Bank of England's Bailey says AI agent concerns 'have got to be addressed'

The Bank of England's governor, Andrew Bailey, told an audience at the Group of Thirty that concerns around AI agents have got to be addressed, the Financial Times reported on October 17. The remark moves AI agent risk from technical debate onto the agenda of central banks and supervisors.

The Group of Thirty is a consultative body of central bankers, regulators and academics. Its discussions carry no binding force, but they often preview where policy consensus is heading. Bailey raising AI agents there suggests the question has shifted from whether models make mistakes to who is accountable when software acts on a person's behalf.

The difference between an agent and a conventional AI tool is autonomy. A chatbot offers a suggestion that a human confirms with a click; an agent is authorised to carry out a sequence of steps, checking an account, initiating a payment, calling an external interface, retrying after a failure. The longer the chain, the harder it becomes for any single checkpoint to cover every action.

That creates two problems for financial systems. The first is explainability and audit: authorities need to reconstruct the path of an automated decision. The second is liability: when an agent does something a supervisor considers unacceptable within its authorised scope, it is not obvious whether the user, the developer or the deploying institution carries the responsibility.

From what is public so far, the remarks look more like agenda-setting than rule-making. Bailey establishes that the issue must be dealt with, while leaving concrete requirements and timetables to supervisors and legislators.

The likely path is incremental. AI agents will enter financial workflows first as assistants, then gradually receive wider authority. The supervisory difficulty is that rules need to be in place before that authority expands, while technology cycles move faster than legislation.

The next signals to watch are whether the Bank of England turns the remark into specific requirements, and whether the Group of Thirty discussion produces a shared international line. For institutions already deploying agents, the window for compliance design may be shorter than it appears.