FTC instructions set 75 percent recall and precision floors for TAR and AI review, law firm reports

Oct 31, 2026

Editor’s Note: The Federal Trade Commission has written 75 percent recall and precision floors into instructions issued with a recent civil investigative demand and a merger second request, for parties using technology-assisted review (TAR) or artificial intelligence-assisted review, according to an Oct. 7 analysis by Bryan M. Marra of the law firm Arnold & Porter.

Recall is the share of responsive documents a review finds; precision is the share of documents marked responsive that truly are. The reported instructions also set a minimum cutoff score of 0.50, the score at which a document counts as responsive, and parties using AI-assisted review must disclose the tool, its underlying model and all prompts given to it. ComplexDiscovery located no FTC document describing the instructions, and the agency’s public Model Second Request, its template for merger information requests, revised in January 2024, asks for recall and precision statistics without setting a minimum.

For cybersecurity, information governance and eDiscovery teams, the terms reach collection timing, platform capabilities and prompt records that some parties argue are protected attorney work product. Watch for published agency language, negotiated modifications and any court test of whether the instructions can be enforced.

Disclosure: Rob Robinson, founder and managing editor of ComplexDiscovery OÜ, also serves as chief marketing officer of HaystackID, a provider of eDiscovery services, including AI-assisted document review. HaystackID, which published its own account of the FTC instructions on Oct. 13, 2026, had no role in the preparation of this article, which does not rely on that account and is based solely on publicly available information. ComplexDiscovery OÜ is an EDRM Guardian Partner, and Robinson sits on EDRM’s Global Advisory Council, where EDRM lists Tara Emory and Maura Grossman, whose paper the article cites.

Content Assessment: FTC instructions set 75 percent recall and precision floors for TAR and AI review, law firm reports

Information - 92%

Insight - 92%

Relevance - 94%

Objectivity - 91%

Authority - 90%

92%

Excellent

A short percentage-based assessment of the qualitative benefit expressed as a percentage of positive reception of the recent article from ComplexDiscovery OÜ titled, "FTC instructions set 75 percent recall and precision floors for TAR and AI review, law firm reports."

News Analysis – eDiscovery Beat

FTC instructions set 75 percent recall and precision floors for TAR and AI review, law firm reports

ComplexDiscovery OÜ Staff

The Federal Trade Commission has written 75 percent recall and precision floors for technology-assisted and artificial intelligence-assisted review into a recent civil investigative demand and a merger second request, a law firm analysis reports.

The analysis names no FTC document, and the model second request the agency posts publicly still asks parties to supply recall and precision statistics without setting a number they must reach.

Bryan M. Marra, a senior attorney in Arnold & Porter’s Washington office, described the instructions in an Oct. 7 post on the firm’s eData Edge blog, writing from the side of parties that respond to compulsory process. He wrote that the FTC had issued them as investigational subpoenas in a civil investigative demand (CID) and in a merger second request investigation. In his account, they tighten the conditions for technology-assisted review (TAR) and give AI-assisted review its first definition. The post does not date either the CID or the second request, name either investigation or quote the instruction text at length. Every published account of the specific thresholds that ComplexDiscovery located credits Marra’s post, including republications on Mondaq and The Legal 500. That means every figure attributed to the instructions below rests on a single account.

Reported terms run from recall to prompts

Recall measures the share of responsive documents a review actually finds, while precision measures the share of documents marked responsive that truly are. Recall must reach at least 75 percent, measured to plus or minus 3 percent at a 95 percent confidence level, according to Marra. Precision must also reach at least 75 percent, a requirement he described as “the first time the FTC has required a set precision” and a break from agency practice that had centered on recall. Parties must disclose the cutoff score that separates responsive from nonresponsive documents, which he said is likewise new, and that score can be no lower than 0.50.

Those numbers come wrapped in a schedule. Disclosures describing the technology, the workflow and the subject matter experts who will train the system are due at least one week before any training or review starts. Each expert must be “a full-time employee of the firm managing the response,” Marra wrote. The same filing must report which kinds of documents will go through TAR or AI and which file types are set aside for special handling, with particular rules for ephemeral messages, text messages and linked, or modern, attachments.

Training cannot begin until every custodian has been collected and FTC staff have had the opportunity to review and confirm the disclosures. In Marra’s reading, the sequencing “seems aimed at discouraging TAR 1.0 in favor of TAR 2.0/Continuous Active Learning.” A calendar of planned rolling productions is due within a week after training starts, and a pre-production elusion sample, delivered as the sample itself rather than as summary statistics, is due at least two weeks before compliance.

Validation also carries consequences under the reported terms. Parties typically offer the FTC a look at a null set when using TAR, Marra said, and the instructions now formalize that practice. A party must retrain the model when the FTC concludes that the validation shows deficiencies, and it must make supplemental productions if the agency’s review of the sample changes the process.

Generative and other AI tools face a further list. A party must name the tool, version, provider and underlying model, and report how often the AI and human reviewers agreed on a statistically valid sample, broken down by classification category. It must disclose all prompts or review protocols given to the AI, along with relevance scores and rationales for every produced document where those exist.

Two limits reach both technologies, by Marra’s account. No analytics, email threading and search terms among them, may shrink the population before deduplication or before TAR or AI runs unless the FTC agrees to the specific process. Nor may manual review or search terms strip out documents identified as responsive by predictive coding, “except to the extent required to identify privileged information, including PII, SPII, and SHI.” Marra said that rule formalizes a position the FTC and the Justice Department have historically taken. The post does not expand those acronyms, although the agency’s model second request separately defines “Sensitive Personally Identifiable Information” and “Sensitive Health Information.”

FTC’s public template still sets no number

The agency’s own posted template asks for less. The FTC’s Hart-Scott-Rodino (HSR) Resources page links a Model Second Request marked “Revised January 2024.” Its Instruction I 5 asks a party using any form of technology-assisted review to confirm that subject matter experts will review the seed set and training rounds and to supply “recall, precision, and confidence-level statistics (or an equivalent).” The same instruction calls for a validation process that lets Commission staff review statistical samples of documents the algorithm marked nonresponsive, and it sets no minimum for any statistic. A text search of the 27-page document on Oct. 30 found no reference to artificial intelligence.

That model is a template, while Marra described instructions attached to particular investigations. Until the agency publishes the language or revises the template, the public description of the floors that ComplexDiscovery located traces to a law firm’s summary of instructions for which it found no public copy.

Why a precision floor may be hard to plan for

Meeting the floors may prove harder than stating them. Marra warned that the thresholds could prove difficult to satisfy, since precision turns heavily on the particulars and richness of a given collection and tends to move against recall. Tuning a system to find more responsive documents usually sweeps in more nonresponsive ones, so a party asked to hold both measures at 75 percent has less room to give up one to gain the other.

The reported margin-of-error language leaves a second question open. If 75 percent is the required point estimate, a measured recall of 75 percent passes. If it is instead the bottom of the confidence interval, a party needs a point estimate near 78 percent, on this article’s reading of the margin as 3 percentage points. Marra’s account does not say whether the 75 percent requirement applies to the point estimate or the lower confidence bound.

Sample size compounds the uncertainty. Under the textbook normal approximation, a ComplexDiscovery calculation shows, estimating recall near 75 percent within 3 percentage points at 95 percent confidence takes about 800 responsive documents in the validation sample. In a collection where 5 percent of documents are responsive, finding that many by simple random sampling means drawing about 16,000 documents on average. Other estimation designs exist, and the instructions as reported do not prescribe one.

The cutoff floor raises a parallel question for generative tools, because the account does not say what scale the 0.50 figure sits on. A stipulated order entered in July in the Northern District of California already anticipates AI systems that generate “scores, probabilities, classifications, or decision boundaries.” A tool that returns a category rather than a score, in this article’s reading, offers no evident 0.50 to report. Teams expecting a second request should ask their platform providers now whether a tool can report agreement by category, export per-document rationales and state precision at a defined cutoff, the outputs the reported instructions list.

Two court orders stop short of recall and precision floors

Two federal orders on AI review this summer stopped short of the reported floors, for different reasons. In Schulte v. LinkedIn Corp., a putative antitrust class action in the Northern District of California, Magistrate Judge Laurel Beeler declined on June 30 to compel LinkedIn to disclose further metrics on its use of Relativity aiR. The plaintiffs had sought elusion estimates, a document error rate and a count of the people validating the generative AI product’s predictions.

The court found that the interim electronically stored information (ESI) order required only disclosure that LinkedIn would use technology-assisted review. A 204,444-document target population, without a more specific showing of deficiency, did not justify discovery on discovery. LinkedIn had told the plaintiffs that it used no seed or training set and that aiR was making final responsiveness calls, with human quality-control review of samples.

A week later, Magistrate Judge Robert M. Illman entered a stipulated ESI order in James v. Cerebras Systems Inc., a putative class action in the same district, whose AI appendix tracks several of the FTC’s reported disclosures. It requires each model’s identity, version and hosting environment, disclosure of all prompts and instruction sets, and redlines of prompt changes within three business days. Validation results must report recall and precision, and the sample of documents the AI excluded as nonresponsive must be sized at a 95 percent confidence level with a 2 percent margin of error. Its only performance expectation is that elusion not exceed 3 percent, and it sets no recall or precision floor. Proprietary prompts may be withheld if logged with particularity and offered for attorneys’-eyes-only inspection.

Prompt disclosure meets an unsettled work-product question

That trade-secret carve-out points to a sharper edge in the reported FTC terms, which, as Marra describes them, call for all prompts with no stated exception for proprietary material. Marra said some parties have argued that prompts, scores and rationales are attorney work product because they reveal legal strategy. A paper by Tara Emory and Maura Grossman, which ComplexDiscovery covered in July as forthcoming in the Columbia Science and Technology Law Review, argued that properly developed review prompts can reflect an attorney’s “factual investigation, mental impressions, and case strategy.” In most cases, the authors wrote, validating the results serves transparency better than exposing the prompts.

Outside document review, one federal court this month declined to treat AI material as categorically protected. On Oct. 21, Magistrate Judge Marty Fulgueira Elfenbein of the Southern District of Florida ruled in Fontainebleau Florida Hotel, LLC v. Botach, a dispute over discovery of a defamation defendant’s AI communications. As to Request Nos. 1, 3, 4 and 7, the order turned away his blanket claim of work-product protection. The order left room for document-specific claims that name the withheld material and explain the protection claimed. It read United States v. Heppner, a Southern District of New York decision, as denying protection to AI material a represented defendant generated without counsel’s direction.

The same order read Warner v. Gilbarco, Inc., an Eastern District of Michigan decision, as accepting that AI-assisted preparation for litigation can qualify for protection and that working through an outside AI tool need not waive it. None of the decisions described here involved review prompts demanded by a federal agency during an investigation.

Negotiate, comply or explain

Marra’s post lays out the practical routes in the meantime. Parties may go back to search terms and manual review, particularly where the review population is small; negotiate compromises with FTC staff; or proceed under a protocol they consider reasonable while giving a statement of reasons for noncompliance. He said the last route carries risk in a second request, where the FTC may dispute that a party has substantially complied and so delay the closing of the transaction. He added that a federal court’s willingness to order compliance is uncertain where a party has followed a reasonable protocol for search and review.

For counsel who expect a second request or CID, the reported terms argue for preparation before either one arrives. Confirm that the review platform can produce the metrics and artifacts the instructions list, and plan for training to wait on complete custodial collection. Keep prompt revision histories and rationale exports in a form that can be logged or produced, and identify in advance the full-time employees who would serve as subject matter experts.

The FTC’s posted template asks a party to show its recall and precision, and if Marra’s account is accurate, the agency has started telling some recipients what those numbers must be. ComplexDiscovery found no public copy of those instructions, and a party that does not follow them as written can negotiate, fall back to search terms or explain itself, which in a second request risks a compliance dispute. When the next second request arrives, will a team’s review platform be able to show 75 percent on both measures for the collection in front of it, and will counsel already know which prompts it is prepared to hand over?

News sources

Request for Additional Information and Documentary Material Issued to [Company] (Federal Trade Commission)
HSR Resources (Federal Trade Commission)
Order on Discovery Letter Brief AND Order on Discovery Letter Brief AND Order on Discovery Letter Brief: #203 in Crowder v. LinkedIn Corporation (N.D. Cal., 4:22-cv-00237) (Schulte v. LinkedIn Corp. order, U.S. District Court for the Northern District of California via CourtListener)
Order on Stipulation: #74 in James v. Cerebras Systems Inc. (N.D. Cal., 4:25-cv-09361) (U.S. District Court for the Northern District of California via CourtListener)
2026-10-21-fontainebleau-florida-hotel-v-botach.pdf (U.S. District Court for the Southern District of Florida via Servient)
FTC Issues Rigorous Requirements on Using AI-Assisted Review or Technology-Assisted Review to Respond to Investigational Subpoenas (Arnold & Porter, Oct. 7, 2026)
FTC Issues Rigorous Requirements On Using AI-Assisted Review Or Technology-Assisted Review To Respond To Investigational Subpoenas (Arnold & Porter via Mondaq, Oct. 11, 2026)
FTC Issues Rigorous Requirements On Using AI-Assisted Review Or Technology-Assisted Review To Respond To Investigational Subpoenas (Arnold & Porter via The Legal 500, Oct. 11, 2026)
Tara S. Emory and Maura R. Grossman: GenAI Prompts in eDiscovery: Protected Work Product or Not? (ACEDS)
Prompt privacy: new scholarship argues GenAI review instructions are attorney work product (ComplexDiscovery)

Assisted by GAI and LLM Technologies

Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more understandable for practitioners, policymakers, and business leaders.

Have a Request?

If you have information or offering requests that you would like to ask us about, please let us know, and we will make our response to you a priority.

ComplexDiscovery OÜ is an independent digital publication and research organization based in Tallinn, Estonia. ComplexDiscovery covers cybersecurity, data privacy, regulatory compliance, and eDiscovery, with reporting that connects legal and business technology developments—including high-growth startup trends—to international business, policy, and global security dynamics. Focusing on technology and risk issues shaped by cross-border regulation and geopolitical complexity, ComplexDiscovery delivers editorial coverage, original analysis, and curated briefings for a global audience of legal, compliance, security, and technology professionals. Learn more at ComplexDiscovery.com.

Generative Artificial Intelligence and Large Language Model Use

ComplexDiscovery OÜ recognizes the value of GAI and LLM tools in streamlining content creation processes and enhancing the overall quality of its research, writing, and editing efforts. To this end, ComplexDiscovery OÜ regularly employs GAI tools, including ChatGPT, Claude, Gemini, Grammarly, Midjourney, and Perplexity, to assist, augment, and accelerate the development and publication of both new and revised content in posts and pages published (initiated in late 2022).

+ Attend

Editor's Note: Estonia hosts the ninth Tallinn Digital Summit on Nov. 5 and 6 around one question: how to thrive in the agentic era. The program pairs Estonian ministers, OECD Secretary-General Mathias Cormann and European Commissioner Ekaterina Zaharieva with AI safety researchers and security leaders. Sessions include a live cyber range demonstration of AI assurance, the promised unveiling of a doctrine for asymmetric cyber defense built on a "Minimum Viable State," and a debate over whether a vendor-supplied "sovereign" technology stack counts as sovereignty.

For cybersecurity, information governance and eDiscovery professionals, the host country supplies a sharper signal. A timetable published in September calls for a December pilot of Estonian digital identities for AI agents, under which a receiving service could recognize when a machine, not a person, is acting. Officials describe the plan in different terms, and legal critics in Estonian media have questioned it. Either way, the line between machine and human action bears directly on how agent activity is logged and attributed.

ComplexDiscovery will attend and report from Tallinn for the third consecutive year. ComplexDiscovery OÜ is an Estonian private limited company, and two staff members are e-residents. Watch the doctrine, the agent code pilot and the sovereignty debate.

[exclude_from_rss]

[taq_review]

[/exclude_from_rss] Industry News - Technology Beat

Tallinn Digital Summit 2026 puts the agentic state to the test

ComplexDiscovery Staff Estonia will convene ministers, EU and OECD leaders, and AI researchers in Tallinn on Nov. 5 and 6 for the ninth Tallinn Digital Summit, built on one question: how to thrive in the agentic era. The host country is not asking in the abstract. A timetable published in September calls for a December pilot of Estonian digital identities for AI agents, and the country's digital affairs minister is calling for "an IAEA for artificial intelligence." ComplexDiscovery will attend and report from the summit as accredited media for the third consecutive year, following its coverage of the 2024 and 2025 editions. ComplexDiscovery OÜ is a private limited company registered in Estonia, and two ComplexDiscovery staff members are e-residents of Estonia, a group the first stage of that AI-agent identity plan would include. What follows is a guide to the program, the people and the country behind it.

