By Sid Newby | April 2026
In twenty-plus years of building and deploying litigation technology, I've learned to be skeptical of vendor hype cycles. I watched "cloud" get slapped on every product for a decade before it actually changed workflows. I watched "blockchain for legal" arrive with fanfare and leave without a forwarding address. So when every major legal tech vendor started stamping "agentic AI" on their products in Q1 2026, my first instinct was to reach for the skepticism. But after spending weeks digging into what's actually shipping -- and what it means for the people who do the work of litigation -- I think this one is different. Agentic AI isn't just a feature upgrade. It's an architectural shift in how legal work gets done, and litigation teams that don't understand it are going to get left behind.
What "agentic" actually means -- and why it matters
Let's start with definitions, because the term "agentic AI" is already being used so loosely that it risks meaning nothing at all. At Legalweek 2026 in March, observers noted that vendors were "relying on buzzwords like 'agentic,' 'streamline,' and 'accelerate'" with limited distinction between offerings.[1] So let me be precise.
Generative AI (what most legal professionals have been using since 2023) takes a prompt and produces an output -- a draft, a summary, an answer. It's a single-turn interaction. You ask, it responds.
Agentic AI is fundamentally different. An agentic system receives a goal, then autonomously plans the steps needed to achieve it, executes those steps across multiple tools and data sources, evaluates its own progress, and adjusts its approach -- all without requiring human input at each stage. Think of the difference between asking a junior associate to "summarize this deposition" (generative) versus telling them to "prepare the fact chronology for our summary judgment motion by pulling from depositions, key documents, and interrogatory responses, then flag any gaps in the record" (agentic). The second task requires planning, tool use, iteration, and judgment.
The reason this matters for litigation is structural. As I wrote in my previous analysis of the legal tech landscape, document review consumes roughly 73% of eDiscovery production costs, and the entire production lifecycle -- collection, processing, review, analysis, production -- has been organized around sequential human labor at every stage.[2] Agentic AI doesn't just speed up one step. It threatens to restructure the entire workflow by allowing machines to orchestrate multi-step processes that previously required teams of people handing work from one stage to the next.
Figure 1: The architectural difference. Generative AI handles discrete tasks; agentic AI orchestrates multi-step workflows with planning, execution, and self-evaluation.
The Q1 2026 agentic arms race: what every major vendor shipped
Between January and March 2026, virtually every major legal technology vendor launched or announced agentic AI capabilities. The velocity was remarkable -- and the competitive pressure behind it was unmistakable. Here's what actually shipped.
Thomson Reuters: CoCounsel Legal goes agentic
Thomson Reuters launched CoCounsel Legal in August 2025 with "Deep Research" capabilities, and by early 2026 had expanded it to over one million professional users -- making it the most widely deployed legal AI product in the market.[3] But the real news came in late 2025 and early 2026 with three new agentic capabilities entering beta:
- Bulk document review of up to 10,000 documents in tabular format, accelerating diligence and risk identification for M&A, litigation discovery, and regulatory compliance
- Agentic workflow planning and execution, where the system independently plans and executes complex, multi-step legal workflows using Westlaw and Practical Law content
- Custom workflow sharing, allowing lawyers to create, save, and share reusable workflow plans that combine Thomson Reuters content with firm knowledge[4]
In February 2026, Thomson Reuters announced the next generation of CoCounsel Legal, promising "more autonomous" AI agents that require less human supervision.[5] The company also expanded CoCounsel Legal to the UK market in January 2026, signaling global ambitions for the agentic platform.[6]
The strategic logic is clear: Thomson Reuters controls the most authoritative legal research corpus in the Western world. By making its agentic AI the default interface to that corpus, it's attempting to ensure that even as AI disrupts how legal work gets done, the content moat remains the competitive advantage.
