FSB AI Sound Practices: A Guide for Financial Firms

On 15 October 2026, the Financial Stability Board (FSB) published 12 “sound practices” to guide responsible AI adoption by financial institutions. Four address organisation-wide governance; eight cover risks across the AI lifecycle, from use-case assessment and data governance to human oversight, cyber resilience and third-party dependencies. The FSB is explicit: these practices are non-prescriptive and are not a binding international standard. Institutions should apply them proportionately while continuing to meet local legal and regulatory requirements.

What the FSB published—and what it did not

The final report groups its 12 practices into two connected layers: board-level governance and lifecycle risk management. It followed a public consultation launched in June 2026. The final version clarifies board and senior-management responsibilities, strengthens guidance on documentation and higher-risk use cases, and expands its treatment of third-party AI and agentic systems (FSB final report; FSB announcement).

That distinction matters. The practices are a supervisory reference and practical toolkit, not a new statute, certification or universal control checklist. The FSB says institutions should adapt the practices to their business model, size, complexity, interconnectedness, risk profile and the materiality of each AI use case. Existing national and regional law still applies.

The 12 practices in plain English

Scope	What the practices ask institutions to address
Organisation-wide governance (1–4)	Board and senior-management strategy and oversight; clear accountability; AI risks within existing risk frameworks and documented use; continuous learning and adaptation.
AI lifecycle (5–12)	Materiality and risk assessment; technology selection; data governance; explainability and transparency; performance monitoring; human oversight; cyber and ICT risk; third-party AI dependencies.

The list is broad by design: a customer-facing credit decision system, an internal document assistant and an agent that can trigger transactions do not have the same impact or authority. The FSB therefore encourages proportionate controls rather than treating every model or workflow as equally risky.

How to turn the framework into operating controls

The following is SignorCrypto’s implementation synthesis of the FSB practices—not a prescribed FSB sequence:

Inventory the actual use. Record the business purpose, affected users, model or provider, data sources, connected tools, human decision points and whether the system can take actions.
Assess materiality and risk before deployment. Consider the consequences of an error, the scale of use, autonomy, critical functions, data sensitivity and dependencies. Reassess when the use case or system changes.
Assign named owners. Make clear who approves use, monitors performance, handles exceptions, accepts residual risk and can stop or roll back a system.
Set lifecycle gates. Define evidence for selection, testing, release, monitoring, material change and retirement. Keep versioned documentation, incident records and review decisions.
Control suppliers and concentration. Map external models, cloud and data services; understand fallback, continuity, audit and exit arrangements; and monitor dependencies that could affect multiple critical workflows.

For a practical evaluation plan that combines model testing, red teaming and user testing, see our NIST ARIA guide. For EU financial firms, these governance practices do not replace applicable law; our EU AI Act high-risk classification guide explains one separate legal test.

Agentic AI raises the stakes—but the guidance is still developing

The FSB identifies risks from highly autonomous systems, including unintended actions, data incidents and disruption to connected systems. It adds considerations for agentic and multi-agent AI, but also says more specific practices for those systems remain nascent and may need fuller treatment as understanding develops. The report should therefore not be read as a complete agent-security standard.

For agents, operational controls should apply to the actions and permissions around the model—not only to the text it generates. Our AI agent security case study examines why isolation, scoped access, monitoring and meaningful human approval matter in connected environments.

What financial firms should take away

Verified fact: the FSB’s final report, published on 15 October 2026, sets out 12 non-binding practices for financial institutions, organised around governance and lifecycle risk management. It encourages a proportionate approach and does not override local law.

SignorCrypto analysis: its most useful contribution is the connection between board accountability and operational controls. Model evaluation alone is insufficient if nobody owns the workflow, supplier exposure, human escalation or response to a material change.

Open point: the FSB says its practices will evolve with AI technologies; it specifically notes that guidance for agentic and multi-agent systems remains at an early stage. The rules applicable to a particular firm still depend on its jurisdiction and activity.

The OECD’s January 2026 paper separately highlights the challenge of how financial supervisors interpret and implement existing AI rules (OECD, Supervision of Artificial Intelligence in Finance). In a September 2026 speech, Bank of France First Deputy Governor Denis Beau also discussed customer, cyber and financial-stability risks associated with AI in finance (speech hosted by the BIS). These sources reinforce why governance frameworks must be applied alongside, not instead of, sector-specific requirements.

Frequently asked questions

Are the FSB AI sound practices legally binding?

No. The FSB describes them as non-prescriptive and not a binding international standard. Financial institutions remain responsible for applicable national and regional legal obligations.

Does the FSB recommend the same controls for every AI system?

No. It recommends proportionality based on the institution’s context and the use case’s materiality and risk. A system affecting a critical function may justify more robust practices than a low-impact internal tool.

Does the report settle how to govern AI agents?

No. It includes agentic-AI considerations, but the FSB says more specific practices for agentic and multi-agent systems remain nascent and may need further development.

Sources

For teams adapting AI governance into everyday employee assistants and controlled business workflows, explore Botchi.