Three pillars and an invitation-only room

The Tallinn-based International Centre for Defence and Security (ICDS) organizes the summit with Estonia's Ministry of Justice and Digital Affairs, Ministry of Foreign Affairs, Ministry of Economic Affairs and Communications and Government Office. ICDS was founded in 2006 by decision of the Estonian government as a foundation, and it describes itself as an independent, nonprofit think tank. Prime Minister Kristen Michal hosts the event at the National Library of Estonia. Organizers expect 1,000 delegates from over 70 countries, according to an Oct. 6 announcement, and attendance remains by invitation. For the first time, the summit is also offering a free Future Delegate Pass to selected university students, who must attend both days in person and participate actively. The 2026 theme carries three headline pillars: competitiveness, security and efficiency. The organizers' 2026 concept describes the moment as one of "rupture, not transition," with the war in Ukraine unresolved and technological change moving faster than governments can respond. It groups the discussion into three challenge areas: new democratic formats and partnerships for the agentic era, defense and digital resilience amid geopolitical tensions, and AI applied to governance and competitiveness. The concept also concedes that, for most organizations and even more governments, AI has yet to deliver substantial financial gains. It calls narrowing the distance between AI's promised and delivered results a priority.

Who takes the stage

That task falls first to a lineup of officials. Michal opens the first day, followed by keynotes from Ekaterina Zaharieva, the European commissioner for startups, research and innovation, and OECD Secretary-General Mathias Cormann, who is scheduled to present key findings from the OECD's digital economy and society outlook. Cormann then joins Estonian State Secretary Keit Kasemets for a fireside chat, and Foreign Minister Margus Tsahkna takes part in a late-morning conversation on digital sovereignty and trusted partnerships. Liisa-Ly Pakosta, Estonia's minister of justice and digital affairs, joins a Digital Nations panel that afternoon and gives the opening address on the second day. The published speaker list mixes government officials with AI safety researchers, security leaders, investors and journalists. It includes Stuart Russell of the University of California, Berkeley; Jaan Tallinn of the Future of Life Institute; Mozilla President Mark Surman; Susan Nesbitt, listed with Cantellus Group and as an OpenAI red team member; Freddy Dezeure, Microsoft's deputy chief information security officer for Europe; and Sasha Rubel of Amazon Web Services. Former Estonian President Toomas Hendrik Ilves also appears, as do Vahur Verte, a state prosecutor with Estonia's Prosecutor's Office, and Kian Vesteinsson, deputy director for research at Freedom House. The Oct. 6 update named further participants, including Anne Le Hénanff, France's minister for artificial intelligence and digital affairs; Nest AI founder Peter Sarlin; and Mohamed Al Kuwaiti, head of cybersecurity for the United Arab Emirates government. The organizers note the agenda remains subject to change.

Sessions where the evidence questions live

Beyond the keynotes, several sessions reach past policy talk and into the daily work of security, governance and discovery teams. On the second morning, "The Tallinn Stress Test" promises a live cyber range demonstration of AI assurance. It will probe two things: how AI can test the security of complex systems, and how an AI system behaves once it is wired into identities, data and tools. The session description frames the aim in terms any investigator would recognize: "identify unexpected behaviour, gather evidence and learn from what happens." CybExer Technologies has three executives on the speaker list. A late first-day session on national cyber resilience says it will unveil a doctrine for asymmetric cyber defense. It starts from what the session description calls the "Minimum Viable State," the few functions a government cannot allow to fail, each paired with a practiced fallback, and it sets one goal: "making every incident non-decisive." The same afternoon brings a session on fighting cybercrime with AI and digital law, and another asking whether handing a "sovereign" government technology stack to a vendor counts as sovereignty at all. That sovereignty question is an information governance question in plain clothes, because it decides who controls the cloud, the models and the data behind public services. A separate session on internet freedom adds a darker data point: Freedom House's Freedom on the Net 2025 report, released in November 2025, recorded a 15th straight year of declining internet freedom worldwide. Freedom House plans to release its 2026 edition Oct. 22, and Vesteinsson, one of its authors, is on the summit's speaker list. Practitioners evaluating AI agents can borrow the Stress Test's premise. The session description argues that a model viewed in isolation reveals only part of its behavior, so a procurement or governance review should watch an agent inside the systems, data stores and permissions it will touch, and should keep the test record as evidence.

Officials set the tone before the doors open

The program's emphasis on accountable agents echoes what Estonian officials have put on the record in recent months. Pakosta argued in an Oct. 6 post published under her name on the summit's site that AI needs its own IAEA, an international institution to back global standards. Her proposal has three parts: a global computing solidarity mechanism, a mandatory watermarking standard for AI-generated content from the largest developers, and a worldwide early-warning system for unusual AI behavior and security flaws, modeled on aviation safety reporting. "Trust and human rights are not constraints on innovation," wrote Pakosta, whose ministry is one of the summit's co-organizers. Closer to home, her government has already moved from argument to design on agent accountability. The Eesti.ai advisory board agreed in June that Estonia will move forward with digital identities for AI agents, and Michal backed the proposal, according to a June 17 government release. Michal said in the release that it "must be clear who is acting on whose behalf with what rights, and who is ultimately responsible." The e-Estonia Briefing Centre, which promotes Estonia's digital government, described the working design Sept. 23. Under the proposal, an "agent code" would be a separate identity linked to the person who authorizes the agent. It would not confer legal personhood, and responsibility would remain with the authorizing person or organization. Its permissions could be narrowed, monitored and withdrawn, and the first stage would focus on natural persons, including e-residents. Registration, authentication, permissions, audit trails and revocation all remained under development, and the timetable published that day called for draft legislation in September, followed by testing in October and November and a December pilot. "AI agent codes can give AI agents a trusted and accountable digital identity," Liina Vahtras said in the e-Estonia article. Vahtras headed the e-Residency program from 2023 and has served as Estonia's minister of economy and industry since Sept. 16, according to her government biography. Not every official describes the plan in the same terms. Anni Lehari, who leads Aruait, a separate public-sector AI project at the Information System Authority, said in a Sept. 8 ERR News report that agents will not receive personal identification codes, and, as ERR summarized her remarks, characterized the June announcement as a communications tactic. What matters, she said, is that an agent acting for a person can be linked to that person, which is the link the agent-code design already describes. Outside government, reaction has split. ERR noted that two attorneys and a University of Tartu constitutional law scholar had criticized the idea in Estonian media. Meelis Kull, a University of Tartu professor of artificial intelligence, saw more promise, calling identification codes for agents a sensible step because they could help establish on whose behalf an agent is acting. "Identity itself does not protect against mistakes, however," Kull said. "That is why authorizations must be revocable and responsibility must remain clear even when something goes wrong." Kull's caveat is the one investigators will recognize, and it is where the design matters for eDiscovery. An agent acting under its own code, rather than a borrowed human login, would let the receiving service recognize that a machine, not the person, was acting. The e-Estonia article says that distinction can vanish today when an agent works through a person's login or a service-specific token. Teams buying agentic tools can ask vendors now whether their agents act under a separate identity, and whether those permissions can be scoped, logged and revoked.

From the song festival to X-Road

Estonia's confidence in building identity infrastructure, now extended to machines, has roots in how the country rebuilt its state. Estonia first declared independence in February 1918, then lost it to Soviet annexation in 1940 after the Molotov-Ribbentrop Pact, according to Visit Estonia. On Sept. 11, 1988, the Tallinn Song Festival Grounds filled with an estimated 300,000 people for songs and independence speeches, a milestone of what became known as the Singing Revolution. On Aug. 23, 1989, nearly 2 million people across Estonia, Latvia and Lithuania joined hands in the Baltic Way, a chain Visit Estonia puts at 670 kilometers. Estonia's Supreme Council established de facto independence on Aug. 20, 1991, and this year marks the 35th anniversary of that restoration. The new state started almost from scratch, and the e-Estonia timeline presents that gap as an opportunity. The Tiger Leap program pushed computers and computer skills into schools in 1996. The X-Road data exchange layer followed in 2001, the mandatory ID card with digital signature in 2002, online voting in 2005 and e-Residency in 2014. In 2017 Estonia approved what it calls the world's first data embassy, which backs up critical databases in a high-security data center in Luxembourg. For records managers, the arrangement offers a public model of continuity planning at national scale, built, in e-Estonia's words, for "worst-case scenarios like critical system failures or external threats." Culture runs through the AI program, too. The kratt is a figure from Estonian mythology, a servant made of hay or old household odds and ends, and the government adopted it as its metaphor for AI. A 2026 article in the journal Imaginations points to a darker trait in the folklore: left idle, the kratt turns on its master. Its authors, Neda Atanasoski of the University of Maryland and Redi Koobak of the University of Strathclyde, argue that the state's use of the myth naturalizes a link between technological progress and liberal democracy. In their reading, it also obscures how Silicon Valley capitalism shaped Estonia's post-socialist transformation. The summit keeps the metaphor in view all the same, naming one of its three stages the Kratt Stage. The metaphor sits alongside newer programs with plainer names. The AI Leap education initiative, which the e-Estonia timeline presents as Tiger Leap's successor, began in 2025 with about 20,000 upper-secondary students and 3,000 teachers. Michal launched Eesti.ai on Jan. 27, 2026. "With AI, we can do more work with a higher added value with fewer people and essentially double the value of Estonian people's work by 2035," Michal said at the launch.

A small state with a large security voice

Those digital ambitions have long come wrapped in security concerns. In April 2007, politically motivated cyberattacks struck Estonia, and the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE) says they served as a wake-up call for other nations and the alliance. Estonia had proposed the center in 2004, after joining NATO, and it was established on May 14, 2008. The center's Tallinn Manual process produced a first edition in 2013 and the expanded Tallinn Manual 2.0 on the International Law Applicable to Cyber Operations in 2017. Its annual Locked Shields exercise has run since 2010, and the center reports 39 member nations in 2026. The government's proposed 2026 state budget, submitted to parliament in September 2025, provided for at least 5 percent of gross domestic product for defense under NATO criteria, plus 0.25 percent of GDP to support Ukraine. Estonia took over the chair of the Nordic-Baltic Eight (NB8) on Jan. 1, 2026. "Nordic-Baltic cooperation is a force multiplier for our shared security, economic strength and political influence," Tsahkna said when the chairmanship was announced in December. The summit organizers say Estonia also chairs the Digital Nations group this year, and Kaja Kallas, a former Estonian prime minister, now serves as the EU's high representative and vice president.

What ComplexDiscovery will watch

That security posture frames the program's AI ambitions, and last year's summit supplied the template this year's sessions will test. At the 2025 edition, former Estonian government chief information officer Luukas Ilves and Manuel Kilian of the Global Government Technology Centre Berlin released "The Agentic State," a vision paper proposing a 12-layer framework for agentic AI across government functions. This year's program includes a session asking what that idea looks like once it reaches a working public service. ComplexDiscovery will report on whether the cyber resilience doctrine is published in usable form, what the agent code pilot looks like on paper, and how the sovereignty debate settles between national, regional and vendor-supplied stacks. If an AI agent acts under its own identity code, who in your organization will own the record that proves what it did, and on whose authority?

News Sources

Tallinn Digital Summit 2026 (International Centre for Defence and Security)
about | Discover Digital Leadership Opportunities: Tallinn Digital Summit 2026 (International Centre for Defence and Security, 2026 concept page)
about | Discover Digital Leadership Opportunities: Tallinn Digital Summit 2026 (International Centre for Defence and Security, summit history page)
agenda | Discover the Event Schedule: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
speakers | Explore Speakers Now: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
One Month to Go: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
For the first time, Tallinn Digital Summit opens its doors to students: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
AI needs its own IAEA: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
Kratt Manifest: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
Government launched the Eesti.ai initiative together with leading entrepreneurs | Eesti Vabariigi Valitsus (Government of Estonia)
Prime Minister Michal: Estonia to become first country to create digital identities for AI agents | Eesti Vabariigi Valitsus (Government of Estonia)
Liina Vahtras | Eesti Vabariigi Valitsus (Government of Estonia)
Government approved the 2026 state budget | Eesti Vabariigi Valitsus (Government of Estonia)
Estonia to chair the Nordic-Baltic (NB8) cooperation format as of 1 January 2026 | Välisministeerium (Estonian Ministry of Foreign Affairs)
About us (NATO Cooperative Cyber Defence Centre of Excellence)
About: International Centre for Defence and Security (International Centre for Defence and Security)
EU HR/VP Kaja Kallas at the European Arctic Summit 2026 | EEAS (European External Action Service)
NEW REPORT: Persistent Authoritarian Repression and Backsliding in Democracies Drive 15th Consecutive Year of Decline in Global Internet Freedom | Freedom House (Freedom House)
AI agents will not get Estonian ID codes and responsibility will stay with humans | News | ERR (ERR News)
Last USSR census 30 years ago counted largest-ever number of residents | News | ERR (ERR News)
Mohamed Al Kuwaiti: Agenda Contributor | World Economic Forum (World Economic Forum)
Automated Myth and the Postsocialist Digital State: The Imperial Trick of the Kratt AI Initiative in Estonia | Imaginations: Journal of Cross-Cultural Image Studies (Imaginations: Journal of Cross-Cultural Image Studies)
REPORT LAUNCH: Freedom on the Net 2026 | Freedom House (Freedom House)
30 years since the most important Singing Revolution concert (Estonian World)
Story: e-Estonia (e-Estonia Briefing Centre)
From e-Residency to agent codes: Estonia´s next digital identity experiment (e-Estonia Briefing Centre)
AI and the Kratt* momentum: e-Estonia (e-Estonia Briefing Centre)
How Estonia regained its independence (Visit Estonia)
Tallinn Digital Summit 2024: Charting a Secure Digital Future Amid Geopolitical Shifts (ComplexDiscovery)
The Agentic State: A Global Framework for Secure and Accountable AI-Powered Government (ComplexDiscovery)

[the_ad_group id="15279"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

[exclude_from_rss]

[taq_review]

[/exclude_from_rss] Industry News - Technology Beat

Tallinn Digital Summit 2026 puts the agentic state to the test

Three pillars and an invitation-only room

Who takes the stage

Sessions where the evidence questions live

Officials set the tone before the doors open

From the song festival to X-Road

A small state with a large security voice

What ComplexDiscovery will watch

News Sources

Tallinn Digital Summit 2026 (International Centre for Defence and Security)
about | Discover Digital Leadership Opportunities: Tallinn Digital Summit 2026 (International Centre for Defence and Security, 2026 concept page)
about | Discover Digital Leadership Opportunities: Tallinn Digital Summit 2026 (International Centre for Defence and Security, summit history page)
agenda | Discover the Event Schedule: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
speakers | Explore Speakers Now: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
One Month to Go: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
For the first time, Tallinn Digital Summit opens its doors to students: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
AI needs its own IAEA: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
Kratt Manifest: Tallinn Digital Summit 2026 (International Centre for Defence and Security)
Government launched the Eesti.ai initiative together with leading entrepreneurs | Eesti Vabariigi Valitsus (Government of Estonia)
Prime Minister Michal: Estonia to become first country to create digital identities for AI agents | Eesti Vabariigi Valitsus (Government of Estonia)
Liina Vahtras | Eesti Vabariigi Valitsus (Government of Estonia)
Government approved the 2026 state budget | Eesti Vabariigi Valitsus (Government of Estonia)
Estonia to chair the Nordic-Baltic (NB8) cooperation format as of 1 January 2026 | Välisministeerium (Estonian Ministry of Foreign Affairs)
About us (NATO Cooperative Cyber Defence Centre of Excellence)
About: International Centre for Defence and Security (International Centre for Defence and Security)
EU HR/VP Kaja Kallas at the European Arctic Summit 2026 | EEAS (European External Action Service)
NEW REPORT: Persistent Authoritarian Repression and Backsliding in Democracies Drive 15th Consecutive Year of Decline in Global Internet Freedom | Freedom House (Freedom House)
AI agents will not get Estonian ID codes and responsibility will stay with humans | News | ERR (ERR News)
Last USSR census 30 years ago counted largest-ever number of residents | News | ERR (ERR News)
Mohamed Al Kuwaiti: Agenda Contributor | World Economic Forum (World Economic Forum)
Automated Myth and the Postsocialist Digital State: The Imperial Trick of the Kratt AI Initiative in Estonia | Imaginations: Journal of Cross-Cultural Image Studies (Imaginations: Journal of Cross-Cultural Image Studies)
REPORT LAUNCH: Freedom on the Net 2026 | Freedom House (Freedom House)
30 years since the most important Singing Revolution concert (Estonian World)
Story: e-Estonia (e-Estonia Briefing Centre)
From e-Residency to agent codes: Estonia´s next digital identity experiment (e-Estonia Briefing Centre)
AI and the Kratt* momentum: e-Estonia (e-Estonia Briefing Centre)
How Estonia regained its independence (Visit Estonia)
Tallinn Digital Summit 2024: Charting a Secure Digital Future Amid Geopolitical Shifts (ComplexDiscovery)
The Agentic State: A Global Framework for Secure and Accountable AI-Powered Government (ComplexDiscovery)

[the_ad_group id="15279"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

Tallinn Digital Summit 2026 puts the agentic state to the test

The 2026 Event Horizon: Early Outlook for eDiscovery, AI, and European Innovation

+ Learn

Editor's Note: ComplexDiscovery OÜ is publishing a three-term Oxford-style tutorial that asks whether, beyond awareness and credibility, marketing drives revenue in eDiscovery and legal technology. The question lands as B2B buyers report ranking their shortlists before they speak to a seller, according to 6sense's 2025 survey of nearly 4,000 buyers.

