Agentic AI represents autonomous legal workflow potential, but also more work for proper implementation; and as an ILTACON panel explored, task delegation, data context, and human-in-the-loop factors are key to AI agent success
Key takeaways:
-
-
-
Goal in mind — Agentic AI functions differently from past technologies in that it works towards a goal, which means understanding the tasks to get to that goal.
-
Importance of context — The more data an AI agent has, the better decisions it will be able to execute, making proper data architecture and interoperability extremely important.
-
Lawyer input required — Despite increased autonomy, agentic AI systems still rely on legal professionals for verification and review — and knowing when to bring in human knowledge will be key to the user experience.
-
-
NATIONAL HARBOR, Maryland — The buzziest term in legal technology currently is that of agentic AI. In theory, an AI agent presents even more technological autonomy than even generative AI (GenAI), presenting a piece of software that can simulate decision-making and even execute certain tasks in the legal workflow.
The prospect of making legal tasks even more streamlined has tantalized legal technologists. As a result, many incumbent legal technology companies are baking agentic AI into their workflows, more tech startups are bringing agentic AI concepts into the legal realm by the day, and technology teams at law firms and corporate law departments are beginning to explore what agentic AI may mean for them.
All of that is well and good, but like any technology, there’s a difference between simply adopting a new technology, and adopting it successfully. Like GenAI, general purpose AI, and a number of technologies before it, agentic AI needs the right conditions to properly thrive. Especially given that agentic AI may be the most complex piece of software that has ever been available for the legal world, simply setting it and forgetting it is not an option.
So, how can law firms and corporate law departments set themselves up for agentic AI success? According to legal technology leaders at the Orchestrating Intelligence: AI Agents in the Legal Space panel at the International Legal Technology Association Conference (ILTACON) 2025, proper agentic AI implementations start with knowing the tasks — yes, plural — at hand, and giving the system a whole lot of contextual data with which to reason.
Chaining tasks
Current technologies, and particularly the GenAI that legal professionals have become familiar with over the past three years, have become very good at executing specific actions. For example, you can ask a GenAI system, “Write me a shopping list for a family of four that gives healthy meals and minimizes costs,” and it will do that specific task very well.
The prospect of making legal tasks even more streamlined has tantalized legal technologists.
However, that’s not what agentic AI is built to do. Rather than execute a specific action, said panelist Joel Hron, Chief Technology Officer at Thomson Reuters, agentic AI instead is built to achieve larger goals. “You don’t tell it to do a certain thing, you tell it to achieve a goal,” Hron explained. “And it, with the context that it has and the information that it has, figures out the best way to achieve that goal.”
This means that rather than one specific action, agentic workflows are made up of tasks that build towards a goal. For example, panelist Adam Ryan, chief product officer at Litera, pointed to the email inbox. Lawyers send and receive business development emails all the time, asking how the firm can help on a given project, or what types of cases a lawyer is working on. Rather than needing to synthesize that information themselves, then draft a response (potentially with the help of GenAI or another tool), an AI agent can begin that work automatically.
“What an agent can do is look at that email, understand the context, and be able to proactively come back with the answer to that question,” Ryan said. Each of these are individual tasks, but they all move towards the larger goal of engaging in business development with this particular client. That way, the lawyer’s time is reduced from drafting to simply reviewing a response.
All of the context
Perhaps just as important, panelists said, is making sure the AI agent has the right context by which to make a decision. Think back to Ryan’s example — a response email will look very different if it’s to a long-standing client, a high-value client the firm is trying to get into the pipeline, or a low-value client that is being cold-called. Just like a first-year associate, the more the agentic AI system understands about the task it’s trying to perform, the better it’s going to be able to execute that task.
The panel noted that while agentic AI is intended to be more autonomous than past technologies, it will not be wholly making decisions independent of legal professionals.
That’s why Hron called context “a very important driver of performance” for those looking to implement agentic AI systems. It’s also why, because of the need for as much data as possible to make that context robust, he said that he expects data interoperability to play a major role as agentic AI becomes more mainstream. “I think what you’ll see with agents is that it will significantly break down barriers between software applications working with each other,” Hron said. “An agent that is going to do its job well is going to have as much context as it possibly can.”
Ryan agreed, noting that he expects a continued emphasis on capturing structured data that can easily be used for agentic systems, across all types of tasks and matters. “I think the firms that will succeed in this race are the firms with really good structured data sets of the entire experience.”
Lawyer in the loop
Finally, the panel noted that while agentic AI is intended to be more autonomous than past technologies, it will not be wholly making decisions independent of legal professionals. Panelist Matt Zerweck, Group Product Manager at Harvey, said his organization thinks about this in two discrete ways: i) The check-in process, in which legal professionals are brought into the agentic workflow to verify or provide input to a prompt; and ii) the review process, in which legal professionals can check the output and convert that output into achieving the overall goal.
Thomson Reuters’ Hron noted that a big part of agentic AI success will be the user experience for the tool, which if executed correctly, will lead to more trust in the system. This even occurs through small visual cues, similar to the way that hyperlinks in online articles let the reader know there is an external citation for what is written, he said.
Just like a first-year associate, the more the agentic AI system understands about the task it’s trying to perform, the better it’s going to be able to execute that task.
Indeed, user experience of a software has been built historically for humans to take actions, but now, Hron said that he thinks “the design principles of the future of software will be much more about maximizing the speed at which optimization and verification will happen.”
In the future, panelists noted, agentic AI will have the ability to learn user preferences and how to refine its workflows over time. Zerwick explained that legal tech developers are already thinking, “how do you go from constantly kicking off an agent, to the agent going without you doing anything, to the agent reaching out directly to you?”
To be sure, we are in the very early days for agentic AI in legal. But the panel said that when agentic AI becomes more commonplace, it will amplify the strong work attorneys already do. Hron said he sees AI agents as a boon for the strongest attorneys in particular, given that a lot of the more tedious work performed by all attorneys will become more automated.
“It makes the harder decisions more important,” Hron said. “So, it amplifies the best people in the organization, because it brings those decisions to the fore more often.”
You can find more of our coverage of recent ILTACON events here
