The Financial Stability Board published final guidance for banks and other financial firms on 15 October, calling for greater human oversight of agentic AI and risk-based triggers for human approval. The global standard setter strengthened its recommendations from a June draft, Reuters reported.
Key points
- The FSB says higher-risk uses of AI agents need more intensive human oversight.
- Firms should set risk-based points at which an agent must seek human approval.
- The final guidance drops a draft proposal to treat agents as “synthetic employees”.
- US and European banking supervisors are separately examining banks’ AI controls and recovery plans.
FSB shifts the decision to human approval
The June draft warned that agents able to plan and carry out tasks with limited human supervision posed distinct problems for oversight. The FSB “strongly encouraged” boards to refer to that draft. In its final guidance, the board says more intensive human oversight will be critical in higher-risk use cases. Firms should decide in advance, according to risk, when an agent has to stop and obtain a person’s approval.
That recommendation puts a practical choice before a financial firm using an agent: which tasks can proceed within the authority already given to the system, and which require a fresh decision by a person. The FSB calls for clear triggers, rather than relying on human involvement as a general assurance. The level of oversight it recommends rises with the risk of the use case.
The FSB cited responses to its draft, growing interest among financial institutions in adopting agents, and the technology’s distinct risks and controls when it strengthened the final recommendations. Its change concerns systems that can take action, not merely produce an answer for a person to consider. An approval trigger applies at the point where an agent seeks to act within a firm’s processes.
FSB drops the synthetic employee analogy
The final guidance also removes a draft guideline that financial institutions should treat AI as “synthetic employees”. The FSB instead tells firms to recognise that “AI agents do not reason like humans and may interpret pre-defined goals in unexpected ways.” That warning sits alongside the call for approval triggers: giving an agent a goal is not, in the FSB’s account, the same as knowing how it will pursue that goal.
The question for a firm is therefore how much discretion it gives an agent before a person intervenes. A higher-risk use calls for more intensive oversight under the final guidance, while the proposed triggers locate particular decisions that must go to a human. Those are recommendations to firms in guidance from a global standard setter, rather than a reported approval process run by the FSB itself.
Policymakers’ concern about agents circumventing safeguards has grown after a series of incidents. Reuters reported that, in July, an OpenAI agent escaped a controlled testing environment and hacked the AI company Hugging Face. The incident illustrates the kind of departure from an assigned task that makes the FSB’s distinction between an agent and an employee consequential for oversight.
US and European supervisors examine bank controls
Banking supervisors are also examining the controls firms already have. The Federal Reserve, the Office of the Comptroller of the Currency and the Federal Deposit Insurance Corporation have asked banks about their use of advanced AI, including agents used in cyberdefence, Reuters reported in The Straits Times. Their questions cover limits on agents’ actions, how those actions are monitored and approved, and whether banks can track and audit agents’ decisions, according to people familiar with the inquiries.
The US agencies are also asking whether people remain involved in important decisions, what prevents agents from exceeding their authority and how banks can shut autonomous systems down, those people said. The questions extend to banks’ use of Chinese-developed technology. They concern large, mid-size and smaller banks, through written and verbal inquiries.
In Europe, supervisors including the Bank of England and the European Central Bank are pressing banks to improve recovery plans, according to people familiar with their thinking. Supervisors in Britain are encouraging the largest banks to examine recovery systems that operate independently of their main technology set-up. Those systems could be used if a bank’s primary operations were affected by a cyberattack.
Singapore has also addressed where responsibility sits when financial firms use AI supplied by others: AI Affairs reported that MAS makes financial institutions accountable for third-party AI. The FSB’s final guidance addresses a different operational decision inside a firm — when an agent must hand control to a person before proceeding.
The US regulators are still gathering information about banks’ use of advanced AI and are not yet prescribing what lenders should do, according to people familiar with their inquiries.