For practitioners, the stakes are concrete. Provider marketing and sales leaders defend budgets with attribution figures, yet researchers analyzing 15 Facebook advertising experiments found that observational methods often failed to match randomized results. Legal operations and procurement teams build vendor lists from reputation and signals they cannot fully inspect. Information governance and cybersecurity buyers may face similar dynamics when they evaluate providers of their own. The tutorial sets 21 propositions, each written to be argued either way, and demonstrates the method with three worked examples. Watch for how providers respond to buyer-led research in their 2027 planning, and whether measurement practice shifts from attribution dashboards toward randomized holdouts. Disclosure: ComplexDiscovery OÜ's founder also serves as chief marketing officer of an eDiscovery provider; the tutorial's propositions are argued either way and are not the publication's position. [exclude_from_rss]

[taq_review]

[/exclude_from_rss] News Analysis - Leadership Beat

Does marketing drive revenue? An Oxford-style tutorial for eDiscovery and legal tech

ComplexDiscovery Staff Many provider marketing budgets in eDiscovery and legal technology rest on an assumption that is seldom argued out loud: that past a certain point, more marketing still produces more revenue. ComplexDiscovery OÜ has put that assumption on trial in the latest installment of The ComplexDiscovery Tutorial, adapting the Oxford tutorial for a field that often moves faster than it reflects. The tutorial is built around one governing question. Once a firm is known and trusted, does further marketing spending cause revenue, or does it mostly take credit for revenue that sales relationships, referrals and reputation would have produced anyway? The tutorial concedes the floor at the outset, noting that few serious observers argue that awareness and credibility are worthless; a provider no buyer has heard of rarely reaches a shortlist. The contested ground is everything above that floor, from demand generation and content programs to events, intent data and attribution dashboards, in a market where much of the work is bought through relationships, panels and procurement. To work that ground, the curriculum sets 21 contestable propositions across three terms, each proposition paired with primary readings and an essay prompt, and closes each term with a collections paper. Three worked examples, one per term, let readers watch the method in use rather than take it on faith.

A disciplined conversation

The Oxford tutorial is less a lecture than a disciplined conversation. As the tutorial document frames it, the form is a weekly argument: a student writes an essay on an assigned proposition, reads it aloud and defends it against a tutor whose job is to find the weakest sentence. The essay is expected to take a side, use its readings as evidence rather than decoration and concede what it cannot defend. The point is not to settle whether marketing works. It is to make every claim about marketing survive a hostile reader, whether that reader is a chief financial officer, a procurement lead or a skeptical colleague in sales.

A method, then three terms

The curriculum follows the Oxford academic calendar, whose three terms are Michaelmas (autumn), Hilary (winter) and Trinity (spring into early summer). Each term presses a different question: what the evidence says marketing does, whether anyone can measure it honestly and whether any of it holds in the legal technology market.

Michaelmas: the evidence

Michaelmas tests whether marketing causes sales at all, drawing on experimental and meta-analytic evidence, most of it from consumer markets. The first proposition (that once a firm is known and credible, further marketing adds little revenue) pairs Byron Sharp's "How Brands Grow" with a 2021 Econometrica study by Bradley Shapiro, Günter Hitsch and Anna Tuchman, which found negative marginal returns on TV advertising for over 80 percent of the 288 consumer brands it examined. The second (that marketing works mainly on buyers who are not yet in the market, so its effect is real but deferred) draws on the 95-5 rule set out by John Dawes of the Ehrenberg-Bass Institute with The B2B Institute at LinkedIn, which holds that up to 95 percent of business buyers are out of the market at any given time, and on Les Binet and Peter Field's "The Long and the Short of It." The third proposition (that measured advertising effects are smaller than marketers believe) rests on a 2011 meta-analysis by Raj Sethuraman, Gerard Tellis and Richard Briesch that reported a mean short-term advertising elasticity of 0.12 and a median of 0.05. The fourth asks whether visible spending signals quality a buyer cannot inspect, through Phillip Nelson's "Advertising as Information" and Amna Kirmani and Akshay Rao's review of signaling research. The fifth (that marketing routinely takes credit for demand it did not create) centers on an eBay field experiment by Thomas Blake, Chris Nosko and Steven Tadelis, in which brand keyword ads showed no measurable short-term benefit. The sixth draws on Leonard Lodish and colleagues' summary of 55 in-market TV experiments, which found that where a TV weight increase lifted sales in its first year, the total effect including the following two years was roughly double the first-year lift, while tests whose first-year effect fell short of statistical significance showed no later effect on average. The seventh turns on the tutorial's own premise, asking whether awareness and credibility can be separated from revenue at all, with readings from Susan Mudambi, Peter Doyle and Veronica Wong on industrial branding and "The B2B Effectiveness Code," produced by The B2B Institute at LinkedIn with WARC and Lions. The term closes with collections, a timed paper on the claim that marketing does not drive revenue but builds the conditions under which selling can.

Hilary: the measure

Hilary asks whether marketing's revenue effect can be isolated in practice, or only asserted. Its first proposition borrows the saying that half of advertising money is wasted and no one knows which half (Quote Investigator names John Wanamaker the leading candidate but finds no direct source) and sets it beside Randall Lewis and Justin Rao's finding that even large digital experiments produce return estimates too noisy to read with confidence. The second (that attribution models report correlation as contribution) draws on 15 Facebook experiments analyzed by Brett Gordon, Florian Zettelmeyer, Neha Bhargava and Dan Chapsky, in which observational methods often failed to match randomized results, and on Ron Berman's critique of last-touch attribution. The third proposition asks whether randomized holdouts are realistic at B2B scale, using the "ghost ads" method developed by Garrett Johnson, Randall Lewis and Elmar Nubbemeyer. The fourth applies the widely quoted phrasing of Goodhart's law from anthropologist Marilyn Strathern ("When a measure becomes a target, it ceases to be a good measure") to marketing-sourced pipeline quotas. The fifth treats "influenced revenue" as a definition rather than a finding and asks students to rewrite it so it could be proved false. The sixth asks whether the honest test of marketing is what happens when it stops. The seventh asks whether, in a long and committee-driven sale, measurement fails before marketing does. In 6sense's 2025 Buyer Experience Report, which surveyed nearly 4,000 B2B buyers, the average buying cycle ran 10.1 months. Collections sets the claim that if marketing cannot be measured honestly, it cannot be managed honestly.

Trinity: the market

Trinity moves the argument into eDiscovery and legal technology. It opens with the proposition that legal buyers buy people rather than brands, read through David Maister, Charles Green and Robert Galford's "The Trusted Advisor" and Matthew Dixon and Brent Adamson's "The Challenger Sale." The second (that the shortlist is set before marketing can touch it) uses 6sense's 2025 finding that B2B buyers made first contact with sellers about 61 percent of the way through their process and that 94 percent of buying groups had already ranked their shortlist. Those figures describe B2B buying broadly, not legal technology specifically. The third asks whether procurement has turned legal technology buying into a specification and price exercise, citing the Buying Legal Council's 2021 survey, which, as reported by Corporate Counsel Business Journal, found procurement influenced 76 percent of legal technology purchases. The fourth proposition tests whether thought leadership is the one form of marketing legal buyers reward, using a 2024 Edelman and LinkedIn study of 3,484 executives in which 86 percent of decision-makers said they would be "moderately likely" or "very likely" to invite an organization that consistently produces high-quality thought leadership into a request for proposal process when they are in the market. It also asks how much weight a report produced with a platform that sells the channel can bear. The fifth uses ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup, which places services at 66 percent of a $19.61 billion market in 2025, to ask whether marketing in a services-heavy market protects price rather than winning share. The sixth sets a Gartner survey of 646 B2B buyers, 67 percent of whom preferred a rep-free experience, against Peter Drucker's often-quoted line that the aim of marketing is to make selling superfluous. A second Gartner release drawn from the same fieldwork adds a counterweight: 69 percent of buyers prefer to validate AI-generated insights with sales representatives. The seventh, drawing on Theodore Levitt's "Marketing Myopia," asks whether "marketing or sales?" is the wrong question altogether. The final collections paper is the governing question itself: that beyond basic awareness and credibility, marketing does not matter in driving revenue.

The form in motion

The tutorial demonstrates its method with three worked examples, one per term, each pairing a model essay with a constructed interrogation. The exchanges are written simulations in composite voices, not transcripts, and are presented in full below. The first, from Michaelmas, argues that marketing routinely takes credit for demand it did not create, building from the eBay experiment to B2B attribution. The tutor presses on whether a mid-market provider resembles a household brand, whether the essay has shown where the money goes or only where measurable engagement lands, and whether a buyer's prior experience with a vendor caused the buyer's ranking or merely accompanied it. The second, from Hilary, argues against treating a marketing blackout as an honest test, citing carryover, cycle length, statistical power and the signal a silent firm sends. The tutor asks what evidence would ever change the student's mind, and pushes the essay from "patient" to a specific test length. The third, from Trinity, argues that buyer-journey statistics show the shortlist is set before sales can touch it, not before marketing can. The tutor asks whether the essay proves marketing's case or only that something before the sales call matters, and follows the argument back to the floor the tutorial concedes at the start.

Running it against your own work

The tutorial is designed to be run alone as well as in pairs. A reader can take any proposition, write the strongest essay for the side they do not hold, then interrogate it as a tutor would: what is the counterfactual, what would change the writer's mind, and does the conclusion claim beyond what the evidence supports? The payoff differs by seat. For a provider's marketing leader, it is a rehearsal for the budget meeting where every figure meets a skeptical reader. For a sales leader, it is a way to separate the deals marketing created from the deals it recorded. For legal operations and procurement teams, it is a lens on how their own shortlists form, and on which signals they trust when they cannot inspect quality in advance. One caution runs through the reading list. Much of the strongest causal evidence comes from consumer markets, the buyer-journey surveys describe B2B buying broadly rather than legal technology in particular, and several of the B2B studies were produced or commissioned by firms that sell or operate the channels they measure. The tutorial asks readers to weigh all three facts, not to discard the evidence. The exercise for any team is the one the tutorial is built on. Put marketing's strongest claim about revenue in front of a hostile examiner. Which sentence breaks first?

The tutorials in full

The three worked examples below are illustrative examples provided by ComplexDiscovery OÜ. Each pairs a model essay with a constructed interrogation: a written simulation of a tutorial exchange in composite, illustrative voices, not a transcript of a live session between real people. The studies, surveys and scholars the essays cite are real and were checked against publisher, journal, repository and news pages. The propositions are written to be argued either way, and the side each essay takes is an exercise, not a position held by ComplexDiscovery OÜ.

Tutorial one, Michaelmas: "That marketing routinely takes credit for demand it did not create"

Explainer This example trains the first habit of the term: asking what would have happened without the intervention. Watch how the essay builds from a single consumer experiment to a claim about B2B attribution, and how the tutor tests whether that bridge can carry the weight placed on it. The essay Michaelmas, week five. The student was asked to argue for the proposition, using Blake, Nosko and Tadelis (2015) and at least two other Michaelmas readings. The click that was already coming When eBay stopped buying search ads on its own brand name on MSN and Yahoo, almost nothing happened. Blake, Nosko and Tadelis report that 99.5 percent of the click traffic those ads had been delivering arrived anyway, through the unpaid search results sitting just below them. The ads had been counting that traffic as their own. I will argue that this experiment is not a curiosity about search engines. It is a precise description of how marketing is credited in most firms, and B2B firms that sell into legal departments and law firms are more exposed to it than eBay, not less. First, the mechanism. Marketing is placed where buyers already are. A brand keyword ad appears only to someone who typed the brand. A retargeting ad follows someone who already visited. A nurture email reaches someone who already gave an address. Each of these touches selects for people who were on their way, and then records their arrival. The eBay authors show what that selection does to the numbers. For non-brand search ads, a simple regression of sales on spending implied returns of over 4,000 percent. Adding controls for day and geography still left an estimate above 1,500 percent. The controlled experiment found a return of minus 63 percent. Nothing about the ads changed between those figures. What changed was the question: from "who bought after seeing an ad?" to "who bought because of one?" Second, the mechanism in eDiscovery. A provider's most likely buyers are its current clients, its former clients and the colleagues those people talk to. They attend its webinars because they already use it. They download its reports because they already trust it. When a matter arrives and a contract follows, the customer relationship management (CRM) system looks back, finds a webinar registration and a report download, and records the deal as marketing-influenced or even marketing-sourced. A 2025 study by 6sense of nearly 4,000 B2B buyers gives a sense of scale: 94 percent of buying groups had ranked their shortlist before engaging sellers, and 85 percent of buyers reported prior experience with the vendor they ultimately chose. A buyer who arrives with a ranked shortlist and a history is the B2B equivalent of someone typing "eBay" into a search bar. The touch is real. The causation is borrowed. Third, the size of careful estimates. Sethuraman, Tellis and Briesch, pooling 751 short-term estimates from published studies, found a mean advertising elasticity of 0.12 and a median of 0.05, well below the 0.22 mean an earlier meta-analysis reported. Shapiro, Hitsch and Tuchman, using a research design built to limit confounding across 288 consumer brands, found negative marginal returns on TV advertising for over 80 percent of them. These are consumer studies, and I will not pretend they transfer cleanly. But they suggest a pattern: the more carefully effects are estimated, the smaller they tend to look. If a provider's dashboard credits marketing with returns far above what careful studies find, that gap is a hypothesis to test before it is a result to report. Until the test is run, the likelier explanation is not that legal technology marketing is unusually potent. It is that the dashboard is counting clicks that were already coming. The incentive makes this worse. A marketing team measured on sourced pipeline has every reason to place touches as close to the buyer's decision as possible, where credit is cheapest to collect. That is the brand keyword strategy, applied to webinars and gated reports. It is rational for the team and expensive for the firm. The strongest objection comes from the same paper. The eBay authors found that paid search had measurable effects on new and infrequent users, with the largest effect among users who had not bought on eBay before, while the effect faded for frequent buyers. Marketing that reaches strangers can create demand. I accept this. But it sharpens the proposition rather than defeating it. Most of a known provider's measurable engagement comes from people who already know it, because that is where engagement metrics are easiest to produce. The stranger is expensive to reach and slow to convert, and so the stranger is underfunded. The proposition is true as a description of how marketing is usually credited. Whether it is true of what marketing could do is a different question, and the answer depends on whether a firm is willing to stop paying for the click that was already coming and spend instead on the buyer who has never heard of it. The interrogation Tutor: You open with eBay. Why should a reader in legal technology care what happens to a household name selling consumer goods? Student: Because the mechanism doesn't depend on the product. Any firm that places touches where its existing buyers already are will over-count. Tutor: But the size of the effect depends on how well known the firm is. eBay's brand ads were redundant because everyone already knew eBay. Is a mid-market eDiscovery provider closer to eBay or to a stranger? Student: For its existing clients, closer to eBay. For the rest of the market, closer to a stranger. Tutor: Then your essay needs two numbers it does not have. What share of the provider's marketing reaches each group? Student: I don't have them. I'd argue most engagement comes from the first group, because they're the ones who open the emails. Tutor: "Most engagement" is not "most spending." A conference sponsorship reaches strangers and clients alike. Have you shown that the money, rather than the measurable engagement, goes to people who already know the firm? Student: No. I've shown the measurement is biased toward them. That's a narrower claim. Tutor: It may be the claim you can defend. Let's test the B2B bridge. You cite 94 percent with a ranked shortlist before contact. Who formed that ranking? Student: Prior experience, mostly. The same study says 85 percent had prior experience with the vendor they chose. Tutor: Experience with the winner. Does that tell you the experience produced the ranking? Student: Not directly. It's consistent with it. Tutor: Consistent with it, and with a relationship that marketing helped open years earlier. Be careful not to read the causation you need into the correlation you have. Now: what is the counterfactual for a webinar attendee who later signs? Student: That they would have signed without the webinar. Tutor: How would you know? Student: Randomize the invitations. Hold some accounts out of the webinar series and compare conversion. Tutor: With how many accounts? Student: A few hundred, for most providers. Tutor: That is Hilary's problem, so I'll leave it there. Last question. Suppose you're right and credit-taking is widespread. Is the remedy less marketing, or better accounting? Student: Better accounting first. Then less of the marketing that the accounting shows was taking credit. Tutor: And more of what? Student: Of the work that reaches people who don't know the firm yet. The concession, really. Tutor: Then your conclusion argues against your title. The proposition says marketing takes credit. Your essay ends by saying some of it creates demand and is underfunded. Rewrite the last paragraph so it owns both halves.

