By Greg Dickason, Executive Vice President and Chief Technology Officer, LexisNexis Legal & Professional
Something shifted in legal AI over the past year. The shift is not about which model is most capable, or which platform has the most features. It is about what lawyers are actually asking AI to do.
From Answers to Outcomes
The way I describe first-generation legal AI is a sage sitting on a hill. You climb up, you ask a question, you get an answer, and you walk back down to do something with it yourself. The AI gave you knowledge. You still had to do the work.
Agentic AI brings the sage off the hill and into the workshop. When you ask them to do something, they act – they research, draft, verify citations, ask follow-up questions, and hand back something close to finished work product. You are not getting a response to carry back and act on. The work is getting done.
This matters because lawyers are not trying to produce AI outputs. They are trying to produce a brief, a contract review, a client memo. The gap between what AI returns and what a lawyer actually needs to deliver is what agentic workflows close. And that is a very different kind of value from what the first generation of legal AI was offering.
Trust Is Not a Feature – It Is the Prerequisite
Here is the challenge: agentic AI that closes the gap only matters if it does so reliably.
Moving from chat to work raises the stakes on accuracy in a specific way. When AI is answering questions, a lawyer can evaluate the response and decide what to do with it. When AI is producing work product, the verification layer has to be part of the workflow itself – not something added at the end.
Before a lawyer relies on any authority, two questions have to be answered: is this citation real, and is it still good law? Getting those answers used to mean a separate research step. With Shepard's® Verify, those signals appear in context, as part of the work. Upload a brief from opposing counsel, and Protégé checks every citation against our database: does it exist, is it still valid, how have subsequent courts treated it? The lawyer does not break flow to go verify. Verification comes to them.
That is the architecture that makes agentic legal AI usable in practice. Lawyers will adopt tools they can move quickly with – and they can move quickly when they can verify everything. Trust does not slow down the work. It is what makes the speed possible.
A Different Kind of Competitive Advantage
One of the most interesting things happening right now is how legal organizations are thinking about the model and platform question. A lot of firms have been experimenting with point solutions: a tool for research, a different one for drafting, another for review. What we are hearing increasingly is that clients want fewer tools, not more. They want something that can hold their way of working, not just a generic capability they have to configure from scratch every time.
The layer that matters in that context is not which foundational model is running. It is how well the platform understands legal work. A good platform should be able to draw on whichever model serves a particular use case best. What does not commoditize is legal content. Two hundred billion documents, four million added daily, 150-plus years of Shepard's intelligence. That is the fuel that makes the model actually useful for legal work.
Firms are also beginning to understand that they can bring their own way of working into the platform. Their playbook – the way their practice group approaches a particular matter type – can become a skill that the agent draws on. That is a meaningful shift from using a generic AI tool to using one that actually reflects how a firm works.
What This Means for the Work Ahead
I am genuinely optimistic about what agentic legal AI makes possible over the next few years. Some of this is work that simply did not exist before. For example, litigation strategy built on judge-specific outcome data and regulatory signals embedded in live deal documents. And some of it is about who gets access to legal expertise at all. Law is underserved in many parts of society, and AI that lawyers can trust and verify gives the profession capacity to serve more people in more places.
Also, junior lawyers are going to develop faster. For example, working through mock trials, mock depositions, richer training environments, getting to senior-level judgment more quickly than previous generations could. The floor of the profession rises.
But none of that happens unless lawyers trust the work being produced. The real shift in legal AI is not from one model to another, or from one feature set to the next. It is the move from answering questions to helping lawyers get work done. As AI moves from chat to work, trust becomes more than a requirement. It becomes the foundation that makes the entire transition possible. The measure of a legal AI platform is not what it can generate. It is what a lawyer can confidently put their name on.
To hear more of my thoughts on this topic, check out my recent conversation on The Geek in Review podcast with hosts Greg Lambert and Marlene Gebauer.
# # #
For media inquiries about this post, please contact Jill Van Nostran, jill.vannostran@lexisnexis.com