AI model risk: what the Bank's Consortium is telling boards to rebuild

The Bank of England's Artificial Intelligence Consortium has delivered one of the more consequential regulatory signals of the summer. Its June minutes, published on 5 August, conclude that generative AI systems strain existing model risk frameworks and that firms should manage risk from an 'AI model system perspective' rather than model by model (Bank of England). That is a structural shift, and it lands directly on senior accountability.

Key Executive Takeaways

• The Bank of England's AI Consortium has concluded that generative AI cannot be governed through traditional single-model risk frameworks and requires oversight of the full interconnected system.
• Boards should expect scrutiny of third-party AI transparency, human-in-the-loop controls, adversarial testing and auditable documentation as the emerging baseline for responsible adoption.
• Co-chairs David Geale and Sarah Breeden are signalling that regulatory support for AI adoption is conditional on firms demonstrating governance across the whole model system, not just individual components.

The Consortium's workshop on explainability and transparency identified recurring problems that will be familiar to any chief risk officer who has tried to onboard a GenAI tool: system complexity, interconnectivity, limited transparency from third-party providers, and variability of outputs (Bank of England). The members' conclusion is that these features make it difficult to apply existing model risk frameworks proportionately. In plain terms, the model risk management approach most large firms spent the last decade refining does not map cleanly onto a system where a foundation model, a retrieval layer, a fine-tuned wrapper and a human reviewer all shape the output.

The response the Consortium sketches is worth reading carefully. It recommends governance across the full AI model system, transparency on how components interact, and testing both the whole system and its individual elements, noting that models can be updated independently (Bank of England). That last point matters. If a third-party provider silently updates a base model, the firm's validation work may be stale within days. Boards that have signed off AI use cases on the basis of a point-in-time assessment are carrying a risk they may not have priced.

The political framing is equally important. Co-chair Sarah Breeden noted that the rapid pace of AI development in financial services underscores the importance of initiatives such as the AIC in identifying how regulators can best support responsible AI adoption (Bank of England). Read alongside co-chair David Geale's welcome, this is the Bank and the FCA jointly positioning themselves as enablers, provided firms can evidence the controls the workshop set out: human in the loop, adversarial testing, and accountability and auditable documentation of decisions (Bank of England). The quid pro quo is explicit even if unstated.

For senior leaders, three stakeholder shifts follow. Model risk teams, historically a second-line function, will need direct board access on AI matters because the risks now span procurement, technology, compliance and conduct. Third-party providers will face harder contractual demands on change notification and interpretability, and those that resist will be quietly deselected. And internal audit functions, which have often lagged on AI, will be expected to test the system as a whole, not tick boxes on individual components.

The Consortium's minutes are not policy. They are a preview of it. Firms that wait for a consultation paper before rebuilding their AI governance will be doing the work under a deadline rather than on their own terms.