Tutorial two, Hilary: "That the honest test of marketing is what happens when it stops"

Explainer This example trains the central discipline of Hilary: separating "unmeasured" from "unmeasurable," and designing a test before defending a verdict. Watch the essay's strongest move (carryover) and the tutor's question about whether the essay has quietly made its own claim impossible to disprove. The essay Hilary, week six. The student was asked to argue against the proposition, drawing on Lodish et al. (1995), Lewis and Rao (2015) and at least one other reading from the term. Silence is not a control group There is a version of this proposition that every chief financial officer finds attractive. Stop the marketing for a quarter. If revenue holds, the spending was waste. If revenue falls, the spending was earning its keep. It sounds like an experiment. I will argue that it is not one, and that a firm that runs it learns less than it thinks and may pay for the lesson later. First, carryover. Lodish and colleagues, summarizing 55 in-market TV experiments, found that when a weight increase lifted sales in its first year, the total effect, including the following two years, was roughly double the first-year lift. Advertising that works keeps paying after the spending ends. Run that logic in reverse. A firm that goes dark is still collecting on past investment, so revenue in the dark quarter reports the decay rate of what was spent before, not the value of what was being spent now. The test is built to report "no harm" early. Second, the calendar. A 2025 study by 6sense of nearly 4,000 B2B buyers puts the average buying cycle at 10.1 months, and 11.1 months in North America. Deals that close in the quarter a provider goes quiet were, on that average, opened most of a year earlier. A one-quarter pause measures the pipeline the firm built last year. Its effect on the pipeline the firm would have built this year shows up after the test has been declared over. Third, statistical power. Lewis and Rao examined 25 large digital advertising field experiments, most reaching over a million customers, and found individual sales so volatile relative to advertising effects that the median retail experiment produced a confidence interval on return on investment over 100 percentage points wide. A provider with a few hundred target accounts, each placing irregular and outsized orders, is in a far worse position. The question "did revenue fall when we stopped?" may be difficult to answer at that scale with confidence, in either direction, unless a power analysis shows otherwise. Fourth, going dark is itself a message. Kirmani and Rao review the argument that visible spending can signal confidence in quality, because only firms expecting repeat business can afford it. If spending can carry that signal, withdrawal may be read the other way, and not only by buyers. That is an inference rather than a finding. Still, the audiences are easy to name: competitors notice an empty booth, recruits notice a silent feed and analysts notice a missed briefing. If any of them reacts, the test changes the market it is trying to measure. The best case for the proposition is that without some holdout, every claim marketing makes about itself is unfalsifiable, and that is worse than an imperfect test. Shapiro, Hitsch and Tuchman's finding that over 80 percent of 288 brands had negative marginal returns on TV advertising suggests many firms could spend less and earn more, and a firm that never stops will never learn whether it is one of them. I accept the principle. I reject the method. The honest test is narrow, randomized and patient. Johnson, Lewis and Nubbemeyer's ghost ads design shows that measurement can compare buyers who saw a campaign with comparable buyers who would have seen it, without darkening the whole market. A provider can hold a random set of accounts out of a named program, keep everything else running and wait. Gordon and colleagues' Facebook experiments show why the randomization matters: observational methods often failed to recover the effects the experiments measured, even with rich data on every user. Stopping everything is a test of nerve, not of marketing. The proposition is right that marketing should be tested by its absence. It is wrong about whose absence, and for how long. The interrogation Tutor: You list four objections to going dark. If going dark is never a fair test, what evidence would ever persuade you that a provider's marketing does nothing? Student: A randomized holdout of accounts from a program, run long enough, showing no difference. Tutor: And if the provider can't find enough accounts to detect a difference? Student: Then it can't prove the marketing works either. Tutor: Which leaves it where? Student: Spending on faith. Tutor: Your essay says the answer is uncertain "in either direction," then treats that uncertainty as a reason to keep spending. Lewis and Rao cut both ways: if you can't detect the effect of stopping, you can't detect the effect of continuing. Why should the default be to keep spending? Student: Because of carryover. Stopping has a delayed cost you won't see in time. Tutor: Lodish also found that tests whose first-year effect fell short of statistical significance showed no later effect on average. Doesn't that weaken your carryover argument for programs that never showed an early signal? Student: It does. It means carryover protects only marketing that was already working. Tutor: So a provider could cut the programs that never showed anything and keep the ones that did. Student: If it could tell which were which. Tutor: Take the signaling point. You say an empty booth tells the market something. How much of that is a revenue effect, and how much is a marketing team's anxiety about being visible? Student: I can't separate them. I'd say the signal matters most to recruits and analysts, which are indirect routes to revenue. Tutor: Indirect routes are where unfalsifiable claims live. Is "it takes too long to show" an argument or a shelter? Student: It's an argument when you can say how long, and a shelter when you can't. Tutor: Then say how long. You have a cycle length in your own essay. Student: The study says 10.1 months on average. As a planning assumption, a program holdout should run at least one full cycle, so roughly a year, adjusted to the provider's own cycle data, before anyone reads the result. Tutor: Good. Put that number in your conclusion, and say where it came from. "Patient" is an adjective. "One full buying cycle" is a test.

Tutorial three, Trinity: "That the shortlist is set before marketing can touch it"

Explainer This example trains a Trinity habit: reading a statistic against the conclusion it is usually recruited to support. Watch how the essay turns the proposition's own evidence around, and how the tutor asks whether the reversal proves marketing's case or only that something before the sales call matters. The essay Trinity, week two. The student was asked to argue against the proposition, using the 6sense 2025 study and at least one reading on legal procurement. Who was in the room before the room The proposition rests on figures that circulate at every sales kickoff. In 6sense's 2025 Buyer Experience Report, which surveyed nearly 4,000 B2B buyers, respondents made first contact about 61 percent of the way through their purchasing process and initiated that contact 79 percent of the time. In 94 percent of buying groups, the shortlist was already ranked by preference, and 77 percent of buyers had their first vendor conversation with the eventual winner. A Gartner survey of 646 B2B buyers, fielded in August and September 2025, found that 67 percent would prefer a buying experience without a sales representative. In legal technology the gate narrows further: the Buying Legal Council's 2021 survey, as reported by Corporate Counsel Business Journal, found procurement influenced 76 percent of legal technology purchases. The usual conclusion is that the deal is decided before anyone can influence it. I will argue that these figures prove less than the proposition claims, and point the other way. They show the shortlist is set before sales can touch it. They do nothing to show it is set before marketing can. First, what a ranked shortlist requires. A buyer who arrives with a favorite has formed a view from something. The candidates are few: prior work with the provider, word of mouth from peers, what the buyer has read, heard and seen, and where the provider stands with analysts and on procurement lists. Of these, only prior work sits entirely outside marketing's reach. Peer conversation is shaped by what peers have read. Analyst coverage and panel inclusion are earned in part through briefings, published material and visibility at the events where legal professionals gather. The share of the journey completed before contact fell from about 69 percent in 2024 to about 61 percent in 2025, but most of it still happens before sales arrives. If the decision is largely made in that stretch, then it is where the contest happens, and it is the part of the journey marketing, not sales, occupies. Second, timing. Dawes, of the Ehrenberg-Bass Institute, argues in work with The B2B Institute at LinkedIn that up to 95 percent of business buyers are not in the market for a given service at any moment. The memory a buyer draws on when a matter arrives was built while they were not buying. A shortlist set before contact was, on this account, set during the long period when the buyer was reading, attending and remembering with no intent to purchase. That is the period the proposition says marketing cannot reach. It is the stretch of the journey that marketing can reach and sales usually does not. Third, the legal market's own structure. ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup places services at 66 percent of a $19.61 billion eDiscovery market in 2025. Services are bought on trust in people and process, qualities a buyer cannot inspect before the work begins. Nelson's distinction between search goods and experience goods applies: when quality can only be judged after purchase, buyers lean on reputation and signals. Procurement does not remove that dependence. It formalizes it, by turning reputation into a vendor list that someone first had to get onto. The objection I take most seriously comes from the same 6sense study: 85 percent of buyers reported prior experience with the vendor they chose, and 68 percent already knew a seller there. If the winner is usually a vendor the buyer has already used, then the shortlist is set by delivery, and marketing is a bystander. I cannot rule this out with the data available. Even so, it does not rescue the proposition. A shortlist set by past delivery is set before this deal's marketing, but the first engagement that led to that delivery had to begin somewhere. The proposition is wrong as written. The shortlist is set before the sales conversation, during hours that belong to reputation, and marketing is one of the few forces at work in those hours. Whether it is the decisive one is a question these surveys cannot answer. The interrogation Tutor: You've turned the statistics around neatly. Now turn them back. What does a preference built by delivering a matter look like, compared with one built by a campaign? Student: The first comes from someone who worked with the provider's team. The second comes from someone who only knows the brand. Tutor: Which kind does the 6sense ranking describe? Student: It doesn't say. It reports a ranking, not where the ranking came from. Tutor: So your essay proves that something before sales matters. Does it prove that marketing is that something? Student: It shows marketing is one of the few things operating in that window. Tutor: One of a few. Name the others and rank them. Student: Prior work, peer referral, analyst and panel position, published reputation. For eDiscovery services, honestly, I'd rank prior work and referral first. Tutor: Your conclusion calls marketing one of the few forces at work; your own ranking puts it behind prior work and referral. The final paragraph should say so. Next: you say the first engagement "had to begin somewhere." Where did it begin for a provider assembled through acquisitions, or founded by a team that left a competitor? Student: With relationships the people brought with them. Tutor: Not with marketing. Student: Not originally. Marketing might keep those relationships warm. Tutor: "Might keep them warm" is a different essay. A harder case: imagine a market where every credible provider markets equally well. What happens to your argument? Student: Marketing stops differentiating. It becomes the price of being considered. Tutor: Which is? Student: Awareness and credibility. The floor in the governing question. Tutor: So in a mature market, your essay collapses into the same concession the tutorial grants at the start. Student: Unless one provider markets better than the others. Tutor: Then the question becomes whether "better" can be measured, which is Hilary's problem again. Last: procurement. You say it formalizes reputation into a vendor list. Who decides which vendors make the list? Student: Legal operations and procurement, usually with input from the lawyers who used them. Tutor: Then the list is the lawyers' experience in a spreadsheet. Where is marketing in that? Student: In getting a provider in front of those lawyers the first time. I can't show it does anything beyond that. Tutor: Then say so. An essay that knows the limit of its evidence is stronger than one that hides it.

An invitation

The tutorial does not end with a verdict. Its final paper sets the governing question as written, and it is meant to be argued from either side. Treat it as an invitation to argue the position you do not hold. If you believe marketing drives revenue, write the essay for "That marketing routinely takes credit for demand it did not create." If you believe it does not, write against "That the shortlist is set before marketing can touch it." Then hand the essay to someone who will interrogate it. The questions cost nothing. The thinking is the work.

Sources and reference materials References are cited in full; titles link to a public version where one exists. ComplexDiscovery OÜ

ComplexDiscovery OÜ. “Complete look: ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup.” June 26, 2026.

Scholarship and commentary

Berman, Ron. “Beyond the Last Touch: Attribution in Online Advertising.” Marketing Science 37, no. 5 (2018): 771-792.
Binet, Les, and Peter Field. The Long and the Short of It. IPA, 2013.
Blake, Thomas, Chris Nosko, and Steven Tadelis. “Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment.” Econometrica 83, no. 1 (2015): 155-174.
Dawes, John. “Advertising Effectiveness and the 95-5 Rule: Most B2B Buyers Are Not in the Market Right Now.” Ehrenberg-Bass Institute for Marketing Science with The B2B Institute at LinkedIn, 2021.
Dixon, Matthew, and Brent Adamson. The Challenger Sale: Taking Control of the Customer Conversation. Portfolio/Penguin, 2011.
Drucker, Peter F. Management: Tasks, Responsibilities, Practices. Harper & Row, 1973.
Gordon, Brett R., Florian Zettelmeyer, Neha Bhargava, and Dan Chapsky. “A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook.” Marketing Science 38, no. 2 (2019): 193-225.
Johnson, Garrett A., Randall A. Lewis, and Elmar I. Nubbemeyer. “Ghost Ads: Improving the Economics of Measuring Online Ad Effectiveness.” Journal of Marketing Research 54, no. 6 (2017): 867-884.
Kirmani, Amna, and Akshay R. Rao. “No Pain, No Gain: A Critical Review of the Literature on Signaling Unobservable Product Quality.” Journal of Marketing 64, no. 2 (2000): 66-79.
Levitt, Theodore. “Marketing Myopia.” Harvard Business Review (July-August 1960): 45-56.
Lewis, Randall A., and Justin M. Rao. “The Unfavorable Economics of Measuring the Returns to Advertising.” Quarterly Journal of Economics 130, no. 4 (2015): 1941-1973.
Lodish, Leonard M., Magid M. Abraham, Jeanne Livelsberger, Beth Lubetkin, Bruce Richardson, and Mary Ellen Stevens. “A Summary of Fifty-Five In-Market Experimental Estimates of the Long-Term Effect of TV Advertising.” Marketing Science 14, no. 3, supplement (1995): G133-G140.
Maister, David H., Charles H. Green, and Robert M. Galford. The Trusted Advisor. Free Press, 2000.
Mudambi, Susan M., Peter Doyle, and Veronica Wong. “An Exploration of Branding in Industrial Markets.” Industrial Marketing Management 26, no. 5 (1997): 433-446.
Nelson, Phillip. “Advertising as Information.” Journal of Political Economy 82, no. 4 (1974): 729-754.
Nelson, Phillip. “Information and Consumer Behavior.” Journal of Political Economy 78, no. 2 (1970): 311-329.
Sethuraman, Raj, Gerard J. Tellis, and Richard A. Briesch. “How Well Does Advertising Work? Generalizations from Meta-Analysis of Brand Advertising Elasticities.” Journal of Marketing Research 48 (June 2011): 457-471.
Shapiro, Bradley T., Günter J. Hitsch, and Anna E. Tuchman. “TV Advertising Effectiveness and Profitability: Generalizable Results from 288 Brands.” Econometrica 89, no. 4 (2021): 1855-1879.
Sharp, Byron. How Brands Grow: What Marketers Don't Know. Oxford University Press, 2010.
Strathern, Marilyn. “‘Improving Ratings’: Audit in the British University System.” European Review 5, no. 3 (1997): 305-321.

Industry research and surveys

The B2B Institute at LinkedIn, with WARC and Lions. The B2B Effectiveness Code. October 2021.
Buying Legal Council. 2021 Legal Procurement Survey. As reported in “Buying Legal Council Releases 2021 Legal Procurement Survey.” Corporate Counsel Business Journal, May 2021.
Edelman and LinkedIn. 2024 B2B Thought Leadership Impact Report.
“Gartner Sales Survey Finds 67% of B2B Buyers Prefer a Rep-Free Experience.” Press release, March 9, 2026.
“Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights.” Press release, May 20, 2026.
2025 B2B Buyer Experience Report. November 2025.
Dodd, David. “Research Round-Up: 6sense Study Provides Critical Insights on B2B Buyer Behavior.” CustomerThink, Nov. 30, 2025.

Reference

Quote Investigator. “Quote Origin: One-Half the Money I Spend for Advertising Is Wasted, But I Have Never Been Able To Decide Which Half.” April 11, 2022.
University of Oxford. “Your Academic Year.”

[the_ad_group id="15279"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

[taq_review]

[/exclude_from_rss] News Analysis - Leadership Beat

Does marketing drive revenue? An Oxford-style tutorial for eDiscovery and legal tech

A disciplined conversation

A method, then three terms

Michaelmas: the evidence

Hilary: the measure

Trinity: the market

The form in motion

Running it against your own work

The tutorials in full

Tutorial one, Michaelmas: "That marketing routinely takes credit for demand it did not create"

Tutorial two, Hilary: "That the honest test of marketing is what happens when it stops"

Tutorial three, Trinity: "That the shortlist is set before marketing can touch it"

An invitation

Sources and reference materials References are cited in full; titles link to a public version where one exists. ComplexDiscovery OÜ

ComplexDiscovery OÜ. “Complete look: ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup.” June 26, 2026.

