By Sid Newby | February 2026
After 20+ years building and deploying technology inside the U.S. litigation machine, I've watched the same pattern repeat: powerful tools get built, powerful firms buy them, and everyone else makes do with keyword searches and prayer. The incumbents haven't failed at innovation -- they've succeeded at gatekeeping. AI is about to change that, and I think it's long overdue.
Where business litigation still bleeds money
In U.S. business litigation, "legal technology" is still, in practice, a cost-allocation machine. It decides how quickly parties can transform electronically stored information (ESI) into a defensible production, and how expensive that translation will be. The governing discovery framework -- Federal Rule of Civil Procedure 26(b)(1), as amended in 2015 -- explicitly recognizes that some ESI sources are "not reasonably accessible" because they impose undue burden or cost, and it expects parties and courts to grapple with those burdens early rather than treating everything as automatically collectible.[1]
That framing matters because modern discovery is not "finding documents." It's moving through a lifecycle that starts at identifying potentially relevant ESI and ends with producing responsive, non-privileged information -- via iterative steps that include collection, processing, review, analysis, and production. That lifecycle (often summarized in the Electronic Discovery Reference Model) is simple to draw but operationally brutal at scale when the data is messy, multi-lingual, privileged, and spread across SaaS systems.[2]
Figure 1: The EDRM lifecycle. Document review (highlighted) consumes the vast majority of production spend.
The economic center of gravity inside that lifecycle remains document review. RAND Corporation studied litigant expenditures for producing electronic discovery across 57 large-volume matters and found review typically consumed about 73% of production costs, dwarfing collection (~8%) and processing (~19%).[3] Just as importantly for the law-firm business model, RAND reported that outside counsel consumed about 70% of total e-discovery production costs -- suggesting discovery spending is not only a vendor problem, but a labor-and-leverage problem.[3]
Figure 2: Where the money goes. Source: RAND Corporation, MG-1208 (2012).
Let me say this plainly, from two decades of watching it happen: the review phase is where fortunes are made and smaller parties are broken. The ABA estimates document review accounts for over 80% of total litigation spend -- roughly $42 billion per year industry-wide.[18] A mid-size company facing a contract dispute doesn't lose because it has a bad case. It loses because it can't afford to find out whether it has a good one. The median cost of a contract dispute for a small or mid-size business is around $91,000 -- and for about 39% of organizations under $100M in revenue, a single matter costs less than $50,000, which sounds manageable until you realize that discovery alone can consume 20-50% of total litigation expenses.[19]
Modern proportionality doctrine is supposed to be the pressure relief valve. The 2015 revisions to Rule 26(b)(1) restored proportionality factors to the scope definition and emphasized that the change was not intended to shift the entire proportionality burden onto the requesting party, nor to let responding parties hide behind boilerplate objections.[1] In other words: the rules anticipate that cost and feasibility will be argued, evidenced, and negotiated -- because the system assumes excess is otherwise the default.
The incumbent stack that still runs discovery and legal research
Today's "big-player" litigation stack splits into two overlapping layers:
- eDiscovery/review platforms that host and operationalize the document universe for productions, privilege decisions, and issue workups
- Legal research and analytics platforms that monetize authoritative corpora and editorial enhancements for case law, statutes, and practical guidance
On the eDiscovery side, platform selection inside law firms and legal departments routinely converges on a familiar shortlist. An ILTA blog post describing a platform evaluation listed five cloud contenders: Casepoint, DISCO, Everlaw, Relativity (RelativityOne), and Reveal.[4] That list matches what third-party award ecosystems tend to surface as well.
This is a market measured in the tens of billions: the global eDiscovery market reached an estimated $20.7 billion in 2026 and is projected to hit $46 billion by 2034.[20]
On the legal research side, the incumbents are even more entrenched because their differentiation is tied to proprietary content, citators, editorial enhancements, and trust positioning. Thomson Reuters (Westlaw) is publicly emphasizing "professional-grade tools built on quality content and deep subject-matter expertise" as AI competition intensifies -- and in August 2025 launched CoCounsel Legal, featuring agentic AI and "Deep Research" capabilities built on Westlaw and Practical Law.[5] Market participants treat the legal information segment as directly exposed to AI disruption: when Anthropic released Claude Cowork plug-ins (including a legal plug-in) on January 30, 2026, RELX and Wolters Kluwer -- both major legal analytics providers -- saw their shares fall sharply amid disruption fears.[6]
| Company | Stock Impact (Jan 30-31, 2026) |
|---|---|
| Thomson Reuters | -18% |
| RELX (LexisNexis parent) | -14% |
| Wolters Kluwer | -13% |
| London Stock Exchange Group | -8%+ |
Table 1: Market reaction to Anthropic's Claude Cowork legal plug-in launch. Sources: Morningstar, Globe and Mail.[^6]
The practical effect of this stack dominance is that most serious business litigations are already "platform wars" underneath the pleading and motion practice: whichever side can filter faster, review more accurately, and build a provable narrative earlier gets leverage in meet-and-confer negotiations, dispositive motions, and settlement posture. That is precisely why the current stack's inefficiencies -- and its pricing -- matter so much.
