An ILTACON 2026 panel explores how law firms can govern agentic AI teams, which can now plan and act on their own, while keeping human oversight front and center.
Listen to this article
Key insights:
- Agentic AI shifts law firms from “AI that answers” to “AI that acts,” raising the stakes on oversight—Unlike generative AI chatbots that respond to prompts, agentic AI systems can independently break down a goal into subtasks and execute them. As a result, legal professionals need new governance frameworks, not just prompting skills, to manage the added risk.
- Every task delegated to an AI agent should be treated as a “one-way” or “two-way” door — Reversible tasks (like an agent summarizing client emails for a lawyer to review) are “two-way doors” that are safer to automate, while irreversible actions (like an agent replying to clients without human review) are “one-way doors” that require much stricter human oversight.
- Law firms should govern agentic AI by defining permissible use cases, mandating observability, and requiring accountability—Experts recommend that law firms create audit trails and decision-tracing for AI agents, restrict which data agents can access, and prevent individual lawyers from creating new agents without firm-wide approval.
NASHVILLE — Not all AI is created equally. It may seem that way, given how quickly new innovations roll out and become a part of legal workflows. However, if legal professionals equate generative AI (GenAI) systems like ChatGPT to agentic AI systems that automate whole parts of the workflow, they may be setting themselves up for failure.
“A few years ago, the state of the art was asking a question of a chatbot and getting a response. Now, we’re talking about AI systems that can take an objective, break it into subparts, and get an answer,” says Valerie McConnell, Vice President of Solutions Engineering at Thomson Reuters. “We’ve gone from AI that answers to AI that acts. That shift raises the stakes of using AI considerably.”
At the 2026 ILTACON conference, one panel, Secret Agentic Man: When AI Starts Acting on Its Own, tackled the resulting question that legal professionals should be asking themselves: How much human intervention is necessary for agentic AI systems to be trusted?
The answer, or course, depends on how the agentic system is being used. What is clear, however, is that the optimal solution combines lawyers’ legal know-how and relationship management with the ability to automate low-level tasks.
Setting boundaries before deploying AI agents
Leyla Samiee, Chief Product and Technology Officer at software company Introhive, breaks down how agentic AI agents have the potential to change the legal workflow. In the past, the lawyer would define the goal, find a path, and execute. With current agentic AI, however, a human defines the goal, but the model finds the path, and the execution trigger is decided by the lawyer up front. And with autonomous AI in the future, even that optimal trigger could be decided by AI.
She then puts this in practical terms: Imagine a partner leaves their firm after 11 years. The clients are impacted by a change in work and communication styles, but the new lawyers on the matter don’t realize it until it’s too late. However imagine instead, she says, an agentic world in which an agent has insight into calls and emails to determine if the pattern of discussion is changing. Or even on day one, the agent sees deactivation of a user ID and can determine where the firm may be at potential risk.
“The agent didn’t make a decision, but it informed you enough to make a decision,” Samiee says.
Of course, this does require a lot of trust. Ali Mohammed, CEO of software company JuristAI, says that when talking governance, he aims to balance risk and costs. “When you are delegating or ceding that type of responsibility of something that is going to go out into the universe, the agent is going to have to touch so many surfaces to develop a competent knowledge base,” Mohammed explains.
We’ve gone from AI that answers to AI that acts. That shift raises the stakes of using AI considerably.
This means that those law firms considering agentic AI need to be proactive about determining what data the agent can and cannot touch. What if it co-mingles with client data? Or export controls? Are what if certain clients are against AI use? “These types of things need to be deterministically outlined before you set the agent loose,” he says.
Building these walls may be easier said than done. Some early firms that have ventured into the agentic space have encountered unexpected ethical dilemmas. “As we start moving into this brave new world of having agents more like co-workers than like a tool, we need to start thinking about, the agents may be co-workers, but they’re not humans,” Mohammed adds. “It’s going to start becoming more of the lawyers’ professional responsibility to teach them what they can and cannot do.”
Reversibility and the one- or two-way door test
As AI continues to advance, however, this type of planning may just be the starting point. Victoria Grech, Chief AI Officer at software provider iCompli by LegalRM, told attendees that “the future of work is nothing like you can picture it right now.” Grech says she expects mixed teams of agents and humans to be a regular part of work within the next few years, with agent accounts already being developed for major platforms in Microsoft, Salesforce, and more.
However, there is a risk with those agentic-centric teams, she adds: If agent swarms decide there’s no leader, they’re even tougher to govern. “If there’s a flock of birds, who’s leading that?”
“As you’re building agents now, you can see where they went wrong. But at the big level, interpretability is an issue,” Grech explains. “The pace we are going with autonomous agents, are we going to get in trouble?”
To combat this, Grech touts the importance of reversibility, the ability to fix an error agentic AI may make. She says she thinks about this as a one-way or two-way door; any task a lawyer delegates to an agent should be a two-way door that allows the lawyer to return through the other side.
The pace we are going with autonomous agents, are we going to get in trouble?
For instance, an agent may be tasked with summarizing emails with clients and developing an action plan. That’s a two way door, because the lawyer can reverse the summary or plan if it’s not correct. However, if the agent is also tasked with answering client emails without human review, that’s a one-way door that’s not as easy to reverse.
Ultimately, panelists agree that now is the time to begin instituting policies that help govern the parameters of these agentic teams. Thomson Reuters’ McConnell provides some practical next steps for firms looking to enter this area, such as:
- define permissible use cases, allowing for reversibility;
- mandate observability, making sure you’re using agentic systems with an audit trail and decision-tracing; and
- create mechanisms for accountability, notably that lawyers can’t create new agents on their own without the firm’s input.
There’s a balance between risk and reward, the panel notes. “We’re at the stage now we can’t possibly control everything the agent does,” McConnell says. “It doesn’t make sense to double-check everything the agent does, because then what’s the point of the more efficient system?”
At the same time, however, the AI agent team member should ultimately serve the needs of the lawyer, adds JuristAI’s Muhammad. “You guys are craftsmen. You are artisans. You have spent years developing your own style of being a lawyer. Human oversight will become more important than it has in the past.”
You can findmore coverage of ILTACON from this year and past years here
Have questions?
Get in touch with one of our solutions experts….
Get in touch
Contributors
Zach Warren
Senior Manager
Thomson Reuters Institute
