Michael James Bommarito (273 Ventures; ALEA Institute; Stanford Center for Legal Informatics; Michigan State College of Law; Bommarito Consulting, LLC), Daniel Martin Katz (Illinois Tech – Chicago Kent College of Law; Bucerius Center for Legal Technology & Data Science; Stanford CodeX – The Center for Legal Informatics; 273 Ventures; ALEA Institute) and Jillian Bommarito (273 Ventures; ALEA Institute) have posted Agentic AI in Law and Finance: Navigating a New Era of Autonomous Systems on SSRN. Here is the abstract:
Agentic AI systems, meaning software capable of autonomous goal pursuit, environmental perception, and iterative action, are rapidly entering legal and financial services. Yet widespread definitional confusion, architectural opacity, and governance gaps leave professionals ill-equipped to deploy them responsibly. We offer a comprehensive treatment of agentic AI for high-stakes professional domains, proceeding in three movements: definition, design, and governance.
We begin by establishing definitional clarity. Drawing on nearly a century of scholarship across eight disciplines, from Anscombe’s philosophy of intention to recent advances in large language models, we propose a three-level hierarchy of agency: Level 1 (Agent), defined by the minimal properties of Goal, Perception, and Action; Level 2 (Agentic System), which adds Iteration, Adaptation, and Termination implemented through traditional algorithms; and Level 3 (Agentic AI), which fulfills these six properties using AI for planning and orchestration. A practical evaluation rubric helps practitioners distinguish genuine agentic systems from sophisticated tools and single-shot chatbots.
We then turn to architecture, because agents are not magic; they are architecture. Organizing the analysis around ten fundamental design questions spanning inputs (triggers, intent, perception, memory), execution (planning, delegation, action tools), and safety (termination conditions, human escalation, governance), we show how each design decision carries real tradeoffs that determine what a system can do, how reliably it performs, and how it fails.
Finally, we propose a governance framework that scales oversight to each system’s risk profile, situating agentic AI within a five-layer regulatory stack spanning foundational law, professional ethics, sector-specific regulation, emerging AI-specific rules such as the EU AI Act, and voluntary assurance standards. In industries built on trust and non-delegable professional duties, we argue, human-in-the-loop and human-in-command architectures are not merely best practices but critical designs for satisfying fiduciary and regulatory obligations. We conclude with organizational models, responsibility-assignment tools, and a maturity-based adoption path that enable legal and financial institutions to capture the benefits of agentic AI while managing liability exposure and reputational risk. Throughout, we argue that responsible adoption requires architectural literacy, the necessary bridge between technical implementation and professional obligation.
Recommended!