eDiscovery AI

When someone says "AI in eDiscovery" today, they could mean almost anything. A chatbot a paralegal opens in a browser. A predictive coding model that ranks documents by probability of relevance. A large language model summarizing a single PDF on demand. They are all real, and they are all different things.

For most of the last decade, the eDiscovery industry rallied around a narrow definition - technology-assisted review built on supervised classifiers, trained on a seed set, refined through rounds of validation. That work is still valuable. But the shape of what AI can do inside a review has changed, and the language we use to describe it has not caught up.

This post is an attempt to name the umbrella. What does eDiscovery AI actually cover in 2026, what kinds of work does it do inside a case, and where does Claira fit?

What eDiscovery AI means now

eDiscovery AI is generative AI applied to the artifacts of litigation - documents, emails, attachments, transcripts, images - in service of the review and analysis phase. It is the layer between collection and production, where decisions get made about every document in the corpus.

Older review tech relied on patterns: keyword lists, regex, TAR seed sets, near-duplicate clustering. Those tools are still in the stack and still useful. The shift with generative AI is that the model reads each document the way a reviewer would, in context, with judgment, against your criteria, and produces an output you can defend in writing.

The work splits into a few recognizable categories. They are not separate products. They are different things you ask the same review engine to do.

First-pass review at corpus scale

The first thing eDiscovery AI does is read every document in the set and form a position on it. Not a relevance score from a classifier. A reasoned answer to your review question, with a justification you can point to.

This is the part that used to be impossible at scale. You either threw bodies at it, or you trained a model on a sample and trusted the extrapolation. Both have known failure modes. The documents you most need to find are often the ones a sample is least likely to surface.

With Claira, your reviewer writes the question, Claira reads every document inside Nuix Discover, and you get a coded decision plus a written rationale for each one. Tens of thousands of documents an hour. The reviewer's job becomes verifying the close calls and confirming the obvious, not doing every read from scratch.

Objective coding

The second category is the structured-fact work that has historically been the most expensive line item on a budget. Author, recipients, date, document type, language, attachment relationships, redaction candidates. The kind of fields that sit in the metadata pane, and that every downstream workflow depends on.

Done by humans, objective coding is slow and inconsistent. Done by a classifier, it covers only what the classifier was trained on. Done by an instruction-following model with the document in front of it, it works the way a careful paralegal would, on every document, without getting tired. The output writes straight back into Nuix Discover fields, ready to be filtered, sorted, and reported on like any other coded data.

Reading for substance: privilege, PII, summaries, chronologies

A third bucket is the work that requires reading for substance rather than form. Flagging documents that may be privileged. Pulling personal information that needs to be addressed before production. Generating short summaries reviewers can use to triage. Building dated timelines that turn a folder of emails into a narrative.

This is where eDiscovery AI starts to feel different from the previous generation of tools. The model is not classifying against a fixed schema. It is responding to a question - your question - on a document. The same review engine that produces a relevance call can produce a privilege flag, a PII extract, a one-paragraph summary, and a chronology entry, depending on what you asked for.

You can see the full map of what falls under that umbrella in Claira's workflow overview.

Case context

The thing that ties all of this together is matter-specific context. A model that reads documents without knowing the case is doing search. A model that reads documents with your case theory, the parties, the issues, the time periods, and the language you care about is doing review.

Claira treats that context as a first-class object on the case. You write it once, you keep refining it, and every scan you run inherits it. That is how you keep tens of thousands of decisions consistent without re-explaining the case to the model each time.

It is also how you keep the work defensible. The criteria are written down. The decisions reference them. The audit trail follows.

What it does not replace

eDiscovery AI is not a substitute for the parts of review that require professional judgment. It does not decide what is privileged in a contested situation. It does not make the call on a borderline trade secret. It does not sign off on a production. The lawyer who is responsible for the file is still the lawyer responsible for the file.

What it does is the high-volume, criterion-driven work that used to consume the bulk of the review hours, and it puts the close calls in front of the people who should be making them. The lawyer's job stays the same. The lawyer's day looks different.

For a fuller walkthrough of where AI fits across the broader eDiscovery workflow, from collection through production, our practical guide on how AI fits into modern eDiscovery is a good companion to this piece.

Where to start

If you are evaluating eDiscovery AI for a case that is already in review, the question is not which category to pick. The categories work together. Most teams start with first-pass review on a defined slice, validate the output against what their reviewers would have done, and then expand to objective coding and the substance-oriented workflows once they trust the engine.

The easiest way to see what that looks like on your data is to walk through it with someone. Book a working session with us and we will set it up against a realistic sample from a real case.

That is the umbrella. Generative review that reads every document in context, codes it, justifies the call, and gives your team back the hours they used to lose to the first pass.