LexisNexis: Protege replaces the old AI with an agent army
LexisNexis made perhaps the boldest move of Q1 2026. On February 24, it launched Lexis+ with Protege, an entirely new platform that replaces Lexis+ AI (its first-generation offering) with an end-to-end agentic workflow system.[7]
The numbers are striking:
- 300+ pre-built, configurable workflows covering litigation, transactional, and everyday legal tasks -- from drafting motions to dismiss, to generating discovery documents, to redlining agreements against internal standards
- A no-code workflow builder that lets firms create custom multi-step workflows without programming
- Infrastructure built on a legal knowledge graph spanning 200+ billion interconnected documents with 4 million new documents added daily
- Support for models from Anthropic, Google, and OpenAI -- a notable multi-model approach[7]
Sean Fitzpatrick, CEO of LexisNexis Global Legal, described the shift: "Legal professionals are increasingly seeking integrated legal AI work environments."[7] The more revealing detail is that advanced agentic workflows -- where Protege can "plan and execute complex, multi-step legal work autonomously" -- are marked as "coming soon," suggesting even LexisNexis views full agent autonomy as a progressive rollout rather than a day-one feature.[7]
The global rollout is ambitious: U.S. first, then Canada, U.K., Europe, and Asia Pacific throughout 2026, with practice-area-specific and additional agentic workflows expanding continuously.[8]
DISCO: first to claim "scaled agentic AI" in eDiscovery
In February 2026, DISCO announced what it called the industry's "first scaled agentic AI tool for fact investigation and eDiscovery."[9] The announcement added an autonomous, multi-step reasoning engine to DISCO's existing Cecilia Q&A tool -- transforming it from a question-answering system into an agent that can independently navigate document sets, build analytical tables, and surface patterns across case materials.
This is particularly significant because DISCO has already established strong benchmarks with its existing AI tools. As I noted in my earlier analysis, DISCO's Auto Review system demonstrated throughput of up to 32,000 documents per hour with published case studies showing 96.9% recall and 70.1% precision.[2] Adding agentic capabilities on top of that performance foundation means the system can now not just classify documents at speed, but reason about them across a multi-step investigative workflow.
Epiq: agentic AI wins industry recognition
Epiq expanded its Epiq AI platform in March 2026 with a suite of agentic solutions, winning the 2026 Legalweek Leaders in Tech Law Award for Best Use of AI in eDiscovery and Litigation.[10] The expanded platform includes:
- Epiq AI for Review (formerly Epiq AI Discovery Assistant): automatically develops review protocols, classifies documents at scale, and provides transparent metrics for defensibility
- Epiq AI for Privilege: automates privilege classification and log creation in compliance with user instructions[11]
Since its initial launch in January 2025, Epiq AI has been adopted by 130 clients including global corporations and law firms.[11] The Legalweek award recognition suggests the industry is already validating agentic approaches in production environments -- this isn't vaporware.
Harvey AI: the $11 billion agent factory
If the incumbents are adding agentic features to existing platforms, Harvey AI represents the opposite approach: building an entire company around the agentic paradigm from day one.
Harvey's Agent Builder, launched in early 2026, enables legal teams to create custom AI agents that handle multi-step tasks autonomously. The platform now processes 400,000+ agentic queries daily, with users extracting over 20 million terms through review tables and generating 445,000+ reports using Deep Analysis.[12]
The scale is staggering. More than 25,000 custom agents now operate on Harvey, executing work across M&A, due diligence, contract drafting, and document review. Customers are deploying what Harvey calls "long-horizon agents" -- systems that handle multi-step workflows over extended periods for complex use cases like fund formation.[12]
Harvey raised $200 million in new funding that values the company at $11 billion, with participation from Andreessen Horowitz, Coatue, and Kleiner Perkins.[13] That's up from the $8 billion valuation just months earlier in December 2025 -- a trajectory that suggests investors see agentic AI as the defining paradigm for legal tech.
The partnership between Harvey and A&O Shearman -- one of the world's largest law firms -- to "roll out agentic AI agents targeting complex legal workflows" is perhaps the clearest signal that this technology is moving from innovation labs into production practice.[14]
| Vendor | Product | Agentic Capability | Scale/Adoption |
|---|---|---|---|
| Thomson Reuters | CoCounsel Legal | Multi-step workflow planning, bulk review (10K docs), Deep Research | 1M+ users |
| LexisNexis | Lexis+ with Protege | 300+ pre-built workflows, no-code builder, multi-model | 200B+ document knowledge graph |
| DISCO | Cecilia Agentic AI | Autonomous fact investigation, multi-step reasoning | 32K docs/hour throughput |
| Epiq | Epiq AI Platform | Auto review protocols, privilege classification | 130 clients, Legalweek Award |
| Harvey | Agent Builder | Custom agent creation, long-horizon workflows | 400K queries/day, 25K+ agents, $11B valuation |
Table 1: The agentic AI landscape in legal technology, Q1 2026.
