How does AI eDiscovery work for legal documents?

AI eDiscovery is the use of machine learning and, increasingly, large language models to identify, collect, process, review, and produce electronically stored information (ESI) in litigation and investigations. Instead of armies of contract attorneys reading documents one at a time, AI systems classify millions of emails, chats, and files in hours, surface the responsive subset, flag privileged material, and even draft summaries. Understanding how the technology actually works — and where it fails — matters because courts increasingly expect parties to use it, while sanctions for careless production remain a real risk.

The Direct Answer: What AI eDiscovery Actually Does

Also worth reading: What is a GenAI eDiscovery validation protocol and how should legal teams build one in 2026? · How is AI legal education reform expected to evolve by 2027, and what does this mean for eDiscovery and document drafting? · How does the EU AI Act classify high-risk legal software and what certification is required for eDiscovery tools?

At its core, AI eDiscovery automates the document review stage of the Electronic Discovery Reference Model (EDRM), the nine-phase framework that governs how ESI moves from identification through collection, processing, review, analysis, and production. Traditional linear review required human reviewers to open every document; at typical contract-attorney rates of $30–$60 per hour and review speeds of roughly 50–100 documents per hour per reviewer, a 500,000-document matter could consume thousands of billable hours before anyone drafted a single motion.

AI changes the economics by predicting which documents are relevant, privileged, or hot before humans read them. Modern platforms apply several techniques simultaneously: supervised machine learning (often called predictive coding or Technology-Assisted Review, TAR), unsupervised clustering that groups similar documents together, email threading that collapses entire conversation chains into single review items, near-duplicate detection that eliminates redundant copies, and — since roughly 2023 — generative AI that can summarize document sets, answer natural-language questions about the corpus, and draft privilege logs. Vendors such as Harvey, Thomson Reuters, OpenText (with its eDiscovery Aviator agents), and Relativity have all shipped generative features aimed squarely at this workflow.

The practical result is that first-pass review populations shrink dramatically. In well-executed TAR workflows, teams frequently review only 5–15% of a collected corpus while still capturing 75–95% of responsive documents, depending on the richness of the data and the quality of training. That reduction is where the cost savings live — but it is also where the risk lives, because every document the model deprioritizes is a document a court may later ask why you did not produce.

How the Technology Works Under the Hood

Supervised TAR follows a training loop. A senior attorney reviews a seed set of perhaps 500–2,000 documents, coding each as responsive or not responsive. The algorithm learns which words, phrases, senders, metadata patterns, and semantic features correlate with responsiveness, then ranks the remaining corpus from most to least likely to be responsive. Reviewers code additional batches, the model retrains after each round, and a statistical measure of separation between the two populations tells the team when the model has stabilized. Once stability is reached, counsel can either stop reviewing at a cutoff (a process some courts have scrutinized) or continue until recall targets are met.

Generative AI adds a different layer. Rather than classifying documents into buckets, large language models read text and generate output: a two-paragraph summary of a 90-page deposition, an answer to "which executives discussed the pricing change after March 2022?", or a draft privilege log entry with author, recipients, date, and general subject matter. These models handle context far better than keyword search — they understand that "the Q3 numbers we discussed" refers to financial projections without needing a keyword list anticipating that phrasing. This is why legal AI adoption accelerated so sharply after 2023: keyword search misses meaning, and LLMs approximate meaning at scale.

Email threading deserves mention because it delivers savings independent of any model training. Threading algorithms identify which messages in a chain contain unique content and suppress those that are merely quoted forwardings, often cutting review volume by 20–40% on email-heavy corpora before any predictive modeling begins. Near-dupe detection performs similar work on attachments and shared files.

The Typical Workflow, Step by Step

A realistic AI-assisted eDiscovery project runs through six stages. First, legal hold and identification: counsel issues preservation notices and maps custodians, data sources, and date ranges. Second, collection: forensic tools pull ESI from mail servers, cloud drives, phones, and collaboration platforms like Slack or Teams, preserving metadata. Third, processing: the platform de-duplicates (often removing 50–70% of raw volume), extracts text from PDFs and Office files, applies threading, and loads everything into a review database.

Fourth, early case assessment: analytics dashboards show who emailed whom, which terms appear where, and what the data looks like before serious money is spent. Fifth, AI-assisted review: this is where TAR training, generative summarization, and privilege prediction happen, with human reviewers validating model outputs on prioritized batches. Sixth, production: responsive documents are converted to the agreed format (typically TIFF or native with extracted text), endorsed with Bates numbers, logged, and delivered, usually under Federal Rule of Civil Procedure 26(b)(5) privilege assertions.

Timing varies enormously by matter size. A small matter of 10,000–50,000 documents can move from collection to production in two to four weeks with AI assistance. A multi-million-document antitrust or product-liability matter typically runs three to nine months, with AI compressing what would have been a year-plus of linear review. Courts' scheduling orders rarely accommodate delays caused by inefficient review choices, which is precisely why judges have grown receptive to TAR — it lets parties meet deadlines that manual review would blow through.

Comparing Your Options: TAR, Generative AI, Keywords, and Manual Review

No single approach dominates every scenario, and honest practitioners will tell you the choice depends on corpus size, budget, deadline pressure, and how much scrutiny the production will face. The table below summarizes the trade-offs:

Keyword search remains the default in many cases because it is cheap and easy to defend — you can show a judge exactly why each document was pulled. Its weakness is well documented: studies going back to the TREC Legal Track evaluations in the late 2000s showed keyword recall often below 50% against known relevant sets, because humans cannot anticipate every synonym, typo, or coded reference. TAR was developed specifically to fix that gap, and courts including Judge Andrew Peck (in Da Silva Moore v. Publicis Groupe, 2012) endorsed its use early, provided parties follow transparent validation protocols.