Scholarship and commentary

Berman, Ron. “Beyond the Last Touch: Attribution in Online Advertising.” Marketing Science 37, no. 5 (2018): 771-792.
Binet, Les, and Peter Field. The Long and the Short of It. IPA, 2013.
Blake, Thomas, Chris Nosko, and Steven Tadelis. “Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment.” Econometrica 83, no. 1 (2015): 155-174.
Dawes, John. “Advertising Effectiveness and the 95-5 Rule: Most B2B Buyers Are Not in the Market Right Now.” Ehrenberg-Bass Institute for Marketing Science with The B2B Institute at LinkedIn, 2021.
Dixon, Matthew, and Brent Adamson. The Challenger Sale: Taking Control of the Customer Conversation. Portfolio/Penguin, 2011.
Drucker, Peter F. Management: Tasks, Responsibilities, Practices. Harper & Row, 1973.
Gordon, Brett R., Florian Zettelmeyer, Neha Bhargava, and Dan Chapsky. “A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook.” Marketing Science 38, no. 2 (2019): 193-225.
Johnson, Garrett A., Randall A. Lewis, and Elmar I. Nubbemeyer. “Ghost Ads: Improving the Economics of Measuring Online Ad Effectiveness.” Journal of Marketing Research 54, no. 6 (2017): 867-884.
Kirmani, Amna, and Akshay R. Rao. “No Pain, No Gain: A Critical Review of the Literature on Signaling Unobservable Product Quality.” Journal of Marketing 64, no. 2 (2000): 66-79.
Levitt, Theodore. “Marketing Myopia.” Harvard Business Review (July-August 1960): 45-56.
Lewis, Randall A., and Justin M. Rao. “The Unfavorable Economics of Measuring the Returns to Advertising.” Quarterly Journal of Economics 130, no. 4 (2015): 1941-1973.
Lodish, Leonard M., Magid M. Abraham, Jeanne Livelsberger, Beth Lubetkin, Bruce Richardson, and Mary Ellen Stevens. “A Summary of Fifty-Five In-Market Experimental Estimates of the Long-Term Effect of TV Advertising.” Marketing Science 14, no. 3, supplement (1995): G133-G140.
Maister, David H., Charles H. Green, and Robert M. Galford. The Trusted Advisor. Free Press, 2000.
Mudambi, Susan M., Peter Doyle, and Veronica Wong. “An Exploration of Branding in Industrial Markets.” Industrial Marketing Management 26, no. 5 (1997): 433-446.
Nelson, Phillip. “Advertising as Information.” Journal of Political Economy 82, no. 4 (1974): 729-754.
Nelson, Phillip. “Information and Consumer Behavior.” Journal of Political Economy 78, no. 2 (1970): 311-329.
Sethuraman, Raj, Gerard J. Tellis, and Richard A. Briesch. “How Well Does Advertising Work? Generalizations from Meta-Analysis of Brand Advertising Elasticities.” Journal of Marketing Research 48 (June 2011): 457-471.
Shapiro, Bradley T., Günter J. Hitsch, and Anna E. Tuchman. “TV Advertising Effectiveness and Profitability: Generalizable Results from 288 Brands.” Econometrica 89, no. 4 (2021): 1855-1879.
Sharp, Byron. How Brands Grow: What Marketers Don't Know. Oxford University Press, 2010.
Strathern, Marilyn. “‘Improving Ratings’: Audit in the British University System.” European Review 5, no. 3 (1997): 305-321.

Industry research and surveys

The B2B Institute at LinkedIn, with WARC and Lions. The B2B Effectiveness Code. October 2021.
Buying Legal Council. 2021 Legal Procurement Survey. As reported in “Buying Legal Council Releases 2021 Legal Procurement Survey.” Corporate Counsel Business Journal, May 2021.
Edelman and LinkedIn. 2024 B2B Thought Leadership Impact Report.
“Gartner Sales Survey Finds 67% of B2B Buyers Prefer a Rep-Free Experience.” Press release, March 9, 2026.
“Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights.” Press release, May 20, 2026.
2025 B2B Buyer Experience Report. November 2025.
Dodd, David. “Research Round-Up: 6sense Study Provides Critical Insights on B2B Buyer Behavior.” CustomerThink, Nov. 30, 2025.

Reference

Quote Investigator. “Quote Origin: One-Half the Money I Spend for Advertising Is Wasted, But I Have Never Been Able To Decide Which Half.” April 11, 2022.
University of Oxford. “Your Academic Year.”

[the_ad_group id="15279"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

Does marketing drive revenue? An Oxford-style tutorial for eDiscovery and legal tech

The one question that reveals whether your marketing plan is actually a plan

+ Analyze

Editor's Note: The eDiscovery Market Size Mashup is a research tool now in its fourteenth annual cycle. Since 2012, it has tracked the worldwide eDiscovery market through three structural eras: the early-cloud era, when subscription consumption began to displace perpetual licensing; the AI-assisted-review era, when predictive coding reset per-document review economics; and the demand-and-response era that defines 2025 through 2030, when generative-AI-assisted review and emerging agentic workflow features compress per-document cost against a data curve growing roughly five times faster than the dollars available to discover it.

The purpose has remained consistent across the cycles: to give legal technology executives, consultants, analysts, service providers, investors, and corporate legal teams a reconciled mid-range view of the market that is internally consistent, methodologically disclosed, and updated each year against the latest available data. The 2025 to 2030 cycle places the worldwide eDiscovery market at an estimated $19.61 billion in 2025 and a projected $28.08 billion by 2030, a reconciled 7.44 percent compound annual growth rate. Underneath the aggregate trajectory sits the central arithmetic of the cycle: a 27.6-percentage-point annual gap between data growth and market growth that compounds across five years into a 3.13-times productivity-per-dollar mandate by 2030. Each of the twelve Market Intelligence installments published across this cycle examined a single segmentation lens. The consolidated synthesis that follows brings those lenses together as one citable reference for procurement, capability planning, market analysis, and vendor-selection decisions through the back half of the decade. [exclude_from_rss]

[taq_review]

[/exclude_from_rss] Industry Research - eDiscovery Market Sizing Beat

Complete look: ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup

A synthesis of the worldwide eDiscovery market across software, services, deployment, geography, sector, delivery, task share, and the demand-side data growth curve, reconciled within the ComplexDiscovery OÜ eDiscovery Marketplace Mashup Model ComplexDiscovery OÜ Staff Two numbers shape worldwide eDiscovery through 2030, and they do not move at the same pace. The dollars to discover potentially relevant information rise from approximately 19.61 billion in 2025 to approximately 28.08 billion by 2030, a multiplier of 1.4 times. The data those dollars must reach rises from approximately 181 zettabytes in 2025 to approximately 812 zettabytes in 2030, a multiplier of 4.5 times. By the end of the decade, the same dollar must cover roughly 3.13 times more data than it does at the start. That arithmetic, larger than any single segment shift or composition change, is the structural force underneath this cycle of the mashup. The 3.13-times productivity-per-dollar mandate that compounds out of the gap is a primary force underneath much of what follows. It pressures software to take share from services as channel billing shifts toward AI-driven workflows. It pressures review's relative share of task spend to decline even as review dollars rise. It pressures cloud-first procurement to become the operational default. And it sits alongside other structural dynamics including subscription consumption migration, direct-buyer maturation, and supplier consolidation. None of those other dynamics is reducible to the productivity mandate alone, but each operates in a market where the mandate sets the demand-side ceiling. What follows is a tour of how the mandate lands across each segmentation lens explored in this cycle of the Market Intelligence series, from the aggregate market line down through software, services, deployment, cloud composition, geography, sector, delivery approach, task share, and the demand-side data growth curve underneath all of it.

The shape of the worldwide market

The reconciled view places the worldwide eDiscovery market at approximately 19.61 billion dollars in 2025, rising to approximately 28.08 billion dollars by 2030, a compound annual growth rate of approximately 7.44 percent. Software, the smaller of the two segments in absolute terms, grows at roughly 10.41 percent. Services, the larger segment, grows at roughly 5.75 percent. The 4.66-percentage-point segment CAGR gap translates into a 5-percentage-point composition shift across the five-year horizon. Software's share of total worldwide eDiscovery spend rises from 34 percent in 2025 to 39 percent in 2030; services' share falls correspondingly from 66 percent to 61 percent. Services remains the larger segment by absolute spend through 2030; the segment crossover point falls outside the 2025 to 2030 window.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/05/eDiscovery-Market-Sizing-Past-and-Projected-2026.pdf" title="eDiscovery Market Sizing - Past and Projected - 2026"] Chart: eDiscovery Market Sizing, Past and Projected (2012 to 2030)

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/eDiscovery-Software-and-Services-Market-2025-2030.pdf" title="eDiscovery Software and Services Market (2025-2030)"] Chart: eDiscovery Software and Services Market (2025 to 2030)

Software's outperformance is not a fluke of the moment. The 4.66-percentage-point CAGR gap is, in large measure, the segment-level expression of an AI-driven channel reallocation: the same review workflow that once generated services revenue at a per-document or per-hour rate increasingly generates software revenue at a SaaS subscription or AI-inference rate. The work has not disappeared, and in many cases has expanded as data volumes grow, but the channel through which the work gets billed has steadily shifted from services to software. Services growth is slower but structurally durable; cross-border data, regulatory exposure, advisory work, and specialized response work sustain the services line even as software automates discrete tasks. The services segment in 2030 will not be a slower-growing version of the services segment of 2025; it will be a different mix.

How the composition is changing

Within the software segment, the cloud-first transition that has been underway for the better part of a decade is now functionally complete for new deployments. Off-premise software grows from approximately 5.29 billion dollars in 2025 to approximately 8.87 billion dollars in 2030, while on-premise software grows from 1.37 billion to 2.08 billion. On-premise solutions persist where security, sovereignty, or contractual constraints require them, but they are no longer the default for new procurement.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/05/eDiscovery-Software-Market-2025-2030-On-Off-Premise.pdf" title="eDiscovery Software Market (2025-2030) - On + Off Premise"] Chart: eDiscovery Software Market, On and Off Premise (2025 to 2030)

The more interesting subplot is inside the cloud category. SaaS holds roughly two-thirds of cloud spend in 2025 (approximately 67 percent) and drifts to approximately 63 percent by 2030 as PaaS and IaaS components compound at faster rates. PaaS rises from 15 percent of cloud spend to 17 percent; IaaS rises from 18 percent to 20 percent. The reason is simple: as advanced eDiscovery workloads incorporate large-scale processing, AI inference, vector search, and complex data engineering, customers and providers are increasingly integrating directly with platform and infrastructure services. Some of that integration appears in vendor SaaS revenue, and some appears as direct PaaS or IaaS spend.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/eDiscovery-Cloud-Software-Market-2025-2030.pdf" title="eDiscovery Cloud Software Market (2025-2030)"] Chart: eDiscovery Cloud Software Market (2025 to 2030)

Services growth lags software growth, but the headline number understates the qualitative shift inside services. Traditional managed-review revenue faces continued pricing compression as AI-assisted review compresses billable hours. Advisory services (litigation readiness, information governance, AI risk advisory) grow on the back of regulatory complexity. Specialized response work (forensic collection, second-request response, cross-border data transfer, regulatory inquiry support) grows at premium rates. The services segment in 2030 will not be a slower-growing version of the services segment of 2025; it will be a different mix.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/05/eDiscovery-Services-Market-2025-2030.pdf" title="eDiscovery Services Market (2025-2030)"] Chart: eDiscovery Services Market (2025 to 2030)

Where the money is going

The United States continues to anchor the worldwide market, accounting for roughly 66 percent of global spend in 2025 and easing to roughly 64 percent in 2030. The shift toward rest-of-world is gradual but real, driven by the maturation of data protection regimes outside the United States, the internationalization of regulatory inquiries, and the gradual buildout of regional capacity. Within rest-of-world, the United Kingdom, Canada, Germany, Australia, and Japan continue to claim the largest sub-shares, with rising activity in Singapore, India, and parts of the Middle East.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/eDiscovery-Market-Geographical-Overview-2025-2030.pdf" title="eDiscovery Market Geographical Overview (2025-2030)"] Chart: eDiscovery Market Geographical Overview (2025 to 2030)

The reconciliation distinguishes between government and regulatory demand on one hand and non-government (private-sector) demand on the other. Both grow over the period; non-government grows faster in both percentage and absolute terms. Non-government growth reflects expansion in civil litigation, internal investigations, corporate compliance, and AI-related risk advisory work. Government and regulatory growth reflects persistent investigative activity, ongoing premerger notification work, parallel inquiries in the European Union and the United Kingdom, and continued cross-border regulatory coordination.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/eDiscovery-Government-and-Non-Government-Market-Overview-2025-2030.pdf" title="eDiscovery Government and Non-Government Market Overview (2025-2030)"] Chart: eDiscovery Government and Non-Government Market Overview (2025 to 2030)

Segmenting worldwide spend by who captures the direct economic transaction (corporations and governments, law firms, or service providers) reveals a clear picture. The corporations-and-governments category remains the dominant channel throughout the period, reflecting continued in-house consumption supplemented by direct vendor procurement. Service providers grow faster than law firms over the forecast period. Law firms increasingly act as orchestrators rather than primary procurement channels, with the work and the dollars flowing around them.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/eDiscovery-Market-By-Direct-Delivery-Approach-2025-2030.pdf" title="eDiscovery Market By Direct Delivery Approach (2025-2030)"] Chart: eDiscovery Market by Direct Delivery Approach (2025 to 2030)

The task shift, and why it matters

The most consequential structural shift in the industry is not visible in the aggregate market line. It is in the composition of where eDiscovery dollars get spent across the three core tasks of collection, processing, and review. Review remains the largest single task expenditure throughout the period, but collection and processing capture increasing absolute and relative shares. Across a longer horizon, from RAND Corporation's 2012 baseline through ComplexDiscovery OÜ's 2025 modeling and 2030 forecast, review's share of total task spend has fallen from 73 percent to a reconciled 62 percent to a projected 52 percent: a 21-percentage-point decline across 18 years. Collection, over the same span, has expanded over threefold, from 8 percent to a projected 25 percent, a 17-percentage-point gain. Processing has been comparatively stable, rising from 19 percent to 23 percent.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/05/eDiscovery-Market-By-Task-2025-2030.pdf" title="eDiscovery Market By Task (2025-2030)"] Chart: eDiscovery Market by Task (2025 to 2030)

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/Relative-Task-Expenditures-for-Core-eDiscovery-Tasks.pdf" title="Relative Task Expenditures for Core eDiscovery Tasks"] Chart: Relative Task Expenditures for Core eDiscovery Tasks (2012, 2025, 2030)

The pace of the rebalance is accelerating. Roughly 47 percent of the 18-year share movement happens in the final five years from 2025 to 2030. Review's 5-year decline of 10 percentage points nearly equals the prior 13-year decline of 11 percentage points. The trend is consistent with a demand-and-response dynamic rather than two independent forces operating in parallel. The demand side is the growth in data volume subject to potential collection. The supply-side response is AI-assisted review's compression of per-document review costs, with predictive coding through the prior decade, generative-AI-assisted review through the current decade, and emerging agentic workflow features as the next compression wave. Data growth raises absolute review work to be done; AI compression compresses per-document review prices. Whether the resulting absolute review spend rises, falls, or stays flat depends on the relative pace of the two. In the reconciled view, review absolute spend continues to grow modestly, from approximately 12.16 billion dollars in 2025 to approximately 14.60 billion dollars in 2030, but materially slower than the aggregate market, which is why review's share declines. For practitioners and providers, the practical consequence is that capacity decisions made today should anticipate the structural drift toward collection-heavy and processing-heavy task profiles.

The demand side, data growth as the underlying force

Worldwide data volumes are projected to grow from approximately 181 zettabytes in 2025 to approximately 812 zettabytes in 2030, a compound annual growth rate of approximately 35 percent. Enterprise-held data, the subset most relevant to discoverable information, expands from approximately 54 zettabytes to approximately 243 zettabytes over the same period at the same rate, holding steady at roughly 30 percent of the global total. The 181 zettabyte 2025 anchor is consistent with IDC's Global DataSphere baseline. The 35 percent CAGR through 2030 reflects the Mashup Model's reconciliation across data-universe forecasts and enterprise-specific projections, and sits on the upper end of published industry growth estimates. It is higher than IDC's headline total-data forecast trajectory and lower than the most aggressive AI-content-driven projections.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/06/Data-Volume-and-Growth-in-Zettabytes-2025-2030.pdf" title="Data Volume and Growth in Zettabytes (2025-2030)"] Chart: Data Volume and Growth in Zettabytes (2025 to 2030)

The 35-percent data CAGR set against the 7.44-percent eDiscovery market CAGR is the central arithmetic of the cycle. Across the five-year horizon, global data multiplies roughly 4.5 times (4.484x, from 181 to 812 zettabytes); the worldwide eDiscovery market multiplies roughly 1.4 times (1.432x, from 19.6 to 28.1 billion dollars). Divide one by the other and the productivity-per-dollar requirement falls out: 4.484 ÷ 1.432 ≈ 3.13. By 2030, the same dollar must process, store, search, review, and produce against roughly 3.13 times more data than it did in 2025 just to maintain the same coverage ratio. That is not a marginal improvement. It is the productivity mandate that defines the decade for the industry. The mandate sits underneath several of the shifts documented above: the segment-level CAGR gap reflects the channel through which the productivity gain flows; the task-share rebalance reflects where the gain lands at the workflow level. AI capability compounding (predictive coding through the prior decade, generative-AI-assisted review through the current decade, emerging agentic workflow features as the next compression wave) is the bridge that must close the gap. Whether the bridge closes the gap fully, partially, or in stages depends on the pace at which the current generation of tooling matures, the pace at which agentic features move from product roadmap to production deployment, and whether the industry holds to full-coverage discovery as the standard or moves toward risk-tiered coverage that reserves the most intensive workflows for the documents that matter most. The mandate sets the ceiling; the answer to how the ceiling gets met is the open question of the decade.

