TAR and Generative AI Team Up to Achieve Success in Large-Scale Discovery

In this case, the solution was not TAR or gen AI.  It was TAR and gen AI, deployed in sequence.  TAR supplied the speed and scale needed to triage millions of documents.  Gen AI then served as a precision-enhancing filter for the subset of documents TAR identified as potentially responsive.  The results were compelling: in a 3.6 million-document population, gen AI eliminated 27.5% of documents it analyzed from the responsive TAR set—more than one in four documents that otherwise would have moved toward production consideration. 

The Solution: A Layered Review Strategy

In these circumstances, our preferred approach was to layer gen AI over TAR, such that gen AI analyzed only the subset of documents where its contextual reasoning would add value by improving precision and reducing overproduction risk.

TAR as the Foundation

At the same time, TAR has a known tradeoff: a document that has a responsiveness score above the TAR cutoff may be a false positive, i.e., it is either not responsive or otherwise out of scope.  In our example, TAR was identifying as responsive documents that otherwise met production criteria except for the fact that they were outside the jurisdictional scope, such as concerning only U.S. activities and not relating to the foreign jurisdiction where the document requests were focused.  Thus, we needed more than TAR alone.

Generative AI as a Precision Layer

The workflow was straightforward.  TAR-eligible documents were scored across six issue-specific TAR models, model-specific control sets established cutoff scores, and documents above the cutoff were treated as TAR-positive for that issue.  Gen AI then analyzed those TAR-positive documents using prompts aligned to the relevant scope.  A document moved forward as responsive only if it cleared both thresholds: it scored above the TAR cutoff and gen AI classified it as responsive.  Responsive document families then went through privilege review.

The Result: Reduced Volume, Increased Precision, and Lower Risk

Of the 375,675 TAR-positive documents that gen AI classified as not responsive, 350,981 were excluded based on issue scope. Critically, however, another 24,694 documents were excluded after gen AI prompts assessed whether the documents related to the regulator’s specific national jurisdiction.  Gen AI’s capacity for contextual analysis allows it to exclude substantively similar documents that may be legally out of scope based on factors such as jurisdiction, product parameters, or the identity or role of the participants.

Defensibility Considerations

A layered workflow does not eliminate the ability to demonstrate transparency and defensibility, but it may change how teams perform validation.  Although TAR control sets were used to determine cutoff scores, once gen AI is added, those control set statistics alone are no longer sufficient to statistically validate the end-to-end process.  If formal validation is required, the discovery team should design a validation method that accounts for the combined workflow, such as a stratified sampling approach drawing from documents treated as both responsive and nonresponsive. 

Key Lessons Learned

Gen AI is a powerful addition to the discovery toolset.  Gen AI brings capabilities in addition to TAR, particularly where responsiveness turns on nuanced contextual analysis.  When deployed thoughtfully, gen AI can materially improve precision without abandoning the scalability of TAR.

Gen AI deployment should match operational realities.  Although gen AI capabilities continue to advance, in matters involving massive populations, rolling data ingestion, and tight deadlines, TAR remains an important consideration.

Gen AI adds value with sophisticated contextual analysis.  Gen AI was effective at identifying documents that look similar but differ in legally significant ways.  This level of nuance is particularly helpful in allowing more precise responsiveness determinations in global investigations that have a specific jurisdictional scope.

Layering technology makes proportionality more practical.  Applying gen AI to the documents TAR determined were most likely responsive reduced cost, improved precision, and helped manage risk.

Conclusion: Consider Combining TAR and Gen AI

Robert Keeling is a co-managing partner of Redgrave and a nationally recognized authority on eDiscovery.  He serves as discovery counsel in complex, data-intensive matters, including litigation, investigations, and regulatory reviews.

Amy Hanke is counsel at Redgrave.  Her practice focuses on eDiscovery in complex litigation, regulatory actions, and investigations, including ESI protocol negotiation, technology assisted review and predictive coding, sensitive document reviews, and privilege issues.

Allison Myers is vice president of data analytics at Consilio, where she leads a team of 45 lawyers and programmers delivering custom analytics solutions for large, complex, and data-intensive eDiscovery matters.  A frequent CLE presenter with more than twenty years in eDiscovery, she combines a legal education with technical expertise to help clients and counsel turn document populations into actionable intelligence.

Mark Resnick is senior director of AI and analytics consulting at Consilio with more than 15 years of experience helping legal teams apply advanced analytics and AI to complex eDiscovery matters.  He works with clients to design, validate, and defend technology-driven approaches to document review and investigations that improve efficiency and reduce cost, and he has played a key role in Consilio's adoption of generative AI and other emerging technologies.

Reprinted with permission from LegalTech News.  © 2026 ALM Media Properties, LLC.  Further duplication without permission is prohibited. All rights reserved.  Please do not redistribute.

The views expressed in this article are those of the authors and do not necessarily represent the views of their law firm or any of its clients.