the latest technologies and the best models "to be able to make sure you were doing your bit against the bad guys" [8]. Both of her categories classify risk by task [6], and software that takes an action on a firm's behalf falls into neither one.
Wharton's summary of the episode lists the evolving role of human oversight, the tools regulators can use to mitigate risk, and what the ultimate measure of regulatory success will look like [4]. The published conversation does not say where accountability for an agent's decision sits internally, or whether a model vendor carries any of it [13]. Those are the points a firm most needs settled.
A firm has always owned the consequences of a tool it chose to run, but Breeden's classification complicates that. It assigned low risk by the task being performed, and an agent changes the task while leaving the model where it was [6].
The page also points readers to "AI-Powered Trading, Algorithmic Collusion, and Price Efficiency" by Winston Wei Dou, Itay Goldstein and Yan Ji [10]. Collusion is a market-level problem, and Breeden's remit is market-level: financial stability, the supervision of financial market infrastructure, international issues, payment innovation and fintech [3].
None of this changes a permission this week. The venue is an academic podcast, in a third-season series Goldstein co-directs with Patrick Harker, the former Wharton dean and former president of the Philadelphia Federal Reserve [2][11]. Wharton's framing of the episode says AI is quickly becoming more autonomous, creating opportunities for financial innovation while raising questions about oversight and systemic risk [12].
What to watch
- Whether the first Bank of England expectation on agentic AI arrives through the public-private partnership Breeden described or through a formal supervisory statement.
- Any Breeden speech or Financial Stability Report passage that names where accountability for an agent's action sits inside a firm.
- Whether the Dou, Goldstein and Ji work on algorithmic collusion gets cited in a UK supervisory document.