The Financial Stability Board has finalised 12 sound practices for financial institutions adopting artificial intelligence. The framework joins board oversight, model and data controls, human intervention, cyber risk and vendor dependence in one lifecycle view. It is presented as a practical toolkit rather than a single binding global rule.[1]
Institutions are encouraged to integrate the practices into existing risk, cyber and third-party frameworks according to their size, complexity and risk profile. The toolkit has no single legal commencement date: applicable obligations continue to come from each institution’s jurisdiction and regulated activities.[1][4]
Four practices set the organisation-wide governance layer
Governance practices
| Practice | Control question | Evidence an institution would need |
|---|---|---|
| Board and senior management | Does AI adoption align with strategy, business model and risk appetite? | Approved risk boundaries, escalation criteria and reporting that distinguish experimentation from material production use.Source 1 |
| Roles and responsibilities | Who owns the use case, model, data, control and final decision? | Named accountability across business, technology, risk, compliance, security and internal audit.Source 1 |
| Risk framework and documentation | Are AI-specific risks integrated into ordinary enterprise controls? | An inventory of use cases, decisions, versions, dependencies, limitations and incidents.Source 1 |
| Continuous learning | Can governance adapt as models and threats change? | Periodic reviews, updated training, incident lessons and revision of controls after material change.Source 1 |
The governance layer prevents AI from becoming a collection of disconnected vendor tools. A board does not need to approve every low-risk model output, but it needs a credible way to know which uses are material, who can intervene and whether the aggregate exposure remains within the institution’s risk appetite.[1][4]
Eight practices follow the AI lifecycle
Lifecycle practices
| Stage or control | FSB focus | Operational implication |
|---|---|---|
| Materiality and risk assessment | Assess the use case at inception and thereafter. | A model can become material when its scale, autonomy, customer impact or market exposure changes.Source 1 |
| Technology selection | Choose an appropriate model and architecture. | The most capable model is not automatically the right model when explainability, latency, privacy or control requirements differ.Source 1 |
| Data governance | Control quality, lineage, access and lawful use. | Training and live input data need ownership, monitoring and handling rules.Source 1 |
| Explainability and compensating controls | Match explanation to the model and decision context. | Where a full technical explanation is impractical, stronger testing, constraints or human review may be needed.Source 1 |
| Performance evaluation | Measure whether the use case continues to work as intended. | Accuracy alone may be insufficient; stability, bias, drift, false positives and business impact can matter.Source 1 |
| Human oversight | Ensure effective intervention and accountability. | A nominal human sign-off is weak if the reviewer lacks time, information or authority to challenge the output.Source 1 |
| Cyber and ICT risk | Protect models, data and connected systems. | Prompt injection, model theft, poisoned data and automated attack scale join ordinary system-resilience risks.Source 1 |
| Third-party dependencies | Manage concentration and outsourced AI services. | The institution needs visibility into provider changes, outages, subcontractors and exit options.Source 1 |
Proportionality follows the authority and impact of the use case
A low-risk internal drafting or search tool can warrant lighter controls than a system making credit, suitability, pricing, trading, fraud or claims decisions. An autonomous workflow able to act across tools or accounts needs limits on its permissions, usable logs and an effective way to intervene. Smaller institutions can adopt simpler arrangements where appropriate, while still identifying who owns each material use and which decisions or data it influences.[1][4][3]
Third-party AI is treated as a financial-stability dependency
Vendor questions behind practice twelve
| Dependency | Risk | Control response |
|---|---|---|
| Shared model provider | Many institutions may rely on the same service or model family. | Assess concentration and common-mode failure rather than only bilateral contract performance.Source 1 |
| Model update | Provider changes can alter outputs without a traditional internal release. | Require notification, testing and rollback or containment arrangements for material changes.Source 1 |
| Subcontractors and data services | The visible provider may depend on other infrastructure. | Map critical fourth parties, data feeds, cloud services and failure points.Source 1 |
| Exit and portability | A service failure can become an operational lock-in. | Maintain data, documentation and alternative processes sufficient to reduce or replace the dependency.Source 1 |
Consultation feedback shaped the final practices
The FSB received around 160 responses to the June consultation, spanning regions and sectors. The response overview records changes made to the final report rather than presenting the consultation draft as unchanged. The FSB also says it will monitor developments and update the practices as AI technologies and applications evolve.[2][1]
Sources & further reading
- FSB finalises sound practices for the responsible adoption of artificial intelligence Financial Stability Board · 15 Oct 2026 · Accessed 16 Oct 2026
- Sound Practices for Responsible Adoption of Artificial Intelligence: overview of consultation responses Financial Stability Board · 15 Oct 2026 · Accessed 16 Oct 2026
- Sound Practices for Responsible Adoption of Artificial Intelligence: consultation report Financial Stability Board · 10 Jun 2026 · Accessed 16 Oct 2026
- Opening remarks on sound practices for artificial intelligence Financial Stability Board — Michelle W. Bowman · 7 Jul 2026 · Accessed 16 Oct 2026