What makes agentic AI different from "AI-assisted review" -- and why the distinction matters for litigation
I want to be precise about what's actually new here, because the legal industry has been doing "AI-assisted review" for over a decade. Technology-Assisted Review (TAR) -- both the older TAR 1.0 (seed-set training) and TAR 2.0 (continuous active learning) models -- already uses machine learning to prioritize and classify documents. The platforms I discussed in my previous post -- Relativity's aiR, DISCO's Auto Review, Everlaw's Deep Dive -- all use generative AI to enhance the review process.[2]
So what's actually different about "agentic" versions of these tools?
1. Planning and decomposition
Traditional AI review tools execute a task you define: "classify these documents as responsive or non-responsive based on these criteria." An agentic system receives a higher-level goal -- "prepare the privilege log for this production" -- and decomposes it into sub-tasks: identify potentially privileged documents, classify the nature of the privilege (attorney-client, work product, joint defense), extract the required metadata fields, generate log entries, and flag edge cases for human review.
Harvey's Agent Builder makes this explicit: "Rather than rigidly moving through a fixed set of steps, a Workflow agent grasps the goal, asks for the right context, and finds the most direct path to a useful outcome."[12]
2. Tool use and data integration
Agentic systems don't just process text -- they use tools. Thomson Reuters' CoCounsel agents can search Westlaw, query Practical Law, analyze uploaded documents, and generate work product, all within a single workflow.[4] LexisNexis's Protege operates across its entire knowledge graph of 200+ billion documents while integrating models from multiple AI providers.[7] This multi-tool orchestration is what elevates the technology from "smart search" to "autonomous workflow."
3. Self-evaluation and course correction
This is the capability that most distinguishes agentic AI from its predecessors. A well-designed agentic system evaluates its own outputs, recognizes when it lacks sufficient information, and either adjusts its approach or escalates to a human. Harvey's agents include "human-in-the-loop checkpoints" that surface decisions and flag moments where human input would improve results.[12]
This matters enormously for legal defensibility. The concern with autonomous AI in litigation has always been accountability: who is responsible when the machine makes a mistake? Agentic systems that are designed to know what they don't know and escalate accordingly address the core ethical requirement that lawyers must supervise and verify AI outputs.
The risks that keep me up at night
I've been building technology for litigators for two decades, and I've seen enough "revolutionary" products to know that the gap between a demo and a defensible production can be enormous. Here are the risks I think the industry needs to confront honestly.
The autonomy-accountability paradox
The entire value proposition of agentic AI is that it requires less human supervision. Thomson Reuters explicitly describes its next-generation CoCounsel as requiring "less human supervision."[5] But the ABA's ethics guidance on generative AI emphasizes confidentiality, accuracy, and the lawyer's responsibility to supervise and verify outputs.[2] The Mata v. Avianca sanctions -- $5,000 for attorneys who submitted fabricated AI-generated case citations -- remain a fresh cautionary tale.[2]
The question litigation teams must answer is: at what point does "agentic" become "unsupervised"? And who bears the malpractice risk when an autonomous agent makes a privilege call that turns out to be wrong?
Harvey's approach -- embedding human-in-the-loop checkpoints as a core design feature -- is the right architectural answer.[12] But the competitive pressure to be "more autonomous" could push vendors to reduce friction in ways that create real liability for the lawyers who deploy these tools.
The measurement problem
At Legalweek 2026, a persistent observation was that vendors struggled to differentiate their offerings beyond buzzwords. One crowdsourced observation noted that "the real product clients are buying is confidence" -- clients seek results they can trust under pressure, valuing "legibility, defensibility, and reliability" alongside automation.[1]
This is the measurement problem in a nutshell: how do you validate an agentic workflow that made dozens of autonomous decisions across thousands of documents? Traditional quality control methods -- sampling, inter-annotator agreement, precision/recall metrics -- were designed for workflows where humans make individual decisions. When an agent orchestrates an entire workflow, you need end-to-end auditability that can trace every decision back to the reasoning that produced it.
