The Financial Stability Board published 12 sound practices for financial institutions using artificial intelligence on October 15, 2026. The final report covers ordinary machine-learning systems, generative AI and agentic AI, but it is not a binding international standard. It is a governance toolkit for banks, insurers, asset managers and other financial firms deciding how much control a particular use case needs.
What the FSB finalised
The first four practices sit above individual models. Boards and senior management are encouraged to align AI adoption with business strategy and risk appetite, define accountability, bring AI risks into the wider risk-management framework and keep adapting as the technology changes.
The remaining eight follow the AI lifecycle:
- assess the materiality and risk of each use case at the start and later;
- select models and systems for the business objective and risk;
- keep training, testing and operating data fit for purpose;
- understand explainability limits and use compensating controls where needed;
- test and monitor performance;
- design human oversight around autonomy, complexity and risk;
- manage cyber and ICT risk; and
- manage third-party AI dependencies.
The report applies proportionality. A small, low-risk document-summary tool does not need the same controls as an autonomous system that can move money or make a credit decision. The institution still needs a documented way to decide which is which.
Why agentic AI gets special attention
The FSB says autonomous systems may take unauthorised or erroneous actions, expose data or disrupt connected systems. That matters in finance because an agent can cross the line from producing information to executing a trade, changing a limit or initiating a payment.
Human oversight therefore cannot be a person nominally “in the loop” after an action is irreversible. The report says oversight should reflect materiality, risk, autonomy, complexity and explainability. In practice, teams need to define which actions require approval, which limits are enforced by systems and how an agent can be stopped.
American Express’s new Intent Intelligence framework illustrates the same issue at checkout: an agent needs boundaries tied to the merchant, budget and approved purchase, not just access to a reusable credential.
Third-party concentration is part of the model risk
Many institutions do not train frontier models. They buy access to models, cloud services and orchestration tools. The FSB’s twelfth practice asks firms to manage the dependencies created by that stack.
Useful questions include:
- Can the provider change the model without notice?
- Which customer or transaction data reaches the provider?
- Can the institution reproduce a material decision later?
- What happens if the model, region or API is unavailable?
- Are several critical use cases dependent on one provider?
- Can the institution test updates before they reach production?
A contract does not replace technical controls. Exit plans, fallbacks and inventories need to be usable when the service fails or its risk changes.
What changed from consultation to final report
The FSB says the final report incorporates public feedback on the June consultation. The final package is non-prescriptive and adaptable to each institution’s size, complexity and role in the financial system. Its case studies are illustrative, not benchmarks or promised outcomes.
That distinction matters. The report does not certify any AI system and does not tell supervisors to accept a particular control. Local law, sector rules and existing model-risk obligations still apply.
What financial firms should do next
The useful first step is an inventory that includes more than model names. For each AI use case, record the owner, business purpose, decisions or actions, data, provider dependencies, users, autonomy, affected customers, testing, monitoring and stop controls.
Then separate three questions:
- Materiality: what could this system change or harm?
- Control: which technical and human limits prevent that outcome?
- Evidence: can the institution show what the system and reviewer did?
The FSB’s list is broad enough to serve as a gap check. It is not permission to treat a completed checklist as proof that a deployment is safe.
