Designing the Governance Around the Queue
A supervisory LLM does not become accountable merely because a human reads its final screen. The question for the BIS-style screening use case is what institutional design surrounds the queue before, during, and after each comparison. This chapter uses the Financial Stability Board’s 2026 consultation framework to connect strategic direction, materiality, ownership, documentation, data lineage, and re-evaluation. The aim is not to turn guidance into a checklist, but to show why a ranked candidate is trustworthy only when the institution can explain who authorised the use, what evidence it used, and when the use must change or stop.
Governance begins before the model call
The FSB consultation proposes 12 sound practices for financial institutions adopting and using AI. Four concern organisation-wide governance; six concern stages of the AI lifecycle; and the remaining two address cyber and information-technology risk and third-party AI risk across those stages. The report describes the practices as a menu, not an international standard or a prescriptive instruction to adopt a particular technology. They are also non-exhaustive and may change as AI develops. That status matters: the consultation is a framework for responsible decision-making, not legal authorisation to use an LLM in supervision.
For a supervisory screening queue, the framework starts with a strategic decision. The board and senior management should be able to state what the use is meant to accomplish and what it is not meant to accomplish. The purpose might be to compare an applicable capital package with AT1 prospectuses, surface possible changes in scope or consequence, and help experts allocate review time. It should not be left as the vague objective of “using AI to improve compliance.” A vague objective makes it difficult to decide which outputs matter, who owns an error, or when the system has exceeded its remit.
Strategic direction also sets the boundary around automation. The BIS procedure assigns the LLM an initial comparison role and leaves permission, materiality, and actionability to supervisory and legal experts. An institution that adopts it should preserve that boundary in its operating policy, training, and system design. It might, for example, prohibit the queue from issuing a finding of breach, communicating a legal conclusion externally, or closing a review without an accountable expert disposition. Those are applications of the rank-not-judge principle, not claims that the FSB consultation imposes those exact prohibitions.
The distinction between a use case and a model is important here. The same underlying LLM could support a low-consequence internal retrieval task or a comparison that influences a critical supervisory process. Governance should attach to what the system does, what information it receives, and how its output affects decisions—not merely to the model’s brand or technical category. A system that only produces a queue may still deserve strong control if the queue determines which contractual risks receive scarce attention.
Materiality is not the same as risk
The FSB separates two ideas that are often compressed into one label. Materiality concerns the importance of an AI use case to an institution’s viability, critical operations, and ability to meet legal and regulatory obligations, including obligations toward customers. Risk concerns the attributes that can cause harm to the institution or its customers. They can be assessed together, but they answer different questions.