Rate Brief ENDE

BoE's Jon Hall in his last FPC speech: AI agents need access limits, monitoring and identity

Jonathan Hall used his final speech as an external member of the Bank of England's Financial Policy Committee to look at a risk he expects the committee to face: AI agents acting in finance. He framed the issue as a principal-agent problem, the same one the Bank itself embodies as a body that holds authority delegated by Parliament.

His starting point is that misalignment between a principal and an agent is not new. Citing Jensen and Meckling's 1976 paper, he set out the three forms of agency cost: the principal's monitoring, the agent's bonding and the residual loss when the agent still diverges. The Bank, he argued, shows how constraints, a narrow mandate and accountability to Parliament keep a powerful delegate in line.

For AI agents he listed safeguards that he said are not all in production at financial firms today. An agent is a model wrapped in a harness that sets its tools, permissions, policies and approvals. For high-stakes use it should also sit in a sandbox built on least privilege, though he noted sandboxes are not fully watertight. To act in markets, an agent must be credentialised with a unique identity; he pointed out that UK rules already require a distinct designation for each trading algorithm in transaction reports, which allows an audit trail and the revocation of access. Monitoring should cover both the model's visible reasoning and its actions, and firms will need scalable, risk-based oversight to avoid alert fatigue.

The financial stability question, in his telling, is about the system rather than single firms. Markets fail to provide liquidity when enough providers withdraw or start demanding it at the same time, so the risk depends on concentration and correlation as well as individual riskiness. Whether delegating trading to AI agents raises those is, he said, not known yet, which argues for experiments, theory and monitoring. He described a planned simulation pairing rules-based portfolio managers with LLM-based agents prompted to copy their style.

Hall noted that a Bank and FCA survey found 80% of respondents using some agentic AI but only 2% of those agents with broad autonomy.

BoE's Jon Hall in his last FPC speech: AI agents need access limits, monitoring and identity
BoE's Jon Hall in his last FPC speech: AI agents need access limits, monitoring and identity — Rate Brief

Why it matters

The speech sets out what the FPC is likely to look for: identity and revocable access for every agent, and evidence on whether AI-driven strategies move together under stress.

Written by Victoria Shinder.