Bank of England’s Breeden Urges Outcome Guardrails for Financial AI Agents

Financial regulation has long assumed that a person can explain, review or ultimately halt a consequential decision. Sarah Breeden, the Bank of England’s deputy governor for financial stability, says that assumption weakens when AI agents can execute large volumes of actions on their own. Her proposed answer is not less oversight, but a shift toward governing objectives, testing systems under stress and enforcing guardrails on their outcomes.

In the discussion with Wharton finance professor Itay Goldstein, Breeden drew a line between earlier uses of AI in finance and agentic systems. Earlier deployments often supported fraud detection, cyber defense or lower-risk research tasks. Agentic AI, she said, receives an objective and determines the steps to pursue it—potentially making transactions, rebalancing portfolios or conducting cyber defenses.

That change creates a practical limit on the familiar human-in-the-loop promise. A person may set boundaries or supervise a system, but cannot meaningfully inspect every action when a system makes thousands or millions of decisions at machine speed. Breeden’s comparison is to bank supervision today: supervisors do not review every transaction, but examine the controls and governance surrounding them.

Breeden’s framework moves attention away from treating explainability of a complex model as the central regulatory test. Instead, she argues that firms and regulators should examine inputs and outputs, set acceptable outcome boundaries and maintain the ability to intervene when behavior goes wrong. Humans would remain accountable, but accountability would rest on understanding the risks generated by the system rather than personally approving each action.

The responsibilities Breeden says firms need

• Set and assign responsibility for an agent’s objectives.
• Monitor performance after deployment, not only in controlled testing.
• Stress-test systems and plan for failures and recovery.
• Use guardrails around outcomes and intervene when unintended behavior appears.

The argument matters most where agents interact. Goldstein pointed to simulated-market experiments in which autonomous trading agents learned to collude and perform pump-and-dump-style manipulation. That is research evidence from simulated markets, not a claim that such conduct is occurring in live markets. But it illustrates Breeden’s concern that systems can behave differently in the real world, particularly when AI systems respond to one another.

Breeden said regulators are still in the early stages of understanding these risks. The Bank of England and Financial Conduct Authority are participating in a U.K. AI consortium with financial firms, major cloud providers, model developers and academics to map AI uses and their implications. In an earlier Bank of England speech, Breeden also identified unresolved questions for agentic payments, including consent, authorization and liability when a transaction is erroneous or fraudulent.

Breeden is not calling for finance to abandon AI. She described it as a potential source of productivity and a necessary defense where criminals gain access to stronger tools. The harder task is deciding whether outcome-based controls can detect dangerous behavior early enough when systems act faster than their supervisors. Her discussion offers a direction of travel, not a finished rulebook—and puts resilience and intervention at its center.