Already, some law firms are beginning to ponder how AI agents may become a part of their workflow; but integrating agentic AI means understanding how it’s different — and the new risks it brings
MYRTLE BEACH, S.C. — The past three years have seen a natural, albeit rapid, progression of AI capabilities in law. First came the large language models (LLMs) and platforms built on top of them such as ChatGPT, which allowed for simple interactions based on pre-existing data. Then came the introduction of retrieval-augmented generation (RAG), introducing more context-specific answers to prompts and the ability to develop code and other types of outputs. This also coincided with the ability for AI tools to call one another, creating an ever more complex system of interactions between AI programs.
Now, however, comes a new question: What if those interactions can happen autonomously, in which an AI program is given a specific task, knows what it needs to accomplish that task, and can call on other programs and systems without human input? Enter AI agents — which many legal technologists believe will be the next step in the AI evolution for the industry — that can develop more complex ways of working throughout autonomous execution of specific tasks.
As these AI agents can make calls over and over, said Rob Saccone, Chief Technology Officer at tech company Lega, something special emerges — the ability to autonomously fix a process that isn’t working. “We also discovered that with carefully crafted prompts, models could not only generate text, or code for us, they could develop plans for us,” Saccone said.
Machines developing autonomous plans may conjure images of the mythic robot lawyer. However, at a panel during the recent 2025 ILTA EVOLVE conference, Agentic AI: How this emerging area may impact the legal profession, the panelists said they view AI agents less as a robot lawyer and more a highly caffeinated assistant — one that can perform specific tasks very well and can interact with the environment on a 24/7 basis, but one that also takes regular oversight and training to function effectively.
What makes AI agentic?
Reanna Martinez, Director of Innovation, Systems, and Data at Munger, Tolles & Olson, said she regularly hears questions from attorneys about the differences between generative AI (GenAI), agentic AI, and other types of technologies in the marketplace. She delineated four key aspects that make a piece of technology agentic in nature:
-
-
- Primary function — Agentic AI is given a full goal to achieve, rather than a specific task.
- Autonomy — Agentic AI has high autonomous ability, as it figures out what is needed to accomplish a goal, rather than a discrete prompt every time.
- Learning and adaptability — Also high, with the ability to respond to new stimuli and the environment.
- Decision-making — Agentic AI has the ability to discern the best outcome to achieve its goal, then execute on that outcome in the real world.
-
Crucially, agentic AI is not the same thing as GenAI, Martinez added. Many GenAI systems aren’t agentic, as “they’re using LLMs to perform certain tasks, but they’re not autonomous, they’re written by software engineers to do something specific.” Agentic, meanwhile, means “churning through a problem to get to the best result.”
AI agents can also accomplish these goals around the clock, as “agents are taking things that are routine and time-consuming and creating 24-hour access,” she explained. In a supermarket, for example, an AI agent can not only determine whether the stock of a product is low but actively research the best method of procurement and place an order, all without human input.
These abilities can also extend to the marketing sphere. For instance, an AI agent can take a new product, research its audience and marketing strategy, conduct A/B marketing campaigns using email and Google, and more. “It’s doing all of these things like a normal marketing professional would, but it’s doing it 24/7 because it has a goal,” Martinez said.
The applicability to law firms
What does that mean for the legal industry? AI agents have a host of potential use cases, noted Sara Miro, Director of Knowledge Solutions at Sullivan & Cromwell. And many of them can cut down on a lot of (non-billable) time within a firm, such as:
-
-
- Marketing and lead generation — AI agents can look at client news, conduct research, and create pitch materials “24/7, doing work while you’re sleeping.”
- Client in-take and follow-up — They can also track client activity and monitor communications, generating reports and suggesting touchpoints, which can be particularly important for lawyers because “the business development part takes a lot of time and effort, and it’s non-billable.”
- Task recommendations — The agents can observe work processes and propose next steps, suggest precedent, and draft documents, in an overall effort to “suggest what you should be working on next to get the best result in this matter.”
- Voice assistants — If allowed, AI agents can listen to communications and book meetings, organize to-do lists, and prepare for the next steps of an engagement. An agent can even conduct an entire series of meetings, immediately booking a follow-up meeting invite “because it knows what you’re looking for.” Then, it can put it into all calendars, book a specific conference room that a partner wants, and collect materials for the meeting.
-
None of this occurs without oversight, of course, as well as some custom engineering to take into account firm desires and preferences. “I can’t imagine a situation where you’re going to give complete control over to these agents, even five years from now,” Miro said, adding that as preferences change over time, “there has to be flexibility.”
Today’s risks
Of course, any new technology is going to have an element of risk that law firms need to evaluate. Munger Tolles’ Martinez said she plots new applications on a risk/value matrix, factoring in security, the readiness of users and the environment, the level of accuracy, the impact if something goes wrong, cost, and ROI.
While the potential value to the firm can sometimes get lost in the nuances of new technology, it is perhaps the most important factor, she added. “The important thing is that you always have to start with your business use case — what is the thing you’re trying to solve?”
Miro agreed, noting that even though agentic AI may be new, “it’s tech, it’s software. The analysis really shouldn’t be that different from how you evaluate things today.” With that in mind, her main risk concerns look similar to other tech systems: a need to evaluate security and privacy, accuracy and reliability, and compliance and ethics.
However, there are a few additional wrinkles to agentic AI’s rollout, the panel explained. For example, autonomy: How much independent decision-making should be allowed? If an AI agent develops a business development use case while sleeping that’s great, but if the agent goes ahead contacts the potential client directly, that can create issues. Or if the agent takes the initiative to set up a follow-up meeting, that can be helpful; but the firm likely will not want the agent to autonomously select to whom the meeting invites get sent.
And as with GenAI itself, agentic AI is supposed to be additive to lawyers and law firm personnel, not a replacement for their skills. And now is the time to begin thinking how to upskill legal professionals to work with the next wave of technology, Sullivan & Cromwell’s Miro said.
After all, agentic AI systems are powerful and autonomous, but they’re not all-encompassing. “Don’t fire your teams,” she said. “There’s still a place for everyone.”
You can find more of our coverage of recent ILTA events here
