Regulation
The Financial Stability Board wants 12 sound practices guiding AI in finance by October, and its consultation window just closed without settling who is accountable when an agentic system goes wrong
The FSB's consultation on responsible AI adoption ran to 22 July 2026, covering board-level governance through model deployment across G20 financial institutions, with a final report due in October and no binding enforcement mechanism attached.
The structure of the FSB's proposal is worth understanding before judging it. Practices one through four are about organisation-wide governance, the board and senior management setting risk appetite and owning the overall approach to AI adoption. Practices five through ten get more specific, covering how individual AI use cases should be managed at each stage of their lifecycle, from development through deployment, with proportionate guardrails rather than a single rulebook applied uniformly. That is a sensible design for a body that has to write guidance covering everything from a chatbot answering customer queries to a model making trading decisions, but it also means the practical bite of any individual practice depends entirely on how national regulators choose to implement it.
What stands out is the explicit attention to generative and agentic AI as a distinct category, not just an extension of the model-risk frameworks banks already run for traditional machine learning. Explainability and human oversight get their own emphasis, which reads as a direct response to the past year of agentic trading and advisory tools shipping into live brokerage accounts faster than anyone has built a shared vocabulary for auditing what those agents actually do. The FSB is trying to write the sound-practices layer before the incident that would otherwise force it to.
The catch is that this is soft law. A consultation report with an October final publication date carries no enforcement teeth of its own, it exists to be picked up piecemeal by national regulators who are already moving at different speeds and with different priorities. The Reserve Bank of India has its own draft model-risk guidance in parallel, the FCA has been running supervised sandbox cohorts rather than blanket rules, and the EU AI Act is only now moving from planning document to active enforcement. The FSB's 12 practices are more likely to become a reference point regulators cite selectively than a uniform standard anyone is bound to.
The real test comes in October, when the final report lands and the consultation feedback either sharpens the practices into something firms can be held to or softens them further into principles nobody can fail. Until then, the gap that matters is the one between a financial institution that can point to a sound-practices document and one that can actually demonstrate, if asked, why a specific agentic decision was made and who signed off on the risk. That gap is exactly where agentic trading tools already live with retail customers are currently operating.
Read the original: Financial Stability Board - The Financial Stability Board wants 12 sound practices guiding AI in finance by October, and its consultation window just closed without settling who is accountable when an agentic system goes wrong. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.