From TAR to GenAI: Rethinking eDiscovery with Large Language Models
A Turning Point in Legal Technology
The legal profession has always had a cautious yet pragmatic approach to technological change. But with the advent of generative AI (GenAI), we are witnessing a shift that is impossible to avoid. Large Language Models (LLMs), such as those behind tools like ChatGPT, have changed the way we interact with information, and their impact on the legal world is only beginning to unfold.
The American Bar Association makes it an issue of professional competence: “lawyers should become aware of the [GenAI] tools relevant to their work so that they can make an informed decision, as a matter of professional judgment, whether to avail themselves of these tools (American Bar Association, Formal Opinion 512, July 29, 2024)". Adopting GenAI is becoming a necessity more than an option.
We've Been Here Before
While the headlines around GenAI may feel unprecedented, the business world has seen this kind of disruption before. The introduction of spreadsheets democratised accounting, moving modelling and analysis to business users. Search engines democratised access to knowledge. Machine learning and predictive coding ushered in a new era for document review in the form of Technology Assisted Review (TAR).
Each wave of innovation brought new challenges – accuracy, auditability and version control are just a few examples – but ultimately, the challenges were resolved. GenAI is no different. It brings novel risks, but also unprecedented opportunities.
Why Use an LLM Instead of Just AI-Powered Applications?
AI in legal technology isn’t new. Many tools already use machine learning, knowledge graphs or natural language processing to deliver specific outcomes, including sorting emails, extracting clauses or predicting case outcomes.
LLMs are different. Like spreadsheets in the 1980s, they are general-purpose tools. Straight out of the box, they do nothing in particular – but when configured and directed, they can do almost anything.
What makes LLMs uniquely valuable?
- Language understanding and generation: They can summarise, translate, draft and interpret.
- Pattern recognition: They identify connections across documents and formats.
- Domain flexibility: A single model can be used across use cases, from contract review to privilege detection.
However, they are not without their flaws. Common risks in the deployment of LLMs include:
- Hallucinations: LLMs can generate plausible but inaccurate text.
- Consistency: Getting the same answers every time can be challenging.
- Opacity: The logic behind their outputs can be hard to follow.
- Data leakage: Data leakage can occur, especially when using public cloud-based tools.
LLMs in the Legal Environment: Direct Use vs. Embedded Use
Legal professionals typically engage with LLMs in one of two ways. The first is direct use – tools like ChatGPT that allow open-ended interaction with the model. The second is through legal applications that embed LLMs behind the scenes as we are now seeing with eDiscovery tools.
At its core, eDiscovery involves trawling through vast volumes of documents to find what matters: identifying relevance, connections and significance. These are precisely the tasks LLMs are built for.
Addressing the Risks: Managing the LLM vs. Managing the User
Lawyers should follow two broad strategies for mitigating the risks of LLM adoption: managing the model itself and managing how users interact with it.
- Managing the LLM: Rather than trust public LLMs, lawyers should deploy models in secure environments – a “walled garden” approach which allows organisations to control what data the model sees and what it can generate, but establishes guardrails, such as limiting outputs to certain formats or data sources. Service providers are beginning to offer these models.
- Managing the user: Lawyers should establish training and governance around the use of AI. Prompt engineering – teaching users how to frame requests effectively – is essential. So is putting in place policies around data handling, approval workflows and audit trails.
Effective prompt engineering is critical to achieving reliable results from an LLM. The best prompts take into account the following:
- Stakeholders: Who is the output for? A partner? A regulator?
- Persona: Who is the model acting as? A junior associate? A data privacy officer?
- Action: What exactly is being asked? A summary? An extraction? A translation?
- Detail: Are there reference outputs that help guide the model? Is more context required?
- Expectations: What form should the result take – bullets, paragraphs, tables? How long? How detailed?
- Sources: What materials or data are being provided?
Understanding these elements enables lawyers to get consistent, useful outputs, reducing the risk of errors or hallucinations.
Rethinking eDiscovery: LLMs in Action
For law firms and in-house counsel, eDiscovery is already familiar ground. Technology professionals spent the last two decades developing workflows around keyword searches, TAR and predictive coding.
Adding an LLM into this process builds on that expertise. It enhances the core task of finding the signal in the noise. And crucially, it does it faster and more intuitively.
From TAR to LLMs: What's New?
Where TAR relied heavily on pre-labelled datasets and statistical modelling, LLMs offer a more fluid, intuitive way to understand content. Their ability to “read” documents much like a human does – and to summarise, contextualise and interpret – is game-changing.
LLM-driven tools in eDiscovery, such as Relativity aiR, bring three standout benefits:
- Faster than TAR: While traditional TAR workflows require time-consuming rounds of training and validation, LLMs can begin generating useful results almost immediately. Initial document triage can be reduced from days to hours, depending on data volume and complexity. For large-scale matters, this speed translates directly into cost savings, both in reviewer time and in broader litigation spend.
- More transparent: One of the long-standing criticisms of TAR has been the “black box” nature of its results. LLMs can improve this dramatically. Many platforms now include “chain of thought” features, offering a clear explanation of how the model reached a particular conclusion. This transparency is especially valuable when defending decisions around privilege or relevance.
- Easier to implement: Unlike TAR, which required carefully curated training sets and dedicated workflows, LLMs can be deployed in more flexible, user-friendly ways. They do not demand a team of data scientists or specialised project managers. Instead, they can slot into existing review platforms and processes with minimal disruption.
The rate of adoption
With many businesses implementing GenAI across their organisations to drive efficiencies, it’s no surprise that service providers are also being asked how GenAI can improve their offerings.
We are seeing that GenAI eDiscovery methodologies have a faster rate of adoption compared to the previous technology advancements such as TAR/CAL. Legal and technology professionals are now using Gen AI technology on cases seated in the U.K. and Cayman courts, and in response to regulatory investigation. GenAI in eDiscovery can lead to both time savings for legal and technology professionals and cost savings for the organisation.
A General-Purpose Tool Within Reach
The relative ease of implementation raises a broader question: is the LLM-enhanced eDiscovery platform still a niche tool, or is it evolving into something more general-purpose?
Many legal teams are beginning to use these same tools for adjacent tasks:
- Internal investigations: Aids in rapid identification of key communications and anomalies in large data sets.
- Regulatory inquiries: Helps pre-filter responsive documents before review or submission.
- DSAR responses: Identifies and summarises personal data in complex document collections.
- Contract audits: Extracts clauses, obligations or risks from legacy contracts en masse.
In each of these use cases, the core challenge is the same: finding what matters within unstructured, heterogeneous and voluminous content. This is where LLMs excel: not by replacing human judgment, but by making it faster and easier to apply.
For lawyers familiar with the evolution of eDiscovery, the value proposition is clear. LLMs do not eliminate the need for expertise, they elevate it. They reduce the friction between insight and action, enabling lawyers to focus on strategy rather than search.
And in doing so, they offer not just a better way to review documents, but a smarter way to manage risk, respond to complexity and deliver value to clients.