Generative AI is the newest entrant and the least settled. It excels at comprehension tasks — summarization, question answering, drafting — but introduces hallucination risk that classification-based TAR does not. The emerging best practice treats generative output as a draft requiring human verification, never as a final work product. Law.com's reporting on generative AI in eDiscovery has emphasized exactly this ground-truth problem: impressive demos, uneven real-world performance, and a profession still building the validation standards that TAR took a decade to develop.

Where AI eDiscovery Goes Wrong: Common Mistakes

The most expensive mistake is treating AI output as self-validating. A model that achieves 85% recall sounds excellent until you realize the missing 15% included the smoking-gun email. Defensible workflows require documented validation: control sets, statistical sampling of the discard pile, and recall estimates a court can inspect. Teams that skip this step discover their vulnerability mid-deposition, when opposing counsel produces a document your model scored irrelevant.

Second, garbage-in problems persist regardless of model sophistication. If collection missed a custodian's personal email account or a Slack workspace, no algorithm recovers documents that were never collected. AI optimizes review, not preservation, and spoliation sanctions under Rule 37(e) attach at the preservation stage. Third, over-reliance on privilege prediction is dangerous: privilege waivers are nearly impossible to claw back under FRE 502 once produced, and automated privilege calls should always sit behind human confirmation, ideally with quick peek workflows and clawback agreements in place.

Fourth, many teams buy generative tools and point them at confidential client data without confirming vendor data-handling terms — whether inputs train future models, where data resides, and what happens on subpoena. Fifth, transparency failures create discovery-about-discovery disputes. If you use TAR, opposing counsel may demand details of your methodology; parties who cannot articulate their process invite motions to compel and credibility damage. Finally, a subtle error is using AI to cut corners on proportionality analysis rather than to enable it — Rule 26(b)(1) requires relevance weighted against burden, and AI gives you better data for that argument than ever before, but only if someone actually makes it.

Costs, Pricing Models, and Budget Reality

Pricing structures vary more than most buyers expect. Legacy review platforms typically charge per gigabyte processed ($15–$40/GB), per user per month for review seats ($150–$400), and hosting fees ($5–$25/GB/month). Generative AI features are usually priced separately — per query, per document summarized, or as a platform add-on tier. Harvey and similar enterprise legal AI products sell annual subscriptions commonly reported in the tens of thousands to low hundreds of thousands of dollars per firm, depending on seat count and usage caps.

For a concrete illustration: a 250,000-document matter might cost $400,000–$800,000 in pure linear review labor, versus $80,000–$200,000 with a validated TAR workflow plus targeted human review of the top-ranked population. Those figures assume competent execution; a botched TAR run that triggers re-review can erase the savings entirely. Small matters under 20,000 documents often do not justify heavy AI tooling at all — flat-fee managed review or a lightweight SaaS platform with built-in analytics may beat an enterprise deployment on both cost and speed.

Budget discipline also means watching scope creep in processing. Over-collection is endemic: broad custodian lists and generous date ranges routinely inflate corpora by multiples, and every unnecessary gigabyte carries hosting and review costs downstream. Early case assessment analytics exist precisely to let you narrow scope before the expensive stages.

When to Act and How Courts View This in 2026

If you are facing a pending or reasonably anticipated dispute, preservation obligations begin immediately — before any AI tooling decisions. The right sequence is: preserve broadly, collect defensibly, then deploy AI to narrow review. Waiting until a scheduling conference to consider TAR wastes leverage; raising it early, ideally in a meet-and-confer under FRCP 26(f), lets you negotiate disclosure protocols and avoid fights later.

By 2026, judicial attitudes have shifted from novelty to expectation. Judges who once required detailed TAR protocols now more commonly ask why parties with massive corpora are not using available technology, and several districts encourage or presume its use in large cases. At the same time, bar regulators and malpractice carriers have sharpened guidance on generative AI: attorneys remain responsible for verifying citations, facts, and productions regardless of what a model generated. The American Bar Association's formal ethics guidance on generative AI, issued in 2024, stressed competence, confidentiality, and supervision — duties that map directly onto eDiscovery practice.

The strategic takeaway is that AI eDiscovery is now table stakes for large matters, a genuine advantage for mid-size ones, and optional overhead for small ones. The winners are not the teams with the flashiest tools but the ones pairing capable technology with rigorous validation, honest documentation, and lawyers who actually read what the machines prioritize. Used that way, AI cuts review time by half or more and surfaces evidence humans would miss. Used carelessly, it produces confident errors at scale — and courts have little patience for either excuse.

Practical Steps to Get Started

Begin with a data audit: inventory where your organization's ESI lives, who controls it, and how quickly it can be preserved under legal hold. Then pilot AI review on a closed matter or internal investigation before trusting it in bet-the-company litigation, measuring recall against a manually reviewed sample. Negotiate vendor contracts with explicit terms on data confidentiality, model training opt-outs, and audit rights. Train your team on validation statistics — precision, recall, elusion testing — so you can defend methodology under oath. Finally, build the human layer deliberately: assign experienced reviewers to validate model-flagged batches, keep privilege determinations human, and document every workflow decision contemporaneously. Organizations that treat AI as a force multiplier for judgment, rather than a replacement for it, consistently get the cost savings without the sanctions exposure.