The legal system has always required the right balance of speed and judgment. Agentic AI is emerging as a tool that can help achieve that balance in ways generative AI cannot; however, it demands a different conversation about accountability, oversight, and how legal work evolves
Key insights:
-
-
-
Agentic AI acts autonomously, creating new accountability challenges — Agentic AI acts and makes decisions with minimal human intervention, and this shift changes everything about responsibility and oversight.
-
-
-
-
-
With intentional design, the risks can be addressed confidently — Silent failures, accountability diffusion, and confidentiality breaches can only be mitigated through governance, testing, and rigorous human oversight.
-
-
-
-
-
AI is changing legal work, not eliminating it — When agentic AI handles routine tasks, legal professionals can move their attention onto higher-value work and increased responsibilities.
-
-
It is no longer useful to treat all AI as a single category or a tool for a single use case. Generative AI (GenAI) has already begun reshaping legal work by drafting documents, researching precedents, and answering questions with remarkable speed. At its core, however, GenAI remains a responsive tool. Agentic AI, on the other hand, represents a distinct evolution. Rather than waiting for a prompt, agentic AI systems can plan workflows, carry them out autonomously, and make decisions along the way.
As technology and the judicial system become increasingly intertwined, it is essential to examine where these more advanced tools intersect and what that convergence means for legal institutions. Ankita Upadhyay, Senior Director of AI Enablement at Thomson Reuters, recently shared her perspective during a recent webinar, Agentic AI in legal settings: Guardrails for responsible innovation, presented by the AI Policy Consortium — a joint effort by the National Center for State Courts (NCSC) and the Thomson Reuters Institute (TRI) — and offered valuable insight into the opportunities and responsibilities that accompany this shift.
One of the key notes to understand from the panel is that “generative AI gives you an answer. Agentic AI takes an action — and that distinction changes everything about the accountability,” Upadhyay said, adding that the distinction is not merely technical. It must reshape how we think about professional responsibility and the integration of AI into institutions that are built on trust and accuracy.
The promise of efficiency and transformation at scale
The potential of agentic AI is already visible in courts across the country. In Palm Beach County, Fla., for example, court officials are using agentic AI to process incoming documents at unprecedented scale. When an attorney files a document, the system autonomously identifies the document type, classifies it, extracts data, and routes it appropriately. The county already has processed up to 5 million documents using this process, operating 20 hours a day, every day of the year.
The most important part isn’t just the volume, however, it’s what happened to the people.
The staff who spent time on routine document processing were not laid off; instead, they were reassigned. Clerk 1 positions were transitioned to Clerk 3 and Clerk 4 roles, and that meant greater responsibility, more complex decision-making, and increased compensation for those making that transition.
“The staff that was doing all the processing of documents has been reallocated to customer experience and more complex tasks,” explained Parik Chokski, Director of IT for Palm Beach, on the webinar. This development reflects a broader truth: AI is not taking jobs in the legal sector; rather, it’s changing what those jobs entail.
You can explore the white paper Judicial Use of Generative AI: Lessons Learned here
As more repetitive work moves to AI, legal professionals move their attention toward the kind of work that demands their judgment, expertise, and accountability.
Risks are real, but not insurmountable
Yet the promise of agentic AI comes with genuine risks that differ from those posed by generative AI. Because agentic AI acts autonomously, for instance, failures can occur silently and invisibly, and sometimes repeatedly before detection.
Thomson Reuters’ Upadhyay identified three predominant risks for legal professionals and their organizations with agentic AI use:
1. Accountability diffusion — When an agentic AI system produces a document through a chain of autonomous decisions, it becomes difficult to determine where human judgment ended, and machine decision-making began. This ambiguity directly challenges professional conduct rules, which assume lawyers make every material decision. The result is an unclear line of responsibility and potential legal exposure for the lawyer.
2. Confidentiality at scale — Agentic AI systems operate across entire databases and multiple use cases simultaneously. A single misconfiguration of permissions can allow an AI agent to access privileged information to which it shouldn’t have access, potentially sharing sensitive client data across unintended matters. The danger lies in the fact that this often happens silently and repeatedly until discovered.
3. Irreversibility — Unlike GenAI, where a flawed draft often gets caught during review, agentic AI can send client communications, file documents, or update records based on faulty reasoning even before human oversight intervenes. The speed of action outpaces the speed of review, and thus, it creates a gap that traditional legal processes weren’t designed to address.
“The risk isn’t that AI gets it wrong,” Upadhyay said. “The problem is agentic AI systems, when it gets things wrong, it happens silently in a black box until you monitor it, and that’s the biggest challenge.”
Guardrails for responsible implementation
Given this, how do courts and legal organizations implement agentic AI thoughtfully? The webinar panelists, drawing on real-world implementations and NCSC research, emphasized several critical actions, including:
Establish clear governance — Begin with centralized registration of all agentic AI agents, conduct rigorous risk classification based on task impact, and start with low-risk workflows before advancing to high-stakes tasks. “Having a proper agentic AI governance is really important,” Palm Beach’s Chokski said.
Commit to rigorous testing — Extensive stress-testing in development and Q&A environments must precede any production deployment. Palm Beach’s implementation required weeks, if not months, of testing before going live — but that investment paid dividends in reliability and organizational confidence.
Design for transparency — Build workflows with built-in checks, balances, and fail-safes. Establish comprehensive logging that tracks what the AI agent does, what permissions it has, and what decisions it makes at each step. Monitor continuously for behavioral drift.
Maintain human oversight —”Trust but verify,” Chokski noted. Agentic AI is powerful and here to stay; but so are human professionals, and they must always retain oversight, the ability to intervene, and ultimate accountability for outcomes.
The conversation continues
The choice legal organizations face today is not whether agentic AI will exist, but how to engage with it responsibly.
Organizations that approach agentic AI with intentionality, clear frameworks, and commitment to human judgment will unlock its potential to expand capability, improve efficiency, and free legal professionals to do work that requires their expertise and accountability. Those that rush forward without guardrails risk silent failures that could undermine trust in both the technology and in the overall institution.
The path forward demands partnership: AI handles scale and speed, while humans provide judgment, accountability, and ethical reasoning. When those work two parts work together intentionally and with clear guardrails, that’s where justice is served.
For more on the impact of AI in courts, visit the TRI/NCSC AI Policy Consortium for Law & Courts
