eDiscovery, short for electronic discovery, is the process by which parties to a lawsuit, investigation or regulatory demand find, preserve, collect, review and hand over electronic information as evidence. It exists because most business records now live in email, chat, shared drives, phones and, increasingly, AI chatbot logs rather than filing cabinets. Courts in the United States allow each side to demand relevant material from the other before trial, and eDiscovery is the machinery that turns that legal right into a working process across billions of digital files.

ComplexDiscovery, an analysis firm that aggregates market estimates, put worldwide eDiscovery spending at $19.61 billion in 2025 and projects $28.08 billion by 2030, according to its 2025-2030 market estimate published in May 2026. Services still account for about two-thirds of that, though software's share is forecast to rise from 34% to 39%.

How the process works

The industry describes the work using the Electronic Discovery Reference Model (EDRM), a framework created in 2005 by consultants George Socha and Tom Gelbmann. According to TechTarget's summary, it has nine stages: information governance, identification, preservation, collection, processing, review, analysis, production and presentation. The model is not linear: teams loop back as new data sources surface.

Identification and preservation come first. Once litigation is "reasonably anticipated", an organisation must stop routine deletion of relevant data. It does this through a legal hold (also called a litigation hold): a notice to named employees, known as custodians, plus technical locks inside email, chat and file systems that override automatic deletion. In Microsoft 365, holds are applied through the eDiscovery tools in Microsoft Purview. In Google Workspace, the equivalent tool is Google Vault.

Collection pulls the preserved data out of its source system. Each file gets a hash value, typically MD5 or SHA-1, proving the copy is unaltered.

Processing turns raw data into something reviewable. Software extracts text and metadata (sender, recipients, timestamps, file paths), strips known system files by comparing hashes against a National Institute of Standards and Technology list ("de-NISTing"), removes duplicates and threads email replies.

Review is where most of the money goes. Lawyers, or software they supervise, tag documents as relevant or not and flag material protected by attorney-client privilege or the work product doctrine. Withheld documents are listed on a privilege log, which US Federal Rule of Civil Procedure 26(b)(5) requires so the other side can challenge the claims. Federal Rule of Evidence 502, added in 2008, lets parties agree on "clawback" orders so that privileged documents disclosed by accident can be retrieved without waiving privilege.

Production is delivery. Documents are usually stamped with sequential Bates numbers and handed over as images with extracted text and a load file, a structured index that lets the receiving party's software rebuild each document and its metadata, or as native files such as spreadsheets. Format and search terms are negotiated at a "meet and confer" required by Rule 26(f).

Origins and evolution

Discovery of paper records is centuries old. The electronic version became a distinct discipline in the early 2000s, driven by email volume and a series of rulings in Zubulake v. UBS Warburg in the Southern District of New York. Judge Shira Scheindlin's opinions between May 2003 and July 2004 set out who pays to restore backup tapes, when the duty to preserve begins and what lawyers must do to make a hold effective. The case ended in an April 2005 verdict of about $29.3 million, after jurors were told they could infer that deleted emails would have hurt UBS.

Rule changes followed. Amendments to the Federal Rules of Civil Procedure that took effect on December 1, 2006 introduced the term "electronically stored information" (ESI), required early discussion of ESI at the Rule 26(f) conference and allowed parties to resist data that was "not reasonably accessible". A second round, effective December 1, 2015, tied the scope of discovery to proportionality and rewrote Rule 37(e): the harshest sanctions for lost ESI, such as an instruction that the jury may presume the data was unfavourable, now require a finding of "intent to deprive". The Sedona Conference, a legal think tank, published its widely cited Sedona Principles in 2004, with a third edition in 2018.