The vendors that win won't be the ones with the most impressive demos. They'll be the ones whose audit trails can survive a Rule 26(f) conference and a hostile motion to compel.
The EU AI Act looms
In August 2026 -- just four months from now -- the EU AI Act reaches full application for high-risk AI systems. Legal AI systems fall squarely within the high-risk category, with penalties reaching EUR 35 million or 7% of global revenue.[15] The Act requires conformity assessments, risk management systems, and human oversight mechanisms.
For global litigation teams, this creates a direct regulatory constraint on how agentic AI can be deployed. An agent that operates with minimal human oversight in a U.S. review workflow might not comply with EU oversight requirements when the same matter involves European data subjects or cross-border discovery. The regulatory arbitrage between jurisdictions will become a practical compliance challenge for every major law firm and corporate legal department.
The access gap could widen before it narrows
Here's what worries me most. Right now, the firms deploying Harvey's Agent Builder with 25,000+ custom agents are the A&O Shearmans of the world -- global firms with the budget to invest in custom AI workflows and the technical staff to manage them.[14] Thomson Reuters' one million CoCounsel users are predominantly at firms that can afford premium Westlaw subscriptions.[3]
The promise of agentic AI is that it makes legal work faster and cheaper. But if the tools are only accessible to firms that are already well-resourced, the technology will widen the access gap before it narrows it. A solo practitioner facing a complex commercial dispute won't have access to Harvey's Agent Builder. A small firm handling employment litigation won't have 300+ pre-built Protege workflows.
This is the same pattern I've watched play out for two decades: powerful tools get built, powerful firms buy them, and everyone else makes do with what they can afford. The technology changes, but the gatekeeping dynamic stays the same -- unless the industry deliberately chooses a different path.
Figure 2: The access gap risk. Large firms and corporate legal departments are deploying agentic AI at scale, while smaller practices face cost and complexity barriers to adoption.
What Legalweek 2026 revealed about the industry's real concerns
Legalweek 2026 (March 9-12, New York) served as the industry's first major gathering since the agentic AI wave hit. The observations from practitioners on the ground are revealing -- and more nuanced than the vendor press releases would suggest.[1]
The "confidence economy"
The dominant insight was that "the real product clients are buying is confidence." Not speed. Not cost savings. Confidence -- the assurance that AI-produced work product can be trusted under pressure, defended in court, and relied upon by the lawyers whose names are on the filings.[1] This reframes the competitive landscape: the winning agentic platform won't be the fastest or cheapest. It will be the one that produces the most trustworthy results.
The EDRM's center of gravity is shifting left
Observers noted significant movement toward the "left side" of the EDRM framework -- data governance, cyber breach response, and analytics -- rather than traditional document review.[1] This is a natural consequence of agentic AI: if review becomes increasingly automated, the human-intensive work shifts upstream to data identification, collection strategy, and preservation decisions -- areas where legal judgment (not just document classification) is the scarce resource.
Vendor differentiation is thin
The most uncomfortable truth from Legalweek: many vendors struggled to articulate differentiation beyond wrapping language models onto existing products.[1] When every platform claims to be "agentic," the question becomes: which ones actually are? Litigation teams evaluating these tools need to probe beyond the marketing and ask hard questions about architecture, auditability, and real-world validation data.
A practical framework for litigation teams evaluating agentic AI
After spending weeks immersed in this research, here's the framework I'd recommend for any litigation team evaluating agentic AI tools in 2026.
1. Demand workflow transparency, not just output quality
Ask vendors: can I see every step the agent took, every source it consulted, and every decision point where it chose one path over another? If the answer is "we show you the final output with citations," that's generative AI with a footnote layer -- not a truly auditable agentic system.
2. Test the failure modes, not just the happy paths
Every vendor will show you the demo where the agent flawlessly prepares a privilege log in minutes. Ask instead: what happens when the agent encounters a document it can't classify? How does it handle contradictory instructions? What does escalation look like? The quality of an agentic system is best measured by how gracefully it handles uncertainty.
