Most lawyers who have used a generative AI tool have used it the same way: they type a question, read the answer, and decide what to do with it. That pattern suits a discrete task such as summarizing a deposition transcript or proposing an outline for a brief. Litigation preparation rarely arrives in discrete tasks, however. Opposing a motion to compel requires the team to reconstruct the meet-and-confer history, reread the disputed requests and the objections served, test the privilege log against what was withheld, research the governing standard in the forum, and draft an opposition whose every citation holds up. The next generation of legal AI is being built to carry out sequences of that kind, and the more important change concerns where attorney judgment has to sit, rather than what the software can do.
Systems designed for this work are usually described as agentic. Instead of answering a single prompt, they pursue a defined objective through a series of steps, often by dividing the work among specialized components that pass results to one another. This post argues that agentic AI will be useful in litigation to the extent firms treat it as a supervised team rather than an autonomous one. In practice, that means designing attorney review into each stage of the workflow instead of inspecting only the finished product.
Agentic systems carry out multistep work rather than single exchanges
A conversational assistant responds to a prompt, and each response stands on its own. The lawyer supplies the context, decides the order of the work, and carries results from one step to the next. An agentic system reverses that arrangement: it takes a goal, breaks it into steps, calls on tools such as a research database, a document repository, or a docketing calendar, and feeds the output of each step into the next. In more elaborate designs, several specialized agents each handle one function while a coordinating process routes work among them. The vocabulary is not standardized, and vendors apply the word "agent" to products ranging from a fixed chain of prompts to systems that choose their own next action, so a firm evaluating these tools should establish which kind it is being offered.
The distinction matters because litigation tasks depend on one another. How a team reads a Rule 34 request determines what gets reviewed; the review determines what is produced and what is logged; the log shapes the motion practice that follows. When a chatbot errs, the error usually appears in the answer the lawyer is reading. When an agent errs at the first step, the mistake may be buried several stages upstream of the document that finally reaches the lawyer's desk, carried forward and built upon by every step that came after it.
Specialized agents can mirror the division of labor on a case team
A hypothetical matter shows how the work might be divided. In Meridian Analytics, Inc. v. Halvorsen Data Partners, LLC, an illustrative federal dispute, Meridian alleges that a former licensee kept and used its proprietary models after the license ended. The parties have exchanged requests for production, Halvorsen has produced documents along with a privilege log, and Meridian's counsel suspects the production is incomplete. Preparing a motion to compel under Fed. R. Civ. P. 37(a) is the kind of project an agentic system is designed to coordinate.
In a well-designed system, each component would take a piece of the work that a case team now divides among associates, paralegals, and litigation support staff. Research would go to one agent, which would assemble authority on proportionality under Rule 26(b)(1) and on the forum's local rules governing discovery motions [1]. A second, devoted to document analysis, would examine the production for gaps, such as custodians who appear in correspondence but contributed no documents, or date ranges that stop at the termination notice, and would compare each log entry against the description Rule 26(b)(5)(A) requires [1]. The chronology of the license term, the termination, and any post-termination access would be built by an evidence organization agent, which would tie each entry to a Bates number. Drafting the motion, including the certification of good-faith conferral that Rule 37(a)(1) demands, would fall to another component [1]. Finally, a verification agent would confirm that every cited authority exists and says what the draft claims.
Much of the time on a project like this is spent on coordination: locating the right document, reconciling the chronology with the log, and tracking which arguments rest on which exhibits. That is where agentic design is most promising. Whether the savings are large in practice remains an open empirical question, and figures published by vendors deserve the same scrutiny a lawyer would give any party's self-reported data.
Supervision has to be built into each handoff
Delegating steps to software does not delegate responsibility for them. ABA Model Rule 1.1, through its comment on technological competence, expects lawyers to understand the benefits and risks of the technology they use, and ABA Model Rules 5.1 and 5.3 impose supervisory duties over the work of others on a matter [2]. The ABA has addressed how these and related duties apply to generative AI in Formal Opinion 512 [2]. In federal court, the attorney who signs a filing certifies under Rule 11(b)(2) that its legal contentions are warranted by existing law, and courts have sanctioned lawyers who filed briefs containing fabricated citations produced by AI tools [3][1]. An automated verification step reduces that risk, but it cannot take the signer's place, because the certification belongs to the lawyer.
The practical consequence is that review should be placed where judgment, not processing, is the work. Mechanical steps can run without interruption, but a step whose output embeds a judgment that later stages will inherit should not proceed until a lawyer has approved it. In the Meridian workflow, a defensible design would require attorney sign-off at the following points:
An attorney approves the interpretation of each disputed request before document analysis begins, because scope decisions govern every later step.
An attorney reviews each privilege or work-product call the system proposes, since those determinations turn on purpose and audience in ways the documents may not reveal.