I've spent my career in this space, and here's what I know to be true: the tools work, but they work for the people who can afford them. That's not a feature. That's a bug in the justice system.
Why the dominant tools still feel inefficient and expensive
What frustrates practitioners is not that eDiscovery tools don't work; it's that the workflow architecture still assumes massive human throughput at the most expensive step. RAND's breakdown -- review at ~73% of production cost -- is a quantification of what litigators experience daily: review labor dominates, and platform costs often scale with volume (data hosted, processed, or reviewed), not with outcome quality.[3]
Three structural inefficiencies show up repeatedly in big-matter reality.
1. Data sources keep multiplying
Many are now collaboration-native (chat, channels, comments, reactions, embedded files) rather than document-native. Even the "default" enterprise ecosystem now includes Exchange Online, OneDrive, SharePoint, and Teams -- all of which Microsoft Purview eDiscovery explicitly targets as discoverable content locations.[7] This pushes review platforms into constant schema adaptation and increases the odds that "collection completeness" and "context reconstruction" become the real disputes.
2. Evidence work is still separated into rigid phases
The EDRM framing explicitly acknowledges an end-to-end, iterative process.[2] Yet in many teams, the tools, staffing models, and budgets still treat these steps as baton passes between different specialists and vendors. The handoffs themselves are expensive: they generate duplicated project management, repeated QC cycles, and slow "time to first insight" -- which is where litigation leverage is often won or lost.
3. "Defensibility" requirements push teams toward conservative workflows
Rule 26 doctrine anticipates detailed burden arguments, sampling, and court involvement when parties can't agree on accessibility and proportionality.[1] That legal reality encourages a preference for tools and processes that can be narrated cleanly to a court -- even when those processes are not computationally optimal. In practice, many review teams accept slower, more manual steps because they can explain them.
This is why generative AI is so threatening to incumbents: it doesn't merely promise "faster keyword search." It targets the heart of what makes matters expensive -- review labor -- and it does so in a way that can be packaged as an auditable workflow rather than an ad hoc shortcut.
The AI shockwave: incumbents adopt GenAI while foundation models pressure the edges
By early 2026, the industry's posture has flipped from "should we allow generative AI?" to "how do we deploy it without malpractice risk?"
The ABA's technology survey found 31% personal use of generative AI at work among legal professionals (up from 27% the prior year), while firm-wide adoption lagged -- and varied dramatically by firm size. Firms with 51+ lawyers reported 39% adoption; firms with 50 or fewer sat at roughly 20%.[8] That gap is the access problem in miniature: the large firms get the tools first.
Meanwhile, insiders are already predicting billing disruption. The 2025 eDiscovery Innovation Report (produced with ACEDS and ILTA) reported that 90% of respondents believed GenAI either already has or will significantly alter conventional billing practices within two years. The same survey found 42% of respondents saving 1-5 hours per week with GenAI -- equivalent to 32.5 working days annually redirected to higher-value work.[9]
Against that backdrop, the incumbents are racing to industrialize "defensible GenAI" inside the platforms where review already happens:
- Relativity announced that aiR for Review and aiR for Privilege would be included in standard RelativityOne pricing effective April 1, 2026 -- explicitly framed as removing "barriers" to adoption. Since launching one year ago, 200+ customers have used these tools to review over 25 million documents. Legal teams report cutting review time by as much as 80%.[10]
- DISCO positions Auto Review as a GenAI-first-pass system that uses "tag descriptions" as instructions and produces explanations for its decisions. DISCO claims throughput of up to 32,000 documents per hour and has published case studies showing 96.9% recall and 70.1% precision on real matters.[11]
- Everlaw describes Deep Dive as a GenAI-driven case preparation tool built to answer questions only when relevant content is found, using a secure RAG framework with citations back to the corpus. It has been proven to handle cases exceeding 10 million documents.[12]
Figure 3: The three-layer competitive landscape. Foundation models (top) pressure both the platform layer and the content/research layer.
And it's not just incumbents adding AI features. Harvey AI -- the most prominent AI-native legal entrant -- raised $806 million across six funding rounds in 2025 alone, reaching an $8 billion valuation by December, with backing from Sequoia, Andreessen Horowitz, and OpenAI's Startup Fund.[21] Legal tech as a whole attracted $5.99 billion in funding in 2025, with fourteen rounds exceeding $100M.[22] That's not incremental investment. That's a signal that capital markets believe the legal industry's cost structure is about to break.
