The Bank of England’s artificial intelligence consortium held its second quarterly meeting, and its first in-person session, in October 2025 to deepen public and private sector dialogue on the capabilities, deployment and risks of artificial intelligence in UK financial services. Co-chair Sarah Breeden opened the meeting, confirming that Sarah Pritchard’s co-chair role would pass to David Geale following her move to deputy CEO of the Financial Conduct Authority, and reminding participants that discussions would follow the Chatham House Rule. Breeden outlined four workshop streams launched in May 2025 to structure the consortium’s work: concentration risk in artificial intelligence infrastructure providers, the evolution of artificial intelligence edge cases, explainability and transparency in generative artificial intelligence, and artificial intelligence accelerated contagion in financial markets.
The first workshop on concentration risk examined how increasing reliance on artificial intelligence within UK financial services is creating risks arising not only from the technology but also from the structure of the wider ecosystem. Participants identified five linked areas of concern, including concentration in third party artificial intelligence providers, contagion from model updates, capacity and scalability constraints, the need for third party assurance and minimum standards, and talent concentration and capability gaps. Members highlighted growing dependence on a small number of providers and discussed how the emergence of artificial intelligence agents, which typically rely on a few dominant models and infrastructure, could deepen concentration risk. The second workshop, focused on edge cases, looked at high value, complex applications of advanced artificial intelligence where novel risks emerge. Members noted that scenarios seen as edge cases today may evolve into business as usual, and warned that pressure to show returns on artificial intelligence investment could compress development timelines, reinforcing the need to balance innovation with safety and robust oversight.
The third workshop addressed explainability and transparency in generative artificial intelligence, recognising that terms such as explainability, interpretability and transparency are often used interchangeably and that inconsistent definitions may hinder effective risk management. To progress, the group adopted an outcomes focused definition and planned to review domestic and international guidance to identify key characteristics of explainability and transparency suitable for financial services use. Some members cautioned against spending too much time on terminology and argued that firms should instead prioritise building inherently explainable models, which they said had not always been standard practice. The fourth workshop on artificial intelligence accelerated contagion explored how automation and interconnected decision making could amplify shocks across the financial system, highlighting three drivers of contagion risk: common vendors and data leading to synchronised market moves, operational resilience challenges from dependency on a few critical providers, and model concentration and homogeneity that could create correlated errors. Participants illustrated these concerns with examples of multiple firms using the same artificial intelligence coding tools, and questioned how more autonomous agentic artificial intelligence might accelerate the spread of flawed updates across systems.
In a broader discussion of technical trends, members considered how emerging standards such as the model context protocol and the rise of small language models are shaping real world applications. They observed that new protocols could expand access to historical and proprietary data but might also introduce data quality issues and, over time, scarcity of suitable training data, with synthetic data suggested as one way to address gaps. The consortium examined the risks of model drift in rapidly evolving agentic artificial intelligence systems and the difficulty of locating drift within complex multimodal chains, concluding that governance and oversight frameworks must evolve in step with such systems. Members compared small language models and large language models, noting potential benefits of small language models for data privacy, guardrail implementation and evaluation, while acknowledging their more limited flexibility and use case specific nature relative to open source frontier models. The meeting closed with concerns about overreliance on large language models for summarising documents without human accountability, and with recognition that extensive artificial intelligence testing and evaluation work underway in the UK will be crucial for scaling proof of concept models. The next consortium meeting is expected to be held virtually in February 2026.