What the reconciled view implies

The reconciled view supports a small set of interpretive points, none of them prescriptive, all of them grounded in the figures. For software vendors, the 10.41 percent reconciled software CAGR outpaces services by 4.66 percentage points a year, and the vendors positioned to capture a disproportionate share of incremental software dollars are those integrating AI-assisted review, modular SaaS delivery, and platform-aware processing into the same product surface. For service providers, slower nominal growth does not imply a less attractive market. It implies a different one. Providers that reposition around higher-value advisory and specialized regulatory response can outpace the segment headline. For law firms, the modest share of direct economic transactions captured by law firms suggests a continued shift toward orchestration and advisory positioning rather than primary procurement. For corporate and government legal teams, in-house consumption continues to dominate the direct delivery approach. Build-versus-buy on internal capabilities, governance over AI use in discovery, and readiness for second-request response remain the central program-level questions. For investors and analysts, the dynamics support a continued investment thesis around cloud-native, AI-enabled software platforms, with margin pressure on traditional services and continued consolidation at the supplier level.

Closing the loop

At the start of this analysis, the cycle was framed around two numbers that do not move at the same pace: 19.61 billion dollars rising to 28.08 billion against 181 zettabytes rising to 812 zettabytes. The productivity-per-dollar mandate that compounds out of that gap, 3.13 times by 2030, is not a forecast of efficiency. It is a measurement of pressure. The segment, task, and channel shifts documented above are the visible places where the pressure shows up. AI capability compounding is the bridge that must close the gap. Whether the industry meets that pressure through AI-assisted tooling alone, or moves toward a structural redefinition of what discovery coverage means in practice, from full coverage of every potentially relevant artifact to risk-tiered coverage that reserves the most intensive workflows for the documents that matter most, is the open question of the decade. The mashup measures the pressure. The industry answers it.

About the Model behind these figures

All quantitative figures in this analysis are drawn from the ComplexDiscovery OÜ eDiscovery Marketplace Mashup Model, a proprietary analyst-aggregation framework that reconciles publicly available third-party research, vendor disclosures, and industry reference work into a single defensible mid-range view of the worldwide eDiscovery market. The Model is a research aggregation tool maintained by ComplexDiscovery OÜ. It is not distributed publicly and does not constitute primary research; figures cited here represent reconciled estimates aligned to a common scope, geography, and timeframe.

This Mashup is the public synthesis vehicle for the Model's 2025 to 2030 cycle. It provides the consolidated reconciliation across software, services, deployment, cloud composition, geography, sector, delivery approach, task composition, long-horizon task share, and data growth, along with a representative listing of the organizations and publications that inform the Model's source aggregation.

Methodology

The scope of this synthesis is the worldwide eDiscovery market, encompassing software and services, expressed in U.S. dollars, across calendar years 2025 through 2030. Reconciliation of varying market definitions, geographic scopes, and source methodologies is presented as ranges with assumptions disclosed where precise alignment is not possible. The 2012 task baseline cited in the long-horizon task share section derives from the RAND Corporation 2012 study by Pace and Zakaras. Compound annual growth rates are derived using the standard formula ((End divided by Start) raised to the power of one over the number of years) minus 1, with five years as the denominator unless otherwise noted. The 3.13-times productivity mandate is a coverage-flat ratio (data multiplier divided by market multiplier), not a forecast of realized productivity gains.

Citing this analysis

The primary citable resource for the figures and analysis in this article is the ComplexDiscovery OÜ eDiscovery Marketplace Mashup Model. The Model is the aggregated research artifact maintained by ComplexDiscovery OÜ since 2012 and is the appropriate citation when referencing data points, projections, segmentation, or analyses derived from any ComplexDiscovery annual eDiscovery market size mashup. Suggested citation: Robinson, R. (2026). 2025 to 2030 eDiscovery Market Size Mashup (H. Robinson, Ed.). ComplexDiscovery OÜ.

Sources informing the Model

The listing below provides an overview of the organizations and publications whose data points have informed the development of the Model over time. The Model itself aggregates publicly available content (including abstracts, excerpts, quotes, references, and data points) from these sources, with inputs collected since the inaugural ComplexDiscovery eDiscovery Market Size Mashup in 2012. Individual entries are presented for transparency about Model construction and are not the appropriate citation for figures appearing in this analysis; readers referencing figures should cite the Model. Market modeling rounding may result in slight differences in aggregate numbers.

360 Market Updates / 360iResearch
Aberdeen
ACG Partners
Allied Market Research
American Medical Association
BMC
Catalyst Investors
ComplexDiscovery OÜ eDiscovery Marketplace Mashup Model (incorporating industry news, editorial analysis, eDiscovery Business Confidence Surveys, eDiscovery Pricing Surveys, and prior Annual eDiscovery Market Size Mashups since 2012)
CS Disco
Discovery & Legal Technology Association (DLTA)
eDiscovery Journal
EY
Facts and Factors
Forbes
FRONTEO (UBIC)
Future Market Insights
Gartner
Georgetown Law Center for the Study of the Legal Profession and Thomson Reuters Legal Executive Institute
Global Industry Analysts
Grand View Research
Greentarget
Harvard Business Review
Houlihan Lokey
i360
IBIS World
IDC
Industry eDiscovery Provider, Analyst, and Investor Briefings and Discussions
Industry Observer Estimations (Multiple Observers)
Industry Research (Company)
KLDiscovery
Markets and Markets
Mordor Intelligence
Nasdaq
Nuix
P&S Market Research
Prescient & Strategic Intelligence
RAND Institute for Civil Justice
Relativity Fest, Industry Panel Discussions
Reports and Data
Research and Markets
Richmond Journal of Law and Technology
Statista
The Conference Board
The Radicati Group
Third-Party Market Studies (Independent Industry Briefings)
Transparency Market Research
U.S. Bureau of Economic Analysis
U.S. Department of Commerce, International Trade Administration
U.S. Securities and Exchange Commission (public company filings)
Zion Market Research

Market Intelligence series reports (2025-30 eDiscovery market size mashup)

[the_ad id="45753"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

[taq_review]

[/exclude_from_rss] Industry Research - eDiscovery Market Sizing Beat

Complete look: ComplexDiscovery OÜ's 2025 to 2030 eDiscovery market size mashup

The shape of the worldwide market

How the composition is changing

Where the money is going

The task shift, and why it matters

The demand side, data growth as the underlying force

What the reconciled view implies

Closing the loop

About the Model behind these figures

Methodology

Citing this analysis

Sources informing the Model

360 Market Updates / 360iResearch
Aberdeen
ACG Partners
Allied Market Research
American Medical Association
BMC
Catalyst Investors
ComplexDiscovery OÜ eDiscovery Marketplace Mashup Model (incorporating industry news, editorial analysis, eDiscovery Business Confidence Surveys, eDiscovery Pricing Surveys, and prior Annual eDiscovery Market Size Mashups since 2012)
CS Disco
Discovery & Legal Technology Association (DLTA)
eDiscovery Journal
EY
Facts and Factors
Forbes
FRONTEO (UBIC)
Future Market Insights
Gartner
Georgetown Law Center for the Study of the Legal Profession and Thomson Reuters Legal Executive Institute
Global Industry Analysts
Grand View Research
Greentarget
Harvard Business Review
Houlihan Lokey
i360
IBIS World
IDC
Industry eDiscovery Provider, Analyst, and Investor Briefings and Discussions
Industry Observer Estimations (Multiple Observers)
Industry Research (Company)
KLDiscovery
Markets and Markets
Mordor Intelligence
Nasdaq
Nuix
P&S Market Research
Prescient & Strategic Intelligence
RAND Institute for Civil Justice
Relativity Fest, Industry Panel Discussions
Reports and Data
Research and Markets
Richmond Journal of Law and Technology
Statista
The Conference Board
The Radicati Group
Third-Party Market Studies (Independent Industry Briefings)
Transparency Market Research
U.S. Bureau of Economic Analysis
U.S. Department of Commerce, International Trade Administration
U.S. Securities and Exchange Commission (public company filings)
Zion Market Research

Market Intelligence series reports (2025-30 eDiscovery market size mashup)

[the_ad id="45753"]

Assisted by GAI and LLM Technologies Additional reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

Complete look: ComplexDiscovery OÜ’s 2025 to 2030 eDiscovery market size mashup

Reset, shift or maturation: what 2026’s 16 eDiscovery deals suggest so far

Editor's Note: Generative AI is no longer a future-state concept in eDiscovery pricing; it is already reshaping how legal, technology, and corporate teams evaluate cost, value, and defensibility. In this Winter 2026 Pricing Pulse analysis, ComplexDiscovery OÜ, in partnership with EDRM, examines a market that is simultaneously stabilizing in traditional service categories and fragmenting in newer AI-driven ones. The findings highlight a clear divide between established pricing norms for forensic collection, processing, hosting, and document review, and the still-developing commercial models emerging around GenAI-assisted review. For cybersecurity, data privacy, regulatory compliance, and eDiscovery professionals, that divide matters. Pricing transparency now directly affects budgeting, vendor selection, matter planning, and risk management—especially as organizations weigh the promise of AI efficiency against unresolved questions around exception handling, quality control, and contract structure. This analysis offers a timely benchmark for understanding where the market stands today and where pricing pressure is likely to intensify next.

[exclude_from_rss]

[taq_review]

[/exclude_from_rss] Industry Research

A Complete Analysis of the Winter 2026 eDiscovery Pricing Survey

ComplexDiscovery Staff

Executive Summary

The Winter 2026 eDiscovery Pricing Survey, conducted by ComplexDiscovery OÜ in partnership with the Electronic Discovery Reference Model (EDRM) across December 2025 through February 2026, captures a market at a pivotal inflection point. Generative AI (GenAI) has moved into operational workflows for a significant and growing segment of the eDiscovery market — but adoption is uneven, pricing frameworks have not kept pace, and a meaningful share of practitioners have not yet engaged with AI-assisted review at any level. That bifurcation between early adopters and the rest of the market is itself one of the survey's defining findings. Drawing on 53 responses from legal professionals, technology providers, corporations, and consultancies, this survey provides a detailed pricing snapshot of the current eDiscovery market, spanning forensic collection, data processing and hosting, document review, and GenAI-assisted review. Several clear signals emerge from the data. Forensic collection and examination rates have stabilized in the $250–$350 per hour range for standard work, with premium rates for testimony and analysis. Data hosting has commoditized meaningfully at the infrastructure level, while analytics-enabled hosting retains pricing differentiation. Document review rates are stable but per-document billing remains opaque. Most critically, GenAI-assisted review pricing is experimentally diverse — hybrid models and per-document billing each claim roughly 28% of reported primary models, with the $0.11–$0.50 per-document range emerging as a competitive zone that directly challenges traditional human review economics. This report covers all 25 survey questions, organized into four thematic sections, with analyst observations and strategic implications throughout. All findings represent self-reported practitioner perceptions of prevailing market pricing — not verified transaction records — and should be read as directional market intelligence. Unlike vendor-produced or client-commissioned pricing guides, the Pricing Pulse is designed and published independently by ComplexDiscovery OÜ in partnership with the Electronic Discovery Reference Model (EDRM), with no commercial interest in any specific pricing outcome.

About the Survey

Survey Design and Purpose The Winter 2026 eDiscovery Pricing Survey was designed and administered by ComplexDiscovery OÜ in partnership with the Electronic Discovery Reference Model (EDRM) as part of its ongoing Pricing Pulse research program. The survey's primary purpose is to provide eDiscovery practitioners, technology providers, and legal operations professionals with empirically grounded pricing benchmarks across the key service categories that define the eDiscovery market. The Pricing Pulse is practitioner-reported and independently produced — it is not sponsored by, or designed to favor, any vendor, platform, or service category. Respondent comments critiquing the survey design itself are actively incorporated into future iterations, as reflected in this report's processing methodology note. This iteration of the survey placed particular emphasis on generative AI-assisted review pricing — a category first addressed formally in prior survey cycles and highlighted significantly in Winter 2026 to reflect the technology's accelerating, if uneven, integration into eDiscovery workflows. The five GenAI pricing questions (Questions 18–22) were designed to capture not just price points but pricing model structures, exception handling practices, and the nascent development of outcome-based pricing — recognizing that practitioners at very different stages of AI adoption would be responding. Respondent Profile The survey received 53 completed responses. By business segment, law firms represented the largest cohort at 43.4% (23 respondents), followed by software and/or services providers at 24.5% (13), corporations at 15.1% (8), consultancies at 9.4% (5), and media, research, or educational organizations at 7.5% (4). By primary function, 67.9% (36) identified as legal/litigation support professionals, 26.4% (14) as business or business support functions, and 5.7% (3) as IT or product development.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Survey-Respondents-by-Organizational-Segment-Winter-2026.pdf" title="Survey Respondents by Organizational Segment - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Survey-Respondents-by-Primary-Function-Winter-2026.pdf" title="Survey Respondents by Primary Function - Winter 2026"]

Geographically, the survey is overwhelmingly U.S.-centric: 92.5% of respondents (49) indicated North America – United States as their primary eDiscovery business geography, with the remaining 7.5% distributed across Europe (United Kingdom and non-UK) and Asia/Asia Pacific. This composition reflects the survey's community of practitioners and should be taken into account when applying results to non-U.S. markets.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Survey-Respondents-by-Geographic-Region-Winter-2026.pdf" title="Survey Respondents by Geographic Region - Winter 2026"]

The respondent pool's composition — heavily weighted toward legal practitioners with meaningful technology provider and in-house corporate representation — lends credibility to the pricing data for legal use cases while also surfacing supply-side perspectives from vendors who see pricing across many client engagements.

Section 1: Forensic Collection, Examination, and Testimony Pricing

Forensic collection and digital examination form the evidentiary foundation of eDiscovery. Unlike commoditized downstream services, forensic work depends on specialized expertise, defensible chain-of-custody protocols, and increasingly complex device environments. Mobile devices, cloud-linked data ecosystems, encrypted storage, and enterprise application footprints have expanded the examiner's scope considerably over the past several years, sustaining rate levels that resist the downward pressure more commoditized services face. Expert witness testimony sits at the highest value tier of forensic work — where practitioner credentials, courtroom experience, and legal exposure command significant premium pricing. Q1 & Q2 — Per Hour Cost for Onsite and Remote Collection The $250–$350 per hour range is the clear market anchor for forensic collection, cited by 56.6% of respondents for both onsite and remote collection. However, the distributions diverge meaningfully at the premium tier: 20.8% of respondents report onsite collection rates exceeding $350 per hour, compared to just 5.7% for remote. Conversely, remote collection skews lower — 18.9% report sub-$250 rates for remote work, versus only 5.7% for onsite. This onsite premium reflects real cost structures: travel, physical access logistics, on-premises security requirements, and the coordination burden of collecting in active enterprise environments. The growth of remote forensic collection tools — driven in part by pandemic-era necessity and now institutionalized in many engagements — has introduced competitive downward pressure on remote rates that onsite services do not face to the same degree. Four respondents (7.5%) indicate alternative pricing models for remote collection, suggesting some providers are moving toward flat-fee or subscription-based remote collection arrangements.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Hour-Cost-for-an-Onsite-Collection-by-a-Forensic-Examiner-Winter-2026-.pdf" title="Collection Pricing - Per Hour Cost for an Onsite Collection by a Forensic Examiner - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Hour-Cost-for-a-Remote-Collection-by-a-Forensic-Examiner-Winter-2026.pdf" title="Collection Pricing - Per Hour Cost for a Remote Collection by a Forensic Examiner - Winter 2026"]

Q3 & Q4 — Per Device Cost for Desktop/Laptop and Mobile Device Collection Device-based pricing skews decisively to the upper tier: 50.9% of respondents report per-device costs exceeding $350 for desktop and laptop collections, and 49.1% report the same for mobile devices. The $250–$350 mid-range captures 18.9% for computers and 24.5% for mobile devices — the higher mobile representation in the mid-range may reflect lower-complexity or volume-based mobile collection engagements where physical access is easier and device configurations are more standardized. Perhaps most notable is the convergence of mobile and computer collection pricing at the upper tier. Mobile device collection — once considered simpler than computer collection due to smaller storage capacities — now commands comparable rates as encryption, cloud sync architectures, third-party application data, and ephemeral messaging platforms have substantially increased examiner effort and risk. Practitioners seeking to budget mobile collection as a lower-cost alternative to computer collection will increasingly find the market does not support that assumption.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Device-Cost-for-a-Desktop-Laptop-Computer-Collection-by-a-Forensic-Examiner-Winter-2026.pdf" title="Collection Pricing - Per Device Cost for a Desktop Laptop Computer Collection by a Forensic Examiner - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Device-Cost-for-a-Mobile-Device-Collection-by-a-Forensic-Examiner-Winter-2026.pdf" title="Collection Pricing - Per Device Cost for a Mobile Device Collection by a Forensic Examiner - Winter 2026"]