An attorney checks the chronology entries the motion relies on against the underlying documents.
An attorney reads each authority cited in the draft before the motion is signed and filed.
Checkpoints of this kind do more than catch errors. They create a record of who approved what, which is the record a supervising partner, a client, or a court may later ask to see. A system that cannot show its sources and its intermediate steps makes that record impossible to produce, and that limitation alone should weigh heavily against adopting it.
Confidentiality raises a separate set of concerns. ABA Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of client information, and an agentic system that moves documents among several tools multiplies the points at which that information can be exposed [2]. Firms should know where client data is stored, whether it is used to train a vendor's models, and who at the vendor can reach it. Whether sharing privileged material with an AI provider risks waiver is not settled and may vary by jurisdiction [4][5]. Fed. R. Evid. 502 and a well-drafted protective order offer some protection against inadvertent disclosure in federal proceedings, but neither answers the vendor question directly.
Coordinated tools could narrow the resource gap between large and small firms
Large firms have long held an advantage in complex litigation that has little to do with the talent of their lawyers. A firm with a research department, a litigation support group, and an established e-discovery vendor can staff a discovery dispute like the one in Meridian with a team; a solo practitioner opposing that firm handles every step personally. Coordinated AI tools could narrow that gap by giving smaller practices the capacity to analyze a large production or assemble a chronology without hiring staff to do it. If that happens, it could change how mid-sized disputes are litigated and widen the pool of clients who can afford capable representation.
The promise comes with a condition that falls hardest on the firms it would help. A small firm's verification burden does not shrink with its headcount, and the Rule 11 obligations of a solo practitioner are identical to those of a partner with 20 associates. Smaller practices will benefit only if the tools they adopt make review efficient, by surfacing sources and flagging uncertain calls, rather than producing polished output that invites a busy lawyer to skip the check. Whether clients see lower costs will also depend on billing arrangements, since efficiency under an hourly model accrues to the client only if the hours actually fall.
The measure of these tools is the quality of the representation
None of this points toward replacing litigators. The decisions that determine outcomes in a matter like Meridian (whether to move to compel at all, how hard to press a privilege dispute, when to trade a discovery fight for a settlement conversation) depend on judgment, negotiation, and knowledge of the client that no workflow automates. What agentic systems change is the allocation of attorney time, moving it away from assembling materials and toward deciding what those materials mean. Firms that use the reclaimed time to produce more work faster will capture part of the benefit. Those that use it to examine the record more closely and advise clients more carefully will capture more.
A sensible first step is modest. Before adopting any agentic tool, a firm should map one real workflow from its own practice, mark the points at which a lawyer's judgment is indispensable, and require the system to stop for review at each of them while preserving a record of its sources and intermediate steps. The open question is how courts and bar regulators will define adequate supervision of systems that act across many steps. Some courts already require parties to certify how AI was used in preparing filings [6], and those requirements will likely be tested as agentic tools become common in discovery and motion practice.
References
Fed. R. Civ. P. 11(b)(2). (This is the December 1, 2024 print. Check for any later amendment.) https://www.uscourts.gov/sites/default/files/2025-02/federal-rules-of-civil-procedure-dec-1-2024_0.pdf
ABA Comm. on Ethics & Pro. Resp., Formal Op. 512 (2024). (Note: I could not fetch the ABA's own PDF because access to americanbar.org was blocked. Cite the ABA original once you have checked it.) https://www.thesedonaconference.org/sites/default/files/meeting_paper/9.8%20aba-formal-opinion-512.pdf
Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. June 22, 2023) (opinion and order on sanctions). (A reporter citation appeared only in an unverified search snippet. Confirm it before adding one.) https://caselaw.findlaw.com/court/us-dis-crt-sd-new-yor/2335142.html
United States v. Heppner, No. 25 Cr. 503 (JSR), slip op. at 8 n.3 (S.D.N.Y. Feb. 17, 2026) (ECF No. 27). https://www.akingump.com/a/web/ssTGsd5NHbtZ1onzXQMTye/1_25-cr-503-27-memorandum.pdf
Cleary Gottlieb Steen & Hamilton LLP, Courts Grapple with Privilege Implications of AI (Feb. 27, 2026) (discussing United States v. Heppner (S.D.N.Y.) and Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026)). (I did not read the Warner order itself. Fetch it from the court docket before citing it directly.) https://www.clearygottlieb.com/news-and-insights/publication-listing/courts-grapple-with-privilege-implications-of-ai
D. Colo., Magistrate Judge Standing Order Requiring Certification re: Use of Artificial Intelligence (AI) in Filings (Oct. 21, 2024). (Confirm the order is still in effect.) http://www.cod.uscourts.gov/Portals/0/Documents/Judges/SP/SP_Standing_Order_AI_Filings.pdf