3. Evaluate the human-in-the-loop design
Not all checkpoints are equal. Some systems pause for human review at every step (which defeats the purpose of autonomy). Others only escalate at pre-defined thresholds. The best designs -- like Harvey's approach of surfacing decisions at critical junctures[12] -- balance efficiency with control. Ask: who decides where the checkpoints go, and can I customize them for my matter?
4. Consider the regulatory trajectory
With the EU AI Act hitting in August 2026,[15] any agentic system deployed on matters with European dimensions needs to satisfy high-risk AI system requirements. Ask vendors: do you have an EU AI Act compliance roadmap? Can you produce the conformity assessments the regulation requires?
5. Price for outcomes, not features
If agentic AI truly delivers the productivity gains vendors claim, the pricing model should reflect it. A system that reduces first-pass review time by 80% but charges per-document processing fees is just transferring cost from labor to software. Look for vendors that are willing to price on outcomes -- successful reviews, accurate privilege logs, defensible productions -- rather than compute consumption.
Where this goes next: the 18-month horizon
Here's what I think happens between now and the end of 2027, based on the trajectory I'm seeing.
Multi-agent orchestration becomes the norm
Today's agentic systems are mostly single agents working on discrete workflows. The next evolution is multi-agent systems where specialized agents collaborate -- one handling document classification, another managing privilege analysis, a third building case chronologies, and a supervising agent coordinating the ensemble. LexisNexis's Protege architecture, with its multi-model approach integrating Anthropic, Google, and OpenAI models, hints at this direction.[7]
The "review architect" becomes a real job title
As I predicted in my earlier analysis,[2] the staffing model for litigation is shifting from armies of contract reviewers to smaller teams of "review architects" who design workflows, set agent parameters, and validate outputs. Agentic AI accelerates this shift dramatically. The firms that invest in training their people to manage AI workflows -- not just use AI tools -- will have a decisive advantage.
Smaller firms get access -- eventually
The current generation of agentic tools is priced for Am Law 200 firms and major corporate legal departments. But the same commoditization dynamic that brought cloud eDiscovery to mid-market firms over the past decade will eventually bring agentic capabilities downstream. Relativity's decision to include AI review tools in standard RelativityOne pricing is an early signal of this trend.[2] I expect similar moves from other vendors within 18 months as competitive pressure forces prices down.
The courts will weigh in
We don't yet have significant case law on the use of agentic AI in litigation. But as these tools move into production on real matters, disputes about their reliability, defensibility, and appropriate scope will reach judges. The first Rule 26(f) conference where both sides present competing agentic AI workflows will be a landmark moment -- and it's coming sooner than most practitioners expect.
The bottom line
Agentic AI in legal technology is real, it's shipping, and it's going to reshape how litigation gets done. The Q1 2026 launches from Thomson Reuters, LexisNexis, DISCO, Epiq, and Harvey represent the most consequential product cycle in legal technology since cloud eDiscovery went mainstream.
But I want to be honest about what I see. The technology is impressive. The vendor claims are aggressive. And the gap between "agentic AI demo" and "defensible agentic AI production" is still wide in many cases. The Legalweek 2026 practitioners who emphasized confidence, legibility, and defensibility over speed and automation have it exactly right.[1]
For litigation teams, the imperative is clear: understand this technology now, evaluate it rigorously, and start building institutional knowledge about how to deploy it responsibly. The firms that wait for "agentic AI 2.0" will find themselves competing against opponents who have already integrated autonomous workflows into their standard operating procedures.
For the legal industry more broadly, the question I keep coming back to is the one I've been asking for twenty years: will this technology open the courthouse door wider, or just make it more profitable for the people already inside?
The answer isn't determined by the technology. It's determined by the choices we make about who gets access to it, how we price it, and whether we prioritize defensibility over dazzle. Agentic AI has the potential to compress the cost of competent legal work by an order of magnitude. Whether that savings flows to clients and to the access-to-justice mission -- or gets captured as margin by the same incumbents who've controlled the market for decades -- is the defining question of this era in legal technology.
I know which outcome I'm fighting for.