Technology moved in parallel. On February 24, 2012, Magistrate Judge Andrew Peck approved computer-assisted review in Da Silva Moore v. Publicis Groupe, writing that it "now can be considered judicially-approved for use in appropriate cases", according to a contemporaneous summary on law firm K&L Gates' eDiscovery blog. That technique, known as technology-assisted review (TAR) or predictive coding, trains a model on a lawyer's coding decisions and ranks the remaining documents by likely relevance. The English High Court accepted it in Pyrrho Investments v MWB Property in 2016.

Platform vendors absorbed the function. Microsoft bought predictive coding company Equivio in January 2015 and built it into what became Microsoft Purview eDiscovery (Premium). Microsoft retired its "classic" eDiscovery tools between August 1 and September 1, 2025 in favour of a unified interface, according to message center notice MC1002399.

Generative AI is the latest layer. Relativity, a leading review platform, made custom AI analyses in aiR for Review generally available on June 16, 2026.

Why it matters for marketing and ad tech

The antitrust cases reshaping digital advertising have been built on eDiscovery. Internal Google emails from 2015, produced in discovery and made public in August 2024 in the Justice Department's ad tech case, showed staff discussing the share of AdX auctions won by the Google Display Network. Exhibit lists for the ad tech remedies trial before Judge Leonie Brinkema covered more than 870 documents spanning 2008 to 2025.

Preservation failures have become part of the story. Google's chat system defaulted to "history off" from 2008, deleting messages after 24 hours. Texas and other states sought sanctions in January 2025, estimating 2.8 million to 4 million deleted chat messages per year and noting that 61 of 202 custodians were not placed under holds until 2022 or later. Advocacy groups filed a California State Bar complaint against Google's chief legal officer Kent Walker in October 2024, citing criticism from three federal judges. Judge James Donato, who had found in March 2023 that Google failed to preserve chats in the Play Store litigation, called it "the most serious and disturbing evidence I have seen in my decade on the bench". In her April 17, 2025 liability ruling, Judge Brinkema suggested the conduct could be sanctionable but found sanctions unnecessary because Google lost on the admissible evidence, according to a client memo by law firm Simpson Thacher.

AI products have turned marketing and consumer conversations into evidence. Magistrate Judge Ona Wang ordered OpenAI to produce 20 million de-identified ChatGPT logs to The New York Times and other publishers in November 2025. Those logs later produced one of the few hard numbers in the copyright fight: an OpenAI expert found 24 instances of verbatim regurgitation, or 0.00012%. Campaign chats, CRM notes and AI prompts are all potential ESI once a dispute is foreseeable.

Limitations and disputes

Cost is the oldest complaint. Review remains labour-intensive, and the 2015 proportionality rule was written because discovery bills could exceed the amount in dispute. TAR and generative AI promise savings, but parties still fight over validation, including how to measure recall, the share of relevant documents actually found.

Privilege and AI is unsettled. In United States v. Heppner, Judge Jed Rakoff ruled from the bench on February 10, 2026 that about 31 documents a defendant generated with Anthropic's Claude before his arrest were protected by neither privilege nor work product, after the Justice Department argued that conversations with Claude were not privileged. The same day, Magistrate Judge Anthony Patti in Michigan held in Warner v. Gilbarco that a self-represented plaintiff's ChatGPT prompts were protected work product, calling AI programs "tools, not persons", according to an analysis by law firm Wilson Elser. The two rulings point in different directions, and no appeals court has resolved the conflict.

Coverage gaps persist. Ephemeral messaging and AI assistants often launch before hold and export tools catch up. Platform dependence is another weakness: when a cloud provider fails, so do searches and exports.

  • eDisclosure is the term used in England and Wales. In the Business and Property Courts it is governed by Practice Direction 57AD, not US federal rules.
  • Digital forensics recovers and analyses data, including deleted files, often for criminal or internal investigations. It can feed eDiscovery but excludes legal review and production.
  • Information governance and records retention decide what an organisation keeps and deletes in normal operation. A legal hold suspends those rules for specific data. A retention policy is not a hold.
  • Discovery ads, Google's former ad format later folded into Demand Gen, share the word but have no connection to legal discovery.

