来源:发现报告(www.fxbaogao.com/view?id=5725989)| 引用请注明出处Governing AFC AgentsHow model governance and validation must evolve for agentic AI in financial crime compliance来源:发现报告(www.fxbaogao.com/view?id=5725989)| 引用请注明出处© Oliver Wyman2AbstractFinancial institutions are beginning to deploy agentic AI in anti-financial crime (AFC) systems while the regulatory landscape continues to evolve. Existing supervisory guidance on model risk management (MRM), emerging horizontal AI governance frameworks such as the EU AI Act, and industry principles for AI in financial crime provide important governance principles. Yet none provides a complete framework for agentic AFC systems, whose autonomous planning, tool use, dynamic workflows, and delegated decision-making introduce a distinctive governance problem.This paper proposes a practical governance framework for agentic AFC systems that institutions can apply today. Rather than replacing existing guidance — including interagency model risk management guidance, the Wolfsberg Principles on AI, and the Financial Action Task Force’s (FATF) work on AI — it identifies the governance principles that remain applicable across existing supervisory guidance and introduces additional controls needed to govern agentic systems operating in AFC environments.Throughout the paper, we use the US Office of the Comptroller of the Currency (OCC) interagency model risk management guidance as a worked example because it provides a governance lifecycle familiar to many compliance and model risk teams, illustrating how existing governance principles can be extended to agentic AFC systems. The resulting recommendations are intended to stand as a governance framework for agentic AFC systems in their own right and can be applied across supervisory and regulatory regimes.Whether developed internally or sourced from a third party, each material agent should connect to an enterprise agent assurance layer that allows the institution to configure, test, observe, challenge, constrain, and stop the system, and to retain the evidence needed for oversight.The central argument is that effective oversight must shift from governing a discrete model to governing the broader agentic system. The inherent materiality of that system is determined by the business use case, but its residual risk, and the corresponding governance and controls it requires, can be shaped through engineering choices around autonomy, tool permissions, workflow constraints and runtime safeguards. Institutions that adopt this broader governance perspective will be better positioned to deploy agentic AI safely, effectively, and in line with evolving supervisory expectations.来源:发现报告(www.fxbaogao.com/view?id=5725989)| 引用请注明出处ContentsAbstract 21. Governing agentic AI for AFC before dedicated guidance exists 42. Defining the agentic AFC system 73. Applying existing governance principles...