No Standalone AI Rulebook
Firms often search for a dedicated PRA AI rulebook and find none exists. The Bank of England, PRA, and FCA have consistently signaled through the AI Public-Private Forum and the 2022 DP5/22 discussion paper that AI and machine learning risks are addressed through existing regulatory frameworks rather than new AI-specific rules. This means compliance obligations for AI systems are derived, not standalone. Supervisors expect firms to extend model risk management (SS1/23), systems and controls requirements (SYSC), and senior manager accountability (SMCR) to cover AI use, including AI agents that make or influence decisions in regulated activities. For compliance leaders, this has a practical consequence: demonstrating AI compliance means demonstrating how multiple existing obligations apply jointly to AI systems, rather than pointing to a single AI certification or checklist.
SS1/23 Model Risk Management Principles Applied to AI
SS1/23 sets out five model risk management principles covering governance, model definition and risk tiering, model development, independent validation, and deployment with ongoing monitoring. Although written in general model risk terms, supervisors apply these principles to AI/ML models without exception. In practice this requires firms to maintain a model inventory that explicitly captures AI and AI agent systems, assign a risk tier based on complexity and materiality of use, and subject higher-tier models to more rigorous development and validation standards. For AI systems, independent validation extends beyond statistical accuracy checks to include assessment of bias, explainability, and robustness. Ongoing monitoring expectations under SS1/23 are not satisfied by a one-time approval; firms must demonstrate continuous performance tracking once an AI system is in production.
Accountability Under Existing Regimes
PRA supervisors expect accountability for AI-related risk to sit with an identified senior manager, consistent with the general accountability approach reflected in AIPPF and DP5/22 materials. This means firms cannot treat AI governance as a distributed or informal responsibility. A named individual should be accountable for the risks arising from a given AI system's systems and controls, in the same way accountability is assigned for other technology and operational risks. Where AI agents operate with a degree of autonomy, firms should document the permitted decision scope and the escalation triggers that route decisions back to human review, so that accountability is traceable even when a human is not directly executing each individual decision.
Applying These Expectations to AI Agents Specifically
AI agents that can take actions, call tools, or interact with other systems introduce a supervisory question that traditional statistical models did not: what happens when the system acts rather than merely predicts. SS1/23's deployment and ongoing monitoring principles apply here, but firms must translate them into controls that constrain agent behavior at runtime, not only at the point of initial validation. This includes agent identity and permissions that reflect least privilege, tool approval workflows that gate what an agent is allowed to invoke, and audit trails that record why a given action was taken. None of this requires new regulation to justify; it follows directly from applying existing model risk, systems and controls, and accountability expectations to systems that operate with greater autonomy than earlier generations of models.