Recent developments

Google has been extending Vault to its AI products. On June 11, 2026 it added retention rules and litigation holds for Gemini app conversations, and on October 22, 2026 it added filtered eDiscovery exports of Gemini conversations inside Gmail, Drive, Docs, Sheets, Slides and Meet, according to the Google Workspace Updates blog. Holds for Gemini inside Workspace apps remain a future release.

Today, a Microsoft 365 outage tied to Exchange Online disrupted 19 services from Outlook to Copilot, and Microsoft's status page said Purview eDiscovery search and export might also be affected. The same Purview suite also runs data loss prevention controls, which govern what leaves an organisation while eDiscovery governs what must be kept.

AI errors are also piling up in court. A tracker kept by researcher Damien Charlotin listed 2,266 cases involving AI-hallucinated content as of early November 2026, according to eDiscovery Today.

Timeline

  • 2003-2004: Judge Shira Scheindlin issues the Zubulake v. UBS Warburg opinions (May 13, 2003 to July 20, 2004) on cost-shifting, the duty to preserve and legal holds.
  • 2004: The Sedona Conference publishes the first edition of the Sedona Principles on electronic document production.
  • April 2005: Zubulake jury returns a verdict of about $29.3 million.
  • 2005: George Socha and Tom Gelbmann create the Electronic Discovery Reference Model.
  • December 1, 2006: Federal Rules of Civil Procedure amendments introduce "electronically stored information".
  • 2008: Federal Rule of Evidence 502 adds clawback protection for inadvertently produced privileged material. Google sets chat "history off" as default the same year, according to later court filings.
  • February 24, 2012: Da Silva Moore v. Publicis Groupe approves computer-assisted review.
  • January 2015: Microsoft acquires Equivio, the basis for its advanced eDiscovery product.
  • December 1, 2015: Federal rule amendments add proportionality and rewrite Rule 37(e) on lost ESI.
  • 2016: Pyrrho Investments v MWB Property approves technology-assisted review in England.
  • 2018: Third edition of the Sedona Principles.
  • March 28, 2023: Judge James Donato finds Google failed to preserve chats in the Play Store antitrust litigation.
  • April 17, 2025: Judge Brinkema's ad tech liability ruling notes Google's chat practices but declines sanctions.
  • August 1 to September 1, 2025: Microsoft retires classic Purview eDiscovery in favour of a unified experience.
  • November 7, 2025: Magistrate Judge Ona Wang orders OpenAI to produce 20 million ChatGPT logs.
  • February 10, 2026: Heppner and Warner v. Gilbarco reach opposite conclusions on protection for AI-generated material.
  • June 11, 2026: Google Vault adds retention rules and litigation holds for the Gemini app.
  • June 16, 2026: Relativity makes aiR for Review custom analyses generally available.
  • October 22, 2026: Google Vault adds filtered exports for Gemini in Workspace.
  • November 13, 2026: Microsoft 365 outage puts Purview eDiscovery search and export among affected functions.

Summary

Who: Litigants, their outside counsel and in-house legal teams, regulators, review vendors such as Relativity, and the platforms that host the data, notably Microsoft (Purview) and Google (Vault). Judges supervise disputes over scope, cost and lost data.

What: The process of identifying, preserving, collecting, processing, reviewing and producing electronically stored information as evidence, increasingly including chat messages and AI conversations.

When: The discipline took shape between 2003 and 2006 through the Zubulake rulings and the December 2006 US federal rule changes, was reshaped by the 2015 proportionality amendments, and is now adapting to generative AI in 2026.

Where: Primarily US federal and state courts, with parallel regimes such as eDisclosure under Practice Direction 57AD in England and Wales, and across cloud systems wherever business data is stored.

Why: Because almost all business evidence is now digital, and failures to preserve it, as in Google's chat deletions, can shape the outcome of cases, including the antitrust and copyright cases redefining advertising and AI.