Devanshu Singh
Volume 2, Issue 2
Published: 16 October 2026
Financial institutions increasingly deploy advanced AI systems in workflows that influence underwriting, surveillance, liquidity management, customer communications, and regulatory reporting. As these systems become more tool-using, adaptive, and goal-directed, the central governance question shifts from whether a model is accurate to whether it remains governable when objectives, constraints, and incentives come into conflict. This paper argues that traditional model risk management remains necessary but is no longer sufficient for agentic AI. Existing frameworks are strong at conceptual soundness, outcome testing, and independent review, yet they were largely designed for models whose behavior is comparatively stable and whose failure modes are easier to observe. Agentic systems introduce behavioral risk, strategic risk, incentive misalignment risk, and control-evasion risk. The paper makes four contributions. First, it defines a practical taxonomy for agentic AI risk in regulated financial settings. Second, it extends validation from static accuracy testing to behavioral stress testing under shutdown, goal-conflict, restriction, and escalation scenarios. Third, it introduces the Agentic AI Risk Control Stack to connect use-case design, objectives, independent challenge, monitoring, and governance ownership. Fourth, it translates these ideas into implementable controls, practical metrics, and a tiered roadmap for institutions and regulators. The aim is not to predict catastrophic outcomes in the near term. Rather, it is to provide a governance architecture for the more ordinary but still material ways advanced AI may fail inside financial institutions: under-escalation, proxy optimization, overconfident rationalization, suppression of weak signals, or strategic adaptation to controls. Those risks are sufficiently important to justify dedicated testing and oversight now.
Advanced artificial intelligence systems increasingly display agentic characteristics: they can plan over multiple steps, adapt to feedback, pursue objectives through tools and workflows, and sometimes produce plausible explanations that conceal brittle reasoning or strategic behavior. Recent frontier-model evaluations suggest that, in controlled settings, some systems may exhibit behaviors consistent with specification gaming, hidden objective pursuit, or efforts to preserve operational continuity when incentives are poorly defined. In financial institutions, where model outputs shape material decisions and reporting processes, these dynamics create governance challenges not fully addressed by traditional model risk management. This paper develops a financial-sector framework for agentic AI governance built on four elements: a practical risk taxonomy, behavioral stress testing, independent validation, and a layered control architecture for monitoring and escalation. It also sets out practical metrics, governance ownership principles, and policy implications for supervisors. The core claim is that the relevant risk question is no longer only whether an AI system is accurate, but whether it remains governable under pressure.
Agentic AI, Model Risk Management, AI Governance, Financial Stability, Validation, Banking Regulation
Devanshu Singh*, Independent Researcher, Hoboken, New Jersey, USA.
Corresponding Author: Devanshu Singh, Independent Researcher, Hoboken, New Jersey, USA.
Devanshu, S. (2026). Agentic AI Risk in Financial Systems: A Model Risk Management Framework for Autonomous Decision-Making Systems. Econ Dev Glob Mark, 2(2), 01-12.