Q5 — Per Hour Cost for Investigation, Analysis, and Report Generation Investigation, analysis, and report generation command a higher hourly rate floor than collection itself. More than half of respondents (54.7%) report rates in the $350–$550 range for this work, compared to the $250–$350 majority for collection. Only 30.2% report rates below $350 per hour for analysis, and 5.7% exceed $550. This premium reflects the cognitive and legal weight of analytical work. Forensic examiners producing reports that will be used in litigation, regulatory proceedings, or internal investigations are exercising expert judgment that creates professional liability — and the market prices that exposure accordingly. Practitioners purchasing forensic services should anticipate that billing rates will escalate from collection through analysis, often within the same engagement.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Hour-Cost-for-Investigation-Analysis-and-Report-Generation-by-an-FE-Winter-2026.pdf" title="Collection Pricing - Per Hour Cost for Investigation Analysis and Report Generation by an FE - Winter 2026"]

Q6 — Per Hour Cost for Expert Witness Testimony Expert witness testimony carries the highest rate profile in the forensic pricing group. While 47.2% report testimony rates in the $350–$550 range — consistent with analysis rates — a notable 26.4% report rates exceeding $550 per hour, the highest proportion in any >$550 category across the survey. The elevated 'do not know' response rate (20.8%) likely reflects that many practitioners engage forensic examiners for collection and analysis but not testimony, creating a meaningful gap in their pricing awareness for this segment. Expert witness rates are driven by factors beyond standard hourly billing — including the examiner's track record, publication history, geographic availability, and the complexity of the matter at issue. The wide distribution, from below $350 to above $550, reflects a market where individual credentials create significant pricing dispersion.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Collection-Pricing-Per-Hour-Cost-for-Expert-Witness-Testimony-In-Person-and-Written-by-an-FE-Winter-2026.pdf" title="Collection Pricing - Per Hour Cost for Expert Witness Testimony (In-Person and Written) by an FE - Winter 2026"]

Analyst Observation — Forensic Collection & Examination The forensic pricing landscape shows a well-established rate structure for collection and a predictable escalation through analysis to testimony. The $250–$350 range for collection hours serves as a reliable negotiation baseline. The key risk for buyers is underbudgeting for analysis and testimony phases — where rates routinely exceed $350/hour and frequently surpass $550. Practitioners with active litigation portfolios should establish explicit rate schedules with forensic vendors for all service tiers at engagement outset, not just collection. Key Takeaways — Section 1

$250–$350/hour is the market anchor for both onsite and remote forensic collection (56.6% each).
Onsite collection carries a measurable premium: 20.8% report >$350/hour vs. 5.7% for remote.
Mobile device collection rates have converged with computer collection at the upper tier (both ~50% report >$350/device).
Investigation, analysis, and report generation rates escalate to $350–$550/hour for 54.7% of respondents.
Expert witness testimony exceeds $550/hour for 26.4% — the highest proportion across all survey categories.

Section 2: Data Processing, Hosting, and Project Management Pricing

Data processing and hosting represent the operational infrastructure of eDiscovery delivery. Processing — transforming raw electronically stored information (ESI) into a reviewable format — has historically been a significant cost driver in large matters. Hosting provides the platform on which review takes place. Both categories have experienced significant commoditization pressure from cloud infrastructure economics, but the emergence of AI-driven early culling and processing tools is beginning to reshape volume dynamics in ways that affect both pricing and billing model design. Q7 & Q8 — Per GB Cost to Process ESI at Ingestion and at Completion Processing pricing at ingestion is relatively compressed: 39.6% of respondents report rates in the $25–$75 per GB range, and 34.0% report rates below $25 per GB. A significant 18.9% indicate alternative pricing models, reflecting the market's movement away from traditional per-GB ingestion billing. Processing pricing at completion of processing tells a different story. The most commonly reported range shifts to 'less than $100 per GB' (37.7%), and the proportion reporting alternative pricing models rises to 22.6%. Another 15.1% report $100–$150 per GB at completion, and 9.4% exceed $150 per GB. The jump from ingestion to completion reflects the data expansion and enrichment that occurs through native processing, deduplication, OCR, and promotion — processes that substantially increase the per-GB cost basis for providers. One respondent offered a methodologically important observation worth acknowledging directly: the survey's two-question processing model may conflate two distinct industry billing philosophies — an 'all-in' per-GB rate that covers ingestion through promotion, versus a staged model with separate per-GB charges for ingestion and native processing or promotion to review. This is a legitimate distinction, and practitioners benchmarking against these results should clarify which model their vendor employs. Future survey iterations will address this more precisely.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-Per-GB-Cost-to-Process-ESI-Based-on-Volume-at-Ingestion-Winter-2026.pdf" title="Processing Pricing - Per GB Cost to Process ESI Based on Volume at Ingestion - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-Per-GB-Cost-to-Process-ESI-Based-on-Volume-at-Completion-Winter-2026.pdf" title="Processing Pricing - Per GB Cost to Process ESI Based on Volume at Completion - Winter 2026"]

Q9 & Q10 — Per GB Per Month Cost to Host ESI Without and With Analytics Data hosting without analytics has substantially commoditized. More than half of respondents (54.7%) report hosting rates below $10 per GB per month, and another 30.2% fall in the $10–$20 range. Less than 2% report rates exceeding $20 per GB per month. This distribution reflects years of cloud infrastructure cost reduction passed through to buyers, as major platform providers compete on storage economics. Analytics-enabled hosting shows a wider and higher distribution. While 43.4% report rates below $15 per GB per month with analytics, 32.1% fall in the $15–$25 range, and 11.3% exceed $25 per GB per month. The premium for analytics-capable hosting reflects platform differentiation: vendors with mature AI search, conceptual clustering, visualization tools, and review workflow automation can sustain higher rates. Undifferentiated platforms — those competing primarily on storage price — face continued downward pressure as infrastructure costs decline. One respondent's comment corroborates this trajectory directly, observing that while overall eDiscovery pricing has been stable, technology costs specifically appear to be coming down — a signal consistent with the commoditization pattern visible in the hosting data.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-Per-GB-Cost-Per-Month-to-Host-ESI-without-Analytics-Winter-2026.pdf" title="Processing Pricing - Per GB Cost Per Month to Host ESI without Analytics - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-Per-GB-Cost-Per-Month-to-Host-ESI-with-Analytics-Winter-2026.pdf" title="Processing Pricing - Per GB Cost Per Month to Host ESI with Analytics - Winter 2026"]

Q11 — User License Fee Per Month for Access to Hosted Data User licensing is in an active state of structural transition. The $50–$100 per user per month range is the most frequently cited (41.5%), but a striking 34.0% of respondents report alternative pricing models — the highest alternative-model proportion among any category in the processing and hosting section. Only 17.0% report rates below $50 per user per month. The high alternative-model rate reflects a market shift away from traditional per-seat licensing toward enterprise agreements, volume tiers, and managed service arrangements that bundle access costs into broader contract structures. For corporate legal departments and law firms managing multi-matter eDiscovery portfolios, these bundled arrangements restructure cost visibility: per-matter spend attribution becomes less granular, which may simplify budgeting at the portfolio level but reduces transparency at the individual matter level. Whether bundled arrangements represent a net financial advantage depends on volume, negotiated terms, and how closely actual usage tracks the contracted scope — variables the survey does not measure.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-User-License-Fee-Per-Month-for-Access-to-Hosted-Data-Winter-2026.pdf" title="Processing Pricing - User License Fee Per Month for Access to Hosted Data - Winter 2026"]

Q12 — Per Hour Cost of Project Management Support for eDiscovery Project management pricing is the most consistent and well-understood category in the processing and hosting group. More than half of respondents (52.8%) report rates in the $100–$200 per hour range, and 26.4% report rates exceeding $200 per hour. The low 'do not know' rate (5.7%) — tied with Q9 for the lowest across all Section 2 questions — indicates that PM pricing is well understood by practitioners and regularly visible in vendor proposals. The 26.4% reporting greater than $200 per hour for project management likely reflects the growing complexity of modern eDiscovery engagements. Today's project managers must coordinate across AI review platforms, multiple review vendor relationships, technical review workflows, and real-time quality control functions — a scope considerably broader than the data management and platform coordination role the title suggested in prior market iterations.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Processing-Pricing-Per-Hour-Cost-of-Project-Management-Support-for-eDiscovery-Winter-2026.pdf" title="Processing Pricing - Per Hour Cost of Project Management Support for eDiscovery - Winter 2026"]

Analyst Observation — Processing, Hosting & Project Management Processing pricing is bifurcating: per-GB billing at ingestion remains common, but completion-phase and analytics-related pricing is shifting toward bundled and alternative models. Practitioners anchored to traditional per-GB benchmarks for TAR, analytics hosting, or managed service arrangements may be negotiating based on outdated frameworks. Hosting has genuinely commoditized at the infrastructure level — the pricing action now lives in analytics differentiation layered above the storage tier. Key Takeaways — Section 2

Processing at ingestion is largely below $75/GB (73.6% combined), but completion-phase pricing climbs with 24.5% reporting $100/GB or more.
Alternative pricing models account for 18.9% at ingestion and 22.6% at completion — signaling a structural shift away from per-GB processing billing.
Basic hosting has commoditized: 54.7% report sub-$10/GB/month. Analytics hosting retains differentiation with 11.3% exceeding $25/GB/month.
User licensing is migrating from per-seat to bundled models — 34.0% report alternative pricing structures.
Project management rates are well understood and rising: 26.4% now exceed $200/hour, reflecting growing engagement complexity.

Section 3: Document Review Pricing

Document review sits at the commercial center of most eDiscovery engagements. It is the largest cost driver in complex litigation, the primary arena in which human expertise meets technology leverage, and the category most directly disrupted by the emergence of GenAI-assisted review. Pricing in this section spans both hourly attorney rates (the traditional billing model) and per-document rates (a model that has gained traction as technology-assisted review has enabled higher throughput). The data in this section provides critical context for interpreting the GenAI pricing data that follows in Section 4. Q13 — Per GB Cost for Predictive Coding / Technology-Assisted Review Predictive coding and technology-assisted review (TAR) pricing has largely migrated away from per-GB billing. The highest single response category (35.8%) is 'alternative pricing model' — the highest alternative-model proportion of any per-GB question in the survey. Among those who do provide per-GB TAR pricing, 30.2% report rates below $75 per GB, 13.2% report $75–$150 per GB, and only 1.9% exceed $150 per GB. The 18.9% 'do not know' rate for TAR pricing suggests that many practitioners receive predictive coding as an embedded capability within their review platform subscription rather than a separately line-itemed service. This bundling trend, combined with the high alternative-model rate, indicates that standalone per-GB TAR billing is becoming the exception rather than the rule as platforms integrate AI-driven prioritization into standard hosting fees.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-GB-Cost-for-Predictive-Coding-in-a-Technology-Assisted-Review-Winter-2026.pdf" title="Review Pricing - Per GB Cost for Predictive Coding in a Technology-Assisted Review - Winter 2026"]

Q14 & Q15 — Per Hour Cost for Onsite and Remote Managed Review Attorneys Hourly managed review attorney rates are well understood and show a consistent onsite premium over remote delivery. For onsite review, 45.3% of respondents report rates exceeding $40 per hour, and 32.1% report $25–$40 per hour. For remote review, the distribution shifts: 41.5% report $25–$40 per hour, and 35.8% report greater than $40 per hour. The onsite premium reflects overhead recovery for physical review facilities, security infrastructure, and on-site supervision costs. Despite the normalization of remote review following the pandemic era, onsite review commands a persistent rate premium that clients with physical review requirements should anticipate. The relatively high 'do not know' rates for both onsite (18.9%) and remote (17.0%) suggest that many practitioners engage review vendors without direct visibility into the underlying attorney billing rates — a transparency gap that can make accurate matter budgeting difficult.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Hour-Cost-for-Document-Review-Attorneys-to-Review-Documents-Onsite-Winter-2026.pdf" title="Review Pricing - Per Hour Cost for Document Review Attorneys to Review Documents Onsite - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Hour-Cost-for-Document-Review-Attorneys-to-Review-Documents-Remote-Winter-2026.pdf" title="Review Pricing - Per Hour Cost for Document Review Attorneys to Review Documents Remote - Winter 2026"]

Q16 & Q17 — Cost Per Document for Onsite and Remote Managed Review Per-document billing for human document review carries significant uncertainty across the respondent pool. For onsite per-document review, 34.0% of respondents indicate they do not know the cost — the highest 'do not know' rate among all document review questions. For remote per-document review, 30.2% report not knowing. Among those with visibility, the $0.50–$1.00 per document range dominates for both onsite (30.2%) and remote (28.3%) delivery, with onsite showing a higher proportion of rates exceeding $1.00 per document (22.6% vs. 18.9% remote). Remote per-document review trends lower at the bottom of the range: 13.2% report sub-$0.50 rates for remote work versus only 3.8% for onsite. This directional difference is consistent with lower overhead costs in remote delivery environments. In this analyst's view, where the $0.50–$1.00 per-document rate for human review meets GenAI-assisted pricing in the $0.11–$0.50 range, the economic case for AI-assisted review becomes direct — provided quality and defensibility standards are met. The per-document rate distribution for human review is strategically important as a baseline against which GenAI-assisted review pricing should be evaluated. Where human review rates run $0.50–$1.00 per document and GenAI-assisted alternatives are priced in the $0.11–$0.50 range, the cost differential is substantial enough to drive adoption decisions — though the economic case ultimately depends on matter-specific quality thresholds and the degree to which AI exception handling costs are controlled.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Document-Cost-for-Document-Review-Attorneys-to-Review-Documents-Onsite-Winter-2026.pdf" title="Review Pricing - Per Document Cost for Document Review Attorneys to Review Documents Onsite - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Document-Cost-for-Document-Review-Attorneys-to-Review-Documents-Remote-Winter-2026.pdf" title="Review Pricing - Per Document Cost for Document Review Attorneys to Review Documents Remote - Winter 2026"]

Analyst Observation — Document Review Traditional document review rates have held relatively stable, but the market's increasing inability to articulate per-document pricing — particularly for onsite review — signals a structural shift away from document-count-based billing toward time-based models that are less directly comparable to AI-assisted pricing. Practitioners should push for per-document rate transparency in vendor proposals to enable genuine cost modeling against AI alternatives. Key Takeaways — Section 3

TAR/predictive coding billing is migrating away from per-GB models: 35.8% report alternative pricing, 18.9% don't know — bundled platform pricing is absorbing this cost.
Onsite managed review attorney rates exceed $40/hour for 45.3% of respondents vs. 35.8% for remote — the onsite premium persists.
Per-document review rates cluster in the $0.50–$1.00 range for both onsite and remote, with significant 'do not know' responses (34% onsite, 30.2% remote) indicating a transparency gap.
The $0.50–$1.00 per-document human review baseline sets up direct economic competition with emerging GenAI-assisted review pricing.

