BoE publishes summary of AI roundtables

The BoE has published a summary of its three recent round tables with firms, at which it wanted to get a better understanding of the constraints firms are facing in adopting AI and what it can do do support responsible AI adoption. It held three meetings, with challenger and UK-focussed larger banks, globally systemically important banks, and insurers.

Key outcomes included:

• the firms generally supported the PRA’s framework of principles-based, outcomes-based policies with supervisory statements – believing this gives space to innovate within clear regulatory guardrails. They didn’t see the need either for more detailed guidance or rules, or for any additional type of BoE or PRA sandbox at the moment, considering the FCA initiatives to be enough;
• that second-line risk functions are still cautious about AI, which might delay deployment. Views were mixed on whether this was a good or a bad thing. Drivers could include bottlenecks in experience and a desire to show comprehensive regulatory compliance;
• that the traditional compliance emphasis on understanding the inner workings of a model is not tenable and can’t be effective in complex AI models – and agentic AI challenges the concept of HITL;
• it is hard for international firms to navigate the differing regulatory approaches, and this slows initiatives and increases cost;
• third party AI providers still fail to understand the compliance requirements of regulated firms, which slows down negotiations. While participants thought this would improve over time, some suggested that some form of minimum standard for third party AI providers to the financial sector could be useful. Others noted that it could get more difficult to substitute between providers as AI models become embedded in agentic systems;
• data protection laws continue to be a challenge, and new data location requirements could prevent scaling AI solutions across borders; and
• data quality can be a barrier, particularly in some areas of insurance, where insurers have relatively little data because they may for instance only have annual interaction with policyholders.