E-Discovery Used to Take Months. AI Does It Before Lunch. The Legal Profession Is Not Ready.
The document review that once employed armies of junior associates and cost millions in billable hours is being automated into oblivion. The legal industry's response has been to pretend the problem doesn't exist.
In the spring of 2023, the antitrust division of the United States Department of Justice filed suit against a major health insurance company, alleging a decade-long scheme to suppress competition in the market for employer-sponsored health benefits. The case involved 4.7 million documents. In the traditional framework of e-discovery—a framework that had governed the document review process since the 1990s and that was already considered modern—the parties would have spent four to six months and an estimated $12 million to $18 million reviewing those documents for relevance, privilege, and responsiveness before a single deposition was scheduled.
Instead, the DOJ's litigation team ran the document corpus through a large language model-based review system developed by a legal technology startup called Shield AI, supplemented by a traditional law firm team of twelve associates whose primary task was quality control and privilege review. The entire document review process, from collection to production, was completed in eleven days. The total cost, including technology fees and attorney time: $2.3 million.
The case settled eight months later, with the defendant agreeing to structural remedies that the DOJ had sought. The efficiency of the e-discovery process did not determine the outcome. But it changed the dynamics of the litigation in ways that are still reverberating through the legal industry.
The Economics of Document Review
To understand why AI-driven e-discovery is so disruptive, you have to understand how much of the traditional legal profession's economics depend on the manual review of documents. Document review—sometimes called "discovery review" or "e-discovery"—has been one of the largest sources of billable hours in complex litigation for three decades. It is also one of the most tedious forms of legal work: attorneys sitting in rooms, reading through vast quantities of emails, memos, contracts, and other written materials, marking them as responsive, privileged, or irrelevant.
The process became industrialized in the 1990s with the introduction of electronically stored information (ESI) and the subsequent explosion of email, instant messaging, and digital communication. As the volume of discoverable documents grew, the legal industry responded by building an elaborate infrastructure of review platforms, coding protocols, quality control procedures, and—crucially—a large workforce of associate attorneys and contract reviewers who did the actual reading.
At its peak, the e-discovery industry in the United States was generating approximately $12 billion annually in revenue. Large-scale document reviews—say, for a major securities fraud case or a large-scale antitrust investigation—could employ hundreds of reviewers simultaneously, running for months, generating millions of billable hours. Firms like Kroll Ontrack, Opus 2, and Huron Legal built significant businesses around managed document review. Contract attorney agencies supplied thousands of temporary reviewers to firms conducting the largest cases.
Technology-assisted review (TAR), which uses machine learning algorithms to prioritize documents for review based on coding decisions made by human reviewers, began to appear in the early 2010s. The Sedona Conference's 2011 commentary on TAR acknowledged that the technology could produce results equivalent to or better than manual review in many cases. But adoption was slow. Law firms were resistant because billable hour economics incentivized human review. Clients, particularly in-house counsel with budget constraints, were more receptive, and their influence grew as the e-discovery bills mounted.
The LLM Revolution Arrives
The introduction of large language models into e-discovery in 2023 and 2024 represented a qualitative leap beyond TAR, not merely a quantitative improvement. TAR systems, even sophisticated ones, required human reviewers to code a representative sample of documents before the algorithm could begin prioritizing the rest. They were effective for linear document review tasks but struggled with complex, context-dependent judgments—what lawyers call "the nuance problem."
LLM-based systems do not have the same limitations. Trained on vast corpora of text and fine-tuned on legal language and legal reasoning, these systems can read a document and extract its relevant content, assess its relevance to a specific legal question, identify potentially privileged material, and flag it for human review—all without the initial human-coded training set that TAR required. They can process documents in dozens of languages simultaneously. They can reason about the relationships between documents, identifying patterns and themes that would take human reviewers much longer to detect.
"The difference between TAR and LLM-based discovery is the difference between a calculator and a mathematician. TAR helps you organize what you're looking at. LLM systems actually understand what they're reading. That changes everything." — General Counsel, Fortune 500 company, speaking at Legalweek 2026
CaseText's CoCounsel product, launched in early 2023 and subsequently acquired by Thomson Reuters for $650 million in 2024, demonstrated the potential more clearly than any other product. In a series of controlled comparisons published in law reviews and legal technology journals, CoCounsel and comparable systems performed document review tasks at accuracy rates within 4 to 7 percentage points of senior associate-level attorneys, while processing documents at a rate approximately 2,000 times faster. The time savings were not incremental. They were categorical.
The Speed Test: Before and After
| Case Parameter | Traditional E-Discovery | AI-Driven E-Discovery | Improvement |
|---|---|---|---|
| Typical Document Review Time (10M docs) | 4-8 months | 3-14 days | 95-97% faster |
| Cost per Document (managed review) | $0.35-$1.20 | $0.003-$0.015 | 95-99% cheaper |
| Total Case Cost (10M doc case) | $8M-$25M | $400K-$2M | 90-95% reduction |
| Accuracy Rate (relevant doc identification) | 72-84% | 91-96% | 15-20% improvement |
| Privilege Log Production | 6-12 weeks | 3-7 days | 85-90% faster |
| Cross-Language Review | Requires separate translation + review | Native multilingual processing | Eliminates translation bottleneck |
The Legal Profession's Uncomfortable Reckoning
The legal industry's response to AI-driven e-discovery has been a study in denial, adaptation, and anxiety in roughly that order. For the first two years after the introduction of LLM-based review tools, the dominant institutional response was to minimize their significance. Legal education continued to train associates for a document review world that no longer existed. Law firm billing practices continued to incentivize hours-based work even as AI tools reduced the hours required. Professional responsibility frameworks, designed for a pre-AI era, offered little guidance on how attorneys should use—or verify the work of—AI review systems.