Section 4: GenAI-Assisted Review Pricing

The Winter 2026 survey's GenAI section was designed to illuminate where pricing clarity exists, where models are still fluid, and where the industry is beginning to form conventions around AI-assisted document review. What the results reveal is not a uniformly mature market but a bifurcated one: a segment of practitioners actively deploying and pricing GenAI review, and a substantial minority — 17.0% reporting it as not applicable or unknown — who have not yet engaged with it at a pricing level. Both cohorts are represented in the data, and the analysis in this section is relevant to each in different ways. This is not surprising. GenAI-assisted review introduces fundamentally different cost economics than traditional review: provider costs are driven by token consumption, GPU infrastructure, and model licensing — not attorney hours. Translating those costs into buyer-facing pricing structures that are transparent, predictable, and defensible has proven more difficult than the technology adoption itself. Q18 — Primary Model for GenAI-Assisted Review The two leading GenAI pricing models are effectively tied: hybrid pricing (combinations of multiple models) and per-document billing each account for 28.3% of primary model responses (15 respondents each). Per-GB billing captures 11.3%, per-token billing 5.7%, flat monthly subscription 5.7%, and outcome-based pricing 3.8%. Notably, 17.0% report that GenAI-assisted review pricing is not applicable or unknown to them — suggesting a meaningful share of the practitioner community has not yet engaged with AI review at a pricing level. The dominance of hybrid models reflects the reality that many providers are constructing bespoke proposals that combine per-document minimums, per-GB infrastructure charges, and platform subscription components. This complexity makes apples-to-apples comparison difficult for buyers — and may be intentional. Per-document pricing's co-equal standing with hybrid models suggests that a document-level unit of value is widely accepted as a conceptual billing anchor, even when the final structure is more complex. One respondent's comment illustrates the breadth of emerging structures not fully captured by the five survey model options: some providers are pricing GenAI review as an hourly professional service — with consultants performing query engineering, model interaction, and attorney collaboration — billed at standard hourly rates with per-matter minimums and not-to-exceed caps. This hourly professional service model sits outside the per-document or per-GB frameworks the market most commonly discusses, and its presence signals that GenAI pricing model diversity is wider than any single survey's categories can fully contain. Per-token pricing — the underlying cost reality for large language model deployments — has not been widely passed through to buyers (5.7%). This indicates that providers are currently absorbing token cost variability and presenting buyers with higher-order pricing units. As token costs evolve with model efficiency improvements, the degree to which providers pass these economics through will be an important market dynamic to watch.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/03/Review-Pricing-Primary-Model-for-Gen-AI-Assisted-Review-in-eDiscovery-Winter-2026.pdf" title="Review Pricing - Primary Model for Gen AI-Assisted Review in eDiscovery - Winter 2026"]

Q19 — Average Cost Per Document for GenAI-Assisted Review (Per-Document Model) Among all survey respondents, the $0.26–$0.50 per-document tier is the most frequently cited GenAI price point (20.8%), followed by both the $0.11–$0.25 and $0.05–$0.10 ranges (15.1% each). Seven and a half percent report per-document GenAI rates exceeding $0.50, and 5.7% report rates below $0.05. A significant 35.8% indicate this pricing model is not applicable to them or that they do not know the cost. The broad distribution among those with pricing visibility — from under a nickel to over fifty cents per document — reflects the wide variance in task complexity, model selection, and quality control overhead that different GenAI review implementations involve. The $0.11–$0.50 range represents the most commercially active zone. At the lower end, GenAI review offers compelling cost efficiency relative to the $0.50–$1.00 range for human per-document review. At the upper end of GenAI pricing (>$0.50), the value proposition requires stronger justification — particularly around accuracy, speed, or reduced downstream review burden. Practitioners should push vendors for specificity on what the per-document fee includes: model inference costs alone, or QC, exception handling, and reporting as well.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/03/Review-Pricing-Average-Cost-Per-Document-in-Per-Document-Model-of-Gen-AI-Assisted-Review-Winter-2026.pdf" title="Review Pricing - Average Cost Per Document in Per Document Model of Gen AI-Assisted Review - Winter 2026"]

Q20 — Average Cost Range for GenAI-Assisted Review (Per-GB Model) Per-GB GenAI pricing is less prevalent in practice — 64.2% of respondents indicate this model is not applicable or unknown. Among those who do report per-GB GenAI pricing, the $25–$50 per GB range is most common (17.0%), followed by below $25 per GB (13.2%). Two respondents (3.8%) report rates exceeding $100 per GB for GenAI review — likely representing specialized, computationally intensive analytical workflows rather than standard review acceleration. Given that data processing at ingestion typically falls below $75 per GB, a per-GB GenAI review charge layered on top represents a meaningful incremental cost. Practitioners evaluating per-GB GenAI pricing should model total matter economics carefully, including whether early data culling through AI reduces the volume that reaches review — potentially offsetting the per-GB GenAI charge with reduced processing and hosting costs downstream.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/03/Review-Pricing-Average-Cost-Range-Per-GB-in-Per-GB-Model-of-Gen-AI-Assisted-Review-Winter-2026.pdf" title="Review Pricing - Average Cost Range Per GB in Per GB Model of Gen AI-Assisted Review - Winter 2026"]

Q21 — Outcome-Based Pricing Structure for GenAI-Assisted Review Outcome-based pricing for GenAI review remains largely theoretical in the current market: 79.2% of respondents report no applicable experience with it. Among the minority with exposure, custom agreements dominate (9.4%), with small numbers reporting tiered pricing based on review speed improvements (3.8%), fixed fees based on achieved accuracy rates (3.8%), a combination of performance metrics (1.9%), and percentage of cost savings compared to traditional review (1.9%). The theoretical appeal of outcome-based pricing is clear — it aligns provider incentives with client results and distributes AI benefit-sharing in a transparent way. The operational mechanisms, however, remain underdeveloped. Defining accuracy baselines, attributing speed gains to AI versus staffing decisions, and calculating savings against hypothetical traditional review costs are all methodologically complex. The custom-agreement dominance (9.4%) reflects that outcome-based structures, where they exist, are negotiated on a bespoke basis without market-standardized frameworks. In this analyst's view, this is an area where the industry is likely to see active experimentation and standardization attempts in coming survey cycles — though the timeline will depend on how quickly buyers begin demanding performance accountability in AI review contracts.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/03/Review-Pricing-Typical-Structure-of-Outcome-Based-Pricing-Models-in-Gen-AI-Assisted-Review-Winter-2026.pdf" title="Review Pricing - Typical Structure of Outcome-Based Pricing Models in Gen AI-Assisted Review - Winter 2026"]

Q22 — How Pricing Models Handle Failed or Exception Documents in GenAI Review Exception document handling — documents that fail AI processing or require human intervention — is a practical and financially significant issue that is significantly underappreciated in headline GenAI pricing discussions. Nearly 40% of respondents (39.6%) cannot speak to how their contracts address this scenario. Among those with visibility, no single approach dominates: 18.9% report that exception documents route to manual review at standard rates; 17.0% say handling depends on the specific issue encountered; 9.4% each report that exceptions are charged as additional processing time or included in the base price (no additional charge); and 5.7% report per-document exception billing. The variability of exception handling approaches — and the high proportion of respondents with no visibility — represents a meaningful contract risk for buyers. In matters where a significant share of documents require human intervention, the effective cost of a GenAI-assisted review engagement can increase substantially depending on which exception pricing structure applies. Buyers negotiating GenAI review engagements should require explicit exception handling clauses that specify the triggering conditions, billing treatment, and quality control obligations for documents that exit the AI workflow.

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/03/Review-Pricing-Accounting-for-Docs-That-Fail-To-Process-or-Require-Special-Handing-Gen-AI-Winter-2026.pdf" title="Review Pricing - Accounting for Docs That Fail To Process or Require Special Handing (Gen AI) - Winter 2026"]

Analyst Observation — GenAI-Assisted Review The GenAI pricing market is operationally engaged but structurally immature. The concentration in hybrid and per-document models reflects practitioners and providers reaching for familiar pricing analogues while the technology matures. The $0.11–$0.50 per-document zone is emerging as a competitive market range — one that creates genuine economic pressure on traditional human review for appropriate document populations. The most important near-term challenge for the market is not the headline per-document or per-GB rate, but the hidden cost variables: exception document handling, quality control overhead, model retraining requirements, and the total cost of ownership of integrating GenAI review into existing workflows. One survey respondent offered a perspective worth placing on record: many vendors are still determining their AI pricing strategies, rushing to market to capture first-mover advantage or market share — and that token-based pricing pressures may cause AI solution costs to increase materially in the future, absent significant reductions in GPU infrastructure costs. This caution deserves attention as buyers evaluate multi-year GenAI review commitments. Key Takeaways — Section 4

Hybrid and per-document models are the dominant GenAI pricing structures, each at 28.3% — the market has converged on document-level units but not uniform delivery structures.
The $0.11–$0.50 per-document range is the emerging competitive zone for GenAI-assisted review, with direct economic implications for traditional human review.
Per-token pricing has not been widely passed to buyers (5.7%) — providers are absorbing LLM cost variability for now.
Outcome-based GenAI pricing is theoretically compelling but operationally undeveloped; 79.2% of respondents have no applicable experience.
Exception document handling is an underappreciated contract risk: 39.6% don't know how their agreements address it, and no standard approach has emerged.

Conclusion and Strategic Implications

The Winter 2026 eDiscovery Pricing Survey paints a picture of a market undergoing layered transitions simultaneously: forensic services have found stable pricing floors; processing and hosting have bifurcated between commoditized infrastructure and differentiated analytics tiers; document review is experiencing pricing model fragmentation as AI alternatives create new economic reference points; and GenAI-assisted review is operationally deployed but commercially immature in its pricing structures. For eDiscovery Buyers and Legal Operations Professionals The $250–$350 per hour range for forensic collection provides a reliable negotiation baseline, but buyers should build explicit rate schedules covering analysis and testimony phases — where rates routinely exceed $350 and frequently surpass $550 per hour. Processing and hosting negotiations should move beyond per-GB benchmarks for analytics-enabled and TAR-related services, where bundled models increasingly dominate. For document review, the critical action item is requiring per-document rate transparency even when hourly billing is the primary model — enabling genuine cost modeling against AI review alternatives. Corporate legal operations professionals face a distinct version of these challenges. Unlike law firms that pass eDiscovery costs to clients, in-house legal departments absorb them entirely — making pricing transparency a budget integrity issue, not just a negotiation tactic. The hosting commoditization finding (54.7% below $10/GB/month for basic hosting) and the user licensing transition (34.0% of respondents on alternative models) both represent leverage points in enterprise vendor negotiations that legal operations teams can use directly. The project management escalation finding (26.4% above $200/hour) warrants particular attention for in-house teams managing multi-matter portfolios: as PM rates rise with engagement complexity, the cost of inadequate internal scoping and vendor coordination compounds. Corporate legal operations teams are well-positioned to offset this by investing in internal eDiscovery program management capability rather than outsourcing all coordination to vendor project managers at premium rates. For GenAI-assisted review engagements, two contractual priorities stand out: first, obtain explicit pricing for exception documents rather than accepting provider discretion; second, require specificity on what is included in per-document or per-GB GenAI rates to enable accurate total-cost modeling. The $0.11–$0.50 per-document range is commercially viable for appropriate document populations, but hidden costs can erode that advantage quickly if not addressed in the agreement. For eDiscovery Service Providers and Technology Vendors The survey data confirms that buyers are engaging with GenAI pricing at a level of sophistication that requires providers to move beyond introductory pricing structures. The dominance of hybrid models reflects buyer uncertainty as much as provider flexibility — and that uncertainty is not sustainable as GenAI review becomes a standard engagement component rather than a premium add-on. Providers who develop clear, reproducible pricing structures with transparent exception handling will differentiate themselves in a market where 39.6% of buyers currently report no visibility into this critical cost variable. The trajectory of outcome-based pricing deserves attention. While only a small minority of respondents currently have exposure to these models, the direction of the market — toward accountability for AI review quality, not just delivery — suggests that providers who invest in outcome measurement frameworks now will be better positioned as client sophistication increases.

Looking Ahead: Open Questions for the Evolving eDiscovery Pricing Landscape

Several questions worth watching in future survey cycles: Will per-token pricing migrate from provider cost basis to buyer-facing billing as LLM economics become more visible? Will outcome-based pricing develop standardized frameworks, or remain bespoke indefinitely? Will the onsite/remote premium for forensic collection and attorney review compress as remote delivery tools mature further? And will the exception document handling gap in GenAI contracts become a litigation issue that forces market standardization? The Pricing Pulse series will continue to track these dynamics. The Winter 2026 results establish a pricing baseline at a pivotal moment — one that future surveys will be measured against as generative AI transforms both the economics and the practice of eDiscovery.

Research Methodology Note

The Winter 2026 eDiscovery Pricing Survey was designed and administered by ComplexDiscovery OÜ in partnership with the Electronic Discovery Reference Model (EDRM) as part of the Pricing Pulse research series. The survey was conducted via an online form distributed through ComplexDiscovery's professional community and partner networks. The survey period ran from December 2025 through February 2026, with the data collection window closing upon reaching the final respondent cohort of 53 individuals. The survey comprised 25 pricing questions organized across four service categories — forensic collection and examination, data processing and hosting, document review, and GenAI-assisted review — plus three respondent classification questions addressing geography, business segment, and primary function. Response options were structured as defined ranges rather than open-ended numeric inputs to facilitate comparative analysis and protect respondent pricing confidentiality. All responses represent self-reported market observations and practitioner experience. Results should be interpreted as directional market intelligence reflecting current practitioner perceptions of prevailing pricing, not as verified transaction records or audited benchmarks. The U.S.-centric geographic distribution (92.5%) should be taken into account when applying findings to non-U.S. markets. ComplexDiscovery OÜ maintains editorial independence in the analysis and publication of survey results. Individual respondent data is treated as confidential; only aggregated findings are reported. ComplexDiscovery and the Electronic Discovery Reference Model (EDRM) thank the 53 practitioners and professionals who contributed their time and market knowledge to this research. Organizations and individuals interested in participating in future Pricing Pulse surveys are encouraged to connect with ComplexDiscovery at complexdiscovery.com. © 2026 ComplexDiscovery OÜ. All rights reserved. Published on ComplexDiscovery.com. Conducted in partnership with the Electronic Discovery Reference Model (EDRM). The Pricing Pulse is an ongoing research series examining pricing dynamics across the eDiscovery market. News Source

Rob Robinson and Holley Robinson, ComplexDiscovery OÜ, "Winter 2026 eDiscovery Pricing Survey," February 2026.

[the_ad_group id="12741"]

Assisted by GAI and LLM Technologies Additional Reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.

[exclude_from_rss]

[taq_review]

[/exclude_from_rss] Industry Research

A Complete Analysis of the Winter 2026 eDiscovery Pricing Survey

ComplexDiscovery Staff

Executive Summary

About the Survey

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Survey-Respondents-by-Primary-Function-Winter-2026.pdf" title="Survey Respondents by Primary Function - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Survey-Respondents-by-Geographic-Region-Winter-2026.pdf" title="Survey Respondents by Geographic Region - Winter 2026"]

Section 1: Forensic Collection, Examination, and Testimony Pricing

$250–$350/hour is the market anchor for both onsite and remote forensic collection (56.6% each).
Onsite collection carries a measurable premium: 20.8% report >$350/hour vs. 5.7% for remote.
Mobile device collection rates have converged with computer collection at the upper tier (both ~50% report >$350/device).
Investigation, analysis, and report generation rates escalate to $350–$550/hour for 54.7% of respondents.
Expert witness testimony exceeds $550/hour for 26.4% — the highest proportion across all survey categories.

Section 2: Data Processing, Hosting, and Project Management Pricing

Processing at ingestion is largely below $75/GB (73.6% combined), but completion-phase pricing climbs with 24.5% reporting $100/GB or more.
Alternative pricing models account for 18.9% at ingestion and 22.6% at completion — signaling a structural shift away from per-GB processing billing.
Basic hosting has commoditized: 54.7% report sub-$10/GB/month. Analytics hosting retains differentiation with 11.3% exceeding $25/GB/month.
User licensing is migrating from per-seat to bundled models — 34.0% report alternative pricing structures.
Project management rates are well understood and rising: 26.4% now exceed $200/hour, reflecting growing engagement complexity.

Section 3: Document Review Pricing

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Hour-Cost-for-Document-Review-Attorneys-to-Review-Documents-Remote-Winter-2026.pdf" title="Review Pricing - Per Hour Cost for Document Review Attorneys to Review Documents Remote - Winter 2026"]

[pdf-embedder url="https://complexdiscovery.com/wp-content/uploads/2026/02/Review-Pricing-Per-Document-Cost-for-Document-Review-Attorneys-to-Review-Documents-Remote-Winter-2026.pdf" title="Review Pricing - Per Document Cost for Document Review Attorneys to Review Documents Remote - Winter 2026"]

TAR/predictive coding billing is migrating away from per-GB models: 35.8% report alternative pricing, 18.9% don't know — bundled platform pricing is absorbing this cost.
Onsite managed review attorney rates exceed $40/hour for 45.3% of respondents vs. 35.8% for remote — the onsite premium persists.
Per-document review rates cluster in the $0.50–$1.00 range for both onsite and remote, with significant 'do not know' responses (34% onsite, 30.2% remote) indicating a transparency gap.
The $0.50–$1.00 per-document human review baseline sets up direct economic competition with emerging GenAI-assisted review pricing.

Section 4: GenAI-Assisted Review Pricing

Hybrid and per-document models are the dominant GenAI pricing structures, each at 28.3% — the market has converged on document-level units but not uniform delivery structures.
The $0.11–$0.50 per-document range is the emerging competitive zone for GenAI-assisted review, with direct economic implications for traditional human review.
Per-token pricing has not been widely passed to buyers (5.7%) — providers are absorbing LLM cost variability for now.
Outcome-based GenAI pricing is theoretically compelling but operationally undeveloped; 79.2% of respondents have no applicable experience.
Exception document handling is an underappreciated contract risk: 39.6% don't know how their agreements address it, and no standard approach has emerged.

Conclusion and Strategic Implications

Looking Ahead: Open Questions for the Evolving eDiscovery Pricing Landscape

Research Methodology Note

Rob Robinson and Holley Robinson, ComplexDiscovery OÜ, "Winter 2026 eDiscovery Pricing Survey," February 2026.

[the_ad_group id="12741"]

Assisted by GAI and LLM Technologies Additional Reading

Source: ComplexDiscovery OÜ

ComplexDiscovery’s mission is to enable clarity for complex decisions by providing independent, data‑driven reporting, research, and commentary that make digital risk, legal technology, and regulatory change more legible for practitioners, policymakers, and business leaders.