The AI you did not buy: a blind spot in the FSB AI practices

The Financial Stability Board's (FSB) new sound practices for artificial intelligence (AI) in finance all rest on one assumption: that the institution knows where AI is running ¹. Most inventories are built from project approvals, so they miss AI that arrives later inside tools the institution already bought. We think this is the practical weak point of the framework. A board that asks "which AI do we use?" should ask a second question: "how would we learn that a supplier switched AI on?"

What the FSB published on 15 October

On 15 October 2026 the FSB published the final version of Sound Practices for Responsible Adoption of Artificial Intelligence, after a consultation that began in June 2026 ¹². It sets out 12 practices in two groups. Four cover organization-wide governance, from board oversight to roles, risk frameworks and adaptability. Eight cover the AI lifecycle, from materiality assessment and data governance to human oversight, cyber and information and communication technology (ICT) risk, and third-party AI risk ².

The practices are non-binding and meant to be tailored. The FSB nevertheless "strongly encourages" boards and senior management to use them, and asks for proportionate application ². Several points matter for security leaders. Materiality and risk of an AI use case are treated as related but different, and both set the level of oversight. Human oversight must be meaningful, meaning that people have the ability, authority and incentive to intervene, not a tick-the-box review. The cyber practice asks institutions to include AI scenarios in tests and exercises, and to check whether patching can keep pace with faster, AI-driven exploitation ¹.

Three AI estates, three ways to miss one

The practices work only on AI that is known. In our reading, an institution's AI falls into three estates, and each is found differently. This is our model, not the FSB's.

The FSB speaks to the third row. Institutions should identify AI components embedded in third-party tools "not explicitly marketed or sold as AI-enabled", and should consider material dependencies in concentration-risk management and contingency planning ¹. The report adds that AI functionality may be introduced after a tool was contracted, and that contracts should require notice of material changes affecting AI behavior ¹.

The reason this row matters is structural. Built and bought AI pass a gate: someone approves a project or a purchase. Embedded AI passes no gate, because the product already has a contract and a risk rating. The assessment made at onboarding describes the product as it was, not as it is after the next release. The same logic applies to human oversight: the FSB asks institutions to consider how much oversight a provider applies when it deploys model updates and new versions ¹.

Three tests show whether an institution has closed this gap:

• Discovery: does the inventory have a source that does not depend on a project team reporting?
• Trigger: does a supplier's change to AI behavior force a renewed risk assessment, or only a release note?
• Ownership: can the institution name the person who answers for AI in its core tools?

Our assessment: a sensible start is to answer these three tests for the ten suppliers behind the most critical processes.

What it means for Switzerland

The FSB's sound practices are not Swiss law, but they are a new international reference point for how supervisors may frame their questions. FINMA's Guidance 08/2024 of 18 December 2024 already asks supervised institutions to identify, assess, manage and monitor AI risks, and lists growing dependence on third parties among them ³. Its 2025 survey of about 400 institutions found that about 50% use AI or have first applications in development, that a further 25% plan to within three years, and that 91% of AI users also use generative AI ⁴.

The survey also reports that smaller institutions often rely solely on externally developed applications, and that reliance on BigTech providers is increasing as generative AI spreads ⁴. FINMA applies the principle "same business, same risks, same rules" and asks institutions that plan AI in critical processes to contact it early ⁴. The survey press release does not quantify embedded AI, so we cannot say how large this estate is in Switzerland. For smaller firms, the practical consequence is that the supplier relationship is the AI governance.

Beyond finance, any Swiss organization that buys software-as-a-service faces the same question. The FSB report is useful as a checklist of what a supervisor may ask, even where it does not apply.

Questions for leadership

1. Which of the three estates does our AI inventory actually cover, and where does its information come from?
2. Do we learn about AI features in existing tools from the supplier, or from our users?
3. Who is accountable for AI behavior in the tools our critical processes already depend on?
4. If a supplier changed a model that supports a critical process, what would trigger a new risk assessment, and how fast?
5. Would our human oversight of AI still count as meaningful if the volume of AI-assisted decisions doubled?