AI model risk: what the Bank's Consortium is telling boards to rebuild
The Bank of England's AI Consortium has concluded that generative AI breaks existing model risk frameworks and requires a system-level approach to governance. For bank, insurer and asset manager boards, that reframing changes who owns AI risk and how third-party dependencies must be documented.
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.
What this reveals
The Consortium's conclusion exposes a structural gap between how firms have been governing AI and how regulators now expect them to. Most boards signed off GenAI use cases on the basis of point-in-time model validation and vendor assurances, assuming existing model risk frameworks would stretch to cover them. That assumption has quietly failed: when a foundation model, retrieval layer, wrapper and human reviewer all shape the output, and any component can be updated independently by a third party, single-model governance produces a false sense of control. Other leadership teams may wrongly believe their MRM framework, refreshed for SS1/23, already handles this, when in fact the accountability, documentation and testing model regulators now expect is materially different.
Questions accountable leaders should ask
- 01Can we produce, today, a documented view of every component in each material AI system we operate, including third-party dependencies, and identify who owns risk at the system level rather than the model level?
- 02How would we know if a third-party provider had silently updated a base model or its guardrails, and what would trigger revalidation of the use cases that depend on it?
- 03Are our human-in-the-loop controls, adversarial testing and explainability evidence designed for individual models, or for the full interconnected system as the Consortium now expects?
- 04If a supervisor asked which SMF holder is accountable for AI model system risk, and asked to see the evidence base behind that accountability, would the answer be clear and defensible?
- 05Where has internal confidence in our AI governance been shaped by vendor assurances or pilot-stage sign-offs that have not been retested against current regulatory expectations?
What accountable leaders should do now
- 1Commission a rapid inventory of material AI use cases mapped at the system level, showing components, third-party dependencies, update rights, and the human controls layered on top, so the board can see what it is actually governing.
- 2Reassign accountability explicitly: name the SMF holder responsible for AI model system risk, and revise Statements of Responsibilities and committee terms of reference so ownership does not fragment across model risk, technology, data and business lines.
- 3Rework third-party AI contracts and assurance processes to require notification of material model updates, transparency on component interaction, and rights to conduct or commission system-level testing, not just point-in-time model reviews.
- 4Pressure-test a small number of high-impact AI use cases against the Consortium's emerging baseline, explainability, adversarial testing, auditable documentation, whole-system testing, and use the gaps to reset the governance framework before the next supervisory conversation.
- 5Brief the board on the shift from model-by-model to system-level governance, including where existing sign-offs may now be stale, so the record shows the board understood and responded to the regulatory reframing rather than absorbing it passively.
Explore the practical guide
This guide sets out what board accountability actually means in regulated financial services and how directors can demonstrate it under regulatory scrutiny. After reading, you will know how to structure oversight, evidence judgement, and avoid the common failures that turn ordinary decisions into personal liability.
Read the guideWhere the operating environment may be moving faster than internal reporting reflects
Polar Insight's Signal Briefings translate emerging regulatory, stakeholder, and market developments into a clear implication for accountable leaders.
Explore Signal BriefingsRelated insights
The Daniel Thomas decision: appointed representative risk returns to the boardroom
The FCA has decided to ban and fine former adviser Daniel Thomas £742,700 for reckless defined benefit pension transfer advice given without the required qualifications, while his firm operated as an appointed representative of Quilter Financial Services. The case reopens hard questions for principal firms about oversight, qualification verification, and the residual liabilities that sit above the AR relationship.
Fuenmayor decision puts SMCR disclosure duties back in the spotlight
The FCA's decision to fine BancTrust CEO Carlos Fuenmayor £99,600 for failing to disclose regulatory investigations and account freezes lands as a pointed reminder that personal disclosure obligations under the Senior Managers regime carry real consequences. For boards and nomination committees, it sharpens the question of what they actually verify about senior hires, not what candidates choose to share.
Agentic finance arrives: the FCA signals testing over consultation
Jessica Rusu's 19 October speech confirms the FCA will embed six of the seven Mills Review recommendations into live supervisory work rather than consult further, as 86% of surveyed firms now use AI. Senior leaders should expect direct engagement, assurance testing and surveillance on AI deployments, not another rulebook.
Stakeholder Signals
Consequential developments in financial services and other regulated markets, with one implication for accountable leaders.