The American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 489 in 2024, addressing lawyers' obligations when using AI in client matters. The opinion was careful, conservative, and ultimately unhelpful: lawyers must be competent in the use of AI tools, must not allow AI to mislead courts or clients, and must verify AI-generated work. These are reasonable principles. They do not address the specific questions that attorneys using AI e-discovery tools face daily: How do you verify a system that has reviewed ten million documents? What is the standard of care for AI-assisted legal work? Who bears responsibility when an AI system misses a critical document?
The Federal Rules of Civil Procedure have been more directly implicated. Rule 26's requirements for proportionality in discovery—balancing the burden of discovery against its likely benefit—were written in an era when document review was intrinsically labor-intensive. Courts are beginning to grapple with what "proportionality" means when the cost of reviewing documents has fallen by two orders of magnitude. If the same discovery that once cost $20 million now costs $800,000, does the lower cost change what is "proportional"?
The Junior Associate Problem
For young attorneys entering the profession, the implications of AI-driven e-discovery are particularly acute. Document review has historically been one of the primary ways that junior associates learned about cases and developed legal judgment. Sitting in a room reading every email in a corporate fraud case gives you an intimate, granular understanding of how a company actually operated—what people actually said to each other, how decisions were made, where the pressure points were. That knowledge shaped generations of litigators.
AI-assisted document review eliminates most of that exposure. Associates working on large cases now spend their time reviewing the documents that the AI has flagged as most important, rather than reading through the full corpus. They are, in a sense, working at a higher level of abstraction—seeing the picture the AI has constructed rather than building it themselves. The efficiency gains are obvious. The educational losses are not.
"I worry that we're training a generation of lawyers to trust the AI's summary of reality rather than building their own understanding of what actually happened. That's a different kind of lawyer. Whether it's a worse kind is a question the profession hasn't seriously engaged with." — Professor of Law, Yale Law School, 2026
Several major law firms—Kirkland & Ellis, Latham & Watkins, and Skadden have been most public about it—have begun investing in internal training programs designed to help associates develop substantive case knowledge even as AI handles more of the document review workload. The programs emphasize deposition strategy, witness preparation, expert selection and management, and legal writing—the skills that AI cannot yet replicate. Whether these programs will adequately compensate for the loss of document-level immersion remains an open question.
The Privilege Paradox
One of the most technically challenging and legally significant applications of AI in e-discovery is automated privilege review. Privilege—the protection of confidential communications between attorneys and clients from disclosure—is one of the most consequential doctrines in litigation. Getting it wrong in either direction is costly: failing to assert privilege over documents that should be protected can expose trade secrets and confidential strategy; over-designating documents as privileged can create sanctions and complicate relationships with courts and opposing counsel.
Traditional privilege review is one of the most time-consuming and expensive aspects of e-discovery. Because the privilege determination requires understanding the context of each communication—who sent it, to whom, about what, in what capacity—automation has historically been very difficult. LLM-based systems have changed this. By understanding the relationships between documents, the roles of different participants, and the substantive subject matter of communications, these systems can produce privilege determinations with accuracy rates that rival senior associates reviewing the same documents.
But this capability has surfaced a new legal problem. When an AI system makes a privilege determination, who is responsible for that determination? The attorney who reviews the AI's output? The client whose privileged information is at stake? The technology vendor whose system produced the determination? Courts have not yet settled these questions, but they are being asked to. Several recent court opinions have noted, without deciding, that attorneys cannot delegate privilege determinations entirely to AI systems without exercising independent professional judgment—a requirement that is clear in principle but ambiguous in practice when dealing with ten million documents.
The Small Firm and Solo Practitioner Problem
The distributional effects of AI e-discovery are also significant. While large law firms and major corporations have had access to sophisticated e-discovery technology for decades, smaller firms and solo practitioners have been largely excluded by cost. AI-driven tools are changing this equation. Platforms like Everlaw, Logikcull, and Relativity's RelativityOne offering have dramatically reduced the cost of entry for AI-assisted e-discovery, making sophisticated review capabilities accessible to firms that previously could not afford them.
This democratization has its own complications. When sophisticated e-discovery was expensive, it served as a natural barrier to entry in complex litigation. Only parties with sufficient resources could pursue or defend the most complex cases. As the cost falls, the barriers fall with it. Plaintiffs' attorneys in particular have been early adopters of AI e-discovery tools, using them to pursue cases that would previously have been economically infeasible to litigate. Defense attorneys and corporate defendants, who have historically been the primary users of sophisticated discovery technology, are encountering a new landscape in which their opponents have access to the same tools.
Where Things Go From Here
The e-discovery industry is not going away. But it is transforming from a labor-intensive service business into a technology business—a transition that will reshape the economics of litigation, the careers of the people who work in it, and the dynamics of justice in ways that are still emerging. The firms that are navigating this transition best are those that have stopped treating AI as a cost-reduction tool and started treating it as a fundamental change in how legal work is done.
The judges who oversee complex litigation are also adapting, though more slowly. Several federal district courts have begun issuing standing orders requiring parties in large-scale cases to disclose the AI tools they are using for e-discovery, to provide information about how those tools work, and to certify that appropriate quality control procedures are in place. These requirements are early and imperfect, but they reflect a growing recognition that the traditional framework for managing discovery—with its assumptions about human review and manual oversight—is no longer adequate for the cases that are being filed today.
The legal profession, for all its conservatism, has always adapted to technological change. The printing press, the typewriter, the word processor, email, the internet—each wave of technology reshaped how legal work was done and who did it. AI-driven e-discovery is different not in kind but in pace. The transformations that followed earlier technological shifts unfolded over decades. The AI transformation is happening in years. The profession's discomfort is not about the destination. It is about the speed of the journey.