AI in eDiscovery: knowing when and how to use it

Maximizing efficiency by combining tools

Well before generative AI, eDiscovery involved selecting and combining the best technologies for each case. Generative AI does not change this. It expands the range of tools available—and makes effective workflow design even more important. The challenge: understanding where AI adds value, where traditional approaches are better suited, and how to combine technologies to deliver more efficient, accurate, and defensible reviews.

Search terms and metadata filtering

Metadata filtering and keyword searching remain foundational even when using AI. The courts recognize this. For example, in Schulte v. LinkedIn Corporation (N.D. Cal. 2026), the court refused the plaintiffs’ request to prevent the defendant from using disclosed search terms to pre-cull documents before generative AI review. The court considered that using search terms to pre-cull documents before technology review satisfies the reasonableness and proportionality standards in the procedural rules (absent any showing that the terms were deficient). 

Metadata filters such as date ranges, custodians, file types, and communication participants provide an objective and defensible means of reducing data volumes before further analysis. In some cases - such as finding email traffic between key individuals around the time of a key event - they can be more efficient and reliable than AI inference. 

Keyword searching also remains important. It can be a blunt instrument, but its greatest strength is transparency (a term is either present or not), making searches easy to understand, explain, and negotiate with others.

AI as a search accelerator

Generative AI tools can, though, help to quickly and effectively develop keyword strategies by identifying related concepts, alternative terminology, key individuals, and patterns that may otherwise be overlooked. AI assistants that are equipped with retrieval-augmented generation (RAG) tools (which the AI can use to retrieve relevant information from a defined data source to answer a user’s question) can generate keyword suggestions based on the language within an actual document population - grounding them in the terminology and context of the documents themselves.

In parallel, AI is changing the role of keyword searching. As with machine learning before it, AI reduces the reliance on search terms alone to cull a dataset. And AI is expanding access to advanced search features; it is increasingly capable of generating complex search queries from natural language prompts.

Further, AI is expanding the types of evidence that can be searched. Vision models can analyze images, screenshots, handwritten notes, and diagrams, reducing reliance on OCR.

Machine learning and generative AI

Continuous active learning (CAL) is one of the most proven and widely accepted eDiscovery technologies. It is well understood by courts and regulators. Now, generative AI can be combined with CAL for more accurate and efficient results.

In the first article in this series, we explained how CAL can prioritize documents identified as potentially relevant by generative AI review workflows.

It’s also possible to use generative AI to identify high-quality “seed” documents for CAL, accelerating and improving CAL training. This can be more effective than using, for example, random sampling, manual searches, or traditional analytics. It can be particularly valuable in low-richness document populations, i.e., where relevant documents are a smaller fraction of the overall population.

When not to use AI

AI is not the answer to every challenge - despite enthusiasm for the technology or pressure to adopt it. As with any technology, success depends on selecting the right tool for the task at hand.

Small and low-richness document populations

For smaller populations, the cost of configuring, testing, and validating generative AI may outweigh the benefits. Similarly, low-richness matters can require extensive validation samples to achieve confidence in the results.

Defensibility risks

Defensibility is often paramount in eDiscovery. Courts and regulators have become increasingly receptive to AI-assisted review, but acceptance in any specific case often depends on the ability to explain and validate it. Where review decisions are likely to be challenged or the stakes are especially high, the additional scrutiny AI may attract should be carefully weighed against its benefits. 

The use of AI introduces the possibility of requests to disclose the prompts used - which can be problematic if they contain sensitive information. The legal position is still evolving on the extent to which generative AI prompts may be disclosable. For more, see our insights on protecting legal privilege when using AI generally and in the U.S. context.

The limitations of retrieval-augmented generation (RAG)

RAG-powered AI tools can effectively assist lawyers to develop an understanding of a matter, surface key documents, and identify relationships between people and events. This can significantly accelerate early case assessment and investigation.

But unlike keyword searches and metadata filters, RAG tools prioritize relevance over completeness. They identify the potentially “most relevant” information, not necessarily “all relevant” information. To generate a response, they retrieve and analyze only a subset of potentially relevant content - typically a fixed number of the highest-ranked passages, or only those scoring above a minimum relevance cut-off (their “retrieval thresholds”). Responses may therefore omit evidence that falls outside these retrieval thresholds, is buried within lengthy documents, or is expressed in unexpected ways.

This means that RAG should not replace other search and review technologies used for data culling, disclosure, and production, where completeness is critical.

Combining technology and expertise

The most defensible workflows combine the speed and contextual understanding of AI with the transparency and completeness of established eDiscovery techniques. As the toolset expands and workflows become more sophisticated, experienced eDiscovery teams play an essential role in selecting the right tools, validating outcomes, and ensuring workflows remain proportionate, defensible, and aligned to the objectives of each matter.

About A&O Shearman’s eDiscovery capability

The A&O Shearman eDiscovery team is one of the most highly certified and experienced in the industry, with nine Relativity Masters - the highest level of accreditation available. Our expertise is grounded in more than a decade of hands-on use of Relativity. During that time, we have worked closely with Relativity to refine workflows, provide structured product feedback, and support the ongoing development of the sophisticated review functionalities needed for our clients’ most complex matters.

This longstanding relationship gives us early access to the latest AI capabilities, enabling us to introduce new features to client matters as soon as they are released. By pairing these innovations with rigorous quality controls and the judgment of seasoned specialists, we help clients surface key facts quickly, manage exponential data growth effectively, and navigate investigative challenges with confidence and strategic focus.

Our teams of lawyers and regulatory advisers work closely with our eDiscovery team to ensure that our clients have the benefit of leading AI tools deployed with our leading expertise on how best they can be used to navigate the most complex and impactful disputes, litigation and investigations matters.

To learn more about our eDiscovery services, please contact Christina Zachariasen, Daniel Smith, or Natalie Lau.