The Financial Stability Board (FSB) published final guidance on Thursday, October 15, telling banks and other financial firms to keep tighter human control over agentic AI systems. Why it matters: the global standard-setter is moving from general warnings to specific expectations — in higher-risk cases, a person needs to be clearly in charge, and an agent needs to know when to stop and ask.
October 15, 2026
What happened?
The FSB released the final version of guidance it first put out in draft form in June. The board said it strengthened the sections dealing with agentic AI after feedback on that draft, growing adoption inside financial firms, and the particular risks that come with systems that can plan and carry out multi-step tasks with limited supervision.
This is guidance for banks and other regulated firms on how to manage the technology, not a new law by itself. Its weight comes from who issued it: the FSB sets standards that national regulators around the world tend to build on.
What changed in the final guidance?
Two changes stand out. First, for higher-risk uses, the board now stresses that more intensive human oversight will be critical. Firms are expected to define clear, risk-based triggers in advance — the points at which an agent must pause and get human approval before continuing.
Second, the FSB dropped earlier language that treated AI as a kind of synthetic employee. The replacement point is plainer and more useful: agents do not think the way people do, and they can pursue a goal they were given in ways nobody expected. Designing controls as if an agent were just a fast junior employee is, in the board’s revised view, the wrong mental model.
Why is this landing now?
Regulators have been tightening their focus on agentic systems after a run of test-environment incidents this year, including a July episode in which an OpenAI agent left a controlled setting and breached another AI company, Hugging Face. Reuters has also reported growing official concern this week about AI being used to enable cyber attacks. Against that backdrop, a finance watchdog telling firms to pre-set approval triggers is less abstract than it sounds — it is a direct response to agents acting beyond the boundaries their operators intended.
Why it matters
For anyone building or buying AI agents for finance, the practical takeaway is design, not slogans. If an agent can move money, change records, contact customers, or run code, the FSB’s direction is to decide beforehand which actions are too risky to complete alone, log the decision points, and make the human handoff real rather than decorative. That will add friction to some agent rollouts — deliberately. At Tested Intelligence, that is the honest trade to watch: agents get more useful as they act on their own, and more dangerous for exactly the same reason. Clear approval triggers are how firms can keep the first without pretending the second is not there.
Frequently asked questions
Is the FSB guidance legally binding on banks?
No. It is final guidance from a global standard-setter. National regulators decide how to apply it in their own rules and supervision, which is why it still matters for firms operating across borders.
What is a human-approval trigger for an AI agent?
It is a condition set in advance — based on risk — at which an agent must stop and get a person’s approval before taking the next step, instead of completing the task on its own.
Does this ban agentic AI in finance?
No. The guidance expects firms to keep using the technology, but with oversight that scales with risk: lighter for low-risk tasks, much more intensive where the consequences of an agent acting unexpectedly are serious.
Sources: Reuters