At the same time, the source-of-model layer is moving up the value chain. The most visible new pressure point is Anthropic's move into "legal workflow" automation via Claude Cowork plug-ins. The launch triggered an immediate market repricing of legal information incumbents.[6] The legal plug-in itself handles tasks like reviewing legal documents and generating briefings -- 11 "starter" plugins were open-sourced on GitHub as customizable templates.[13]
On the OpenAI side, the technical trajectory that matters for litigation is not "chatbots" but long-context reasoning. GPT-4.1 supports up to 1 million tokens of context (up from 128,000 in prior GPT-4o models), explicitly positioned for processing large volumes of long documents.[14] That capability -- combined with embedding-based semantic retrieval -- makes "stuff the entire deposition transcript into the prompt" a real architectural option, not a fantasy.
The accuracy question is resolving faster than expected
For years, the knock on AI-assisted review was accuracy. That objection is eroding quickly. A Vals AI benchmark study (with participants including Harvey, Thomson Reuters CoCounsel, and others, and test cases created by firms like Reed Smith and Fisher Phillips) found that the top AI tool produced a reliable first draft 73.3% of the time -- compared to 70% for the top human lawyer.[23] In document review specifically, TAR (Technology-Assisted Review) systems have consistently demonstrated ~77% recall versus ~60% for manual review, with similar precision advantages.[24] Meanwhile, human reviewers show error rates of 10-20%, and the classic Blair and Maron study found that professionals retrieved just 20% of relevant documents while believing their recall exceeded 75%.[24]
This doesn't mean AI is infallible. General-purpose chatbots still hallucinate on legal queries at alarming rates. But purpose-built legal AI tools operating within controlled review environments are now demonstrably competitive with -- and in many cases superior to -- human review accuracy. The question has shifted from "can AI do this?" to "can we audit and defend it?"
Two constraints prevent an immediate "AI replaces Westlaw/Relativity" narrative
First, ethics and reliability. The ABA's formal ethics guidance on generative AI emphasizes confidentiality, accuracy, and the lawyer's responsibility to supervise and verify outputs.[8] The caution is not academic: in Mata v. Avianca, Judge P. Kevin Castel imposed a $5,000 sanction on attorneys who submitted filings containing fictitious AI-generated case citations. ChatGPT had fabricated the cases entirely, and even falsely assured the lawyers they could be found on LexisNexis and Westlaw.[15]
Second, content rights and defensible sourcing. In Thomson Reuters v. Ross Intelligence (D. Del., Feb. 2025), the court found Ross was not permitted to copy Thomson Reuters' Westlaw headnotes to build a competing AI-based legal platform, rejecting the fair use defense. The court found Ross' product was a "market substitute" for Westlaw.[16] That kind of ruling strengthens the incumbents' gatekeeping advantage: even if an LLM can reason well, it still needs licensed, authoritative corpora to produce outputs litigators will rely upon.
A "perfect case" scenario: smart embeddings, long context windows, and agentic review
Here is what a high-performing, defensible, near-future discovery-and-research pipeline plausibly looks like by 2027-2028, given what vendors and model providers are already shipping.
Context-first ingestion and normalization (hours, not weeks)
Discovery begins with connectors that pull from collaboration and content systems into a workspace where chain-of-custody, audit logs, and deduplication live as default objects -- not "project management chores." Microsoft Purview already centrally targets Exchange Online, Teams, OneDrive, and SharePoint as first-class eDiscovery sources.[7]
Embedding-native retrieval: keyword is still there, but it's no longer the steering wheel
The first search interface isn't a Boolean query builder; it's a semantic retrieval layer that maps requests ("communications about revenue recognition issues in Q2") into vector queries and returns clusters plus exemplars, with keyword filters used mainly for constraint and defensibility. Embedding models enable semantic search that captures meaning and relationships in text, improving concept matching compared to traditional keyword approaches.[14]
A crucial practical shift follows: instead of sampling documents to design search terms, teams sample semantic neighborhoods to design issue tags, responsiveness definitions, and privilege heuristics. That flips early case assessment from "guess-and-check keyword lists" to "inspect the latent structure of the corpus" -- which is closer to how litigators actually think about stories, actors, and timelines.
Long-context review and synthesis: fewer chunks, fewer lost facts
Long context windows reduce the chunking tax that has historically made legal GenAI feel brittle. GPT-4.1 processes up to 1 million tokens; Anthropic's Claude similarly supports extended context windows for dense document analysis.[14]
The "perfect case" workflow uses long context selectively:
- A long-context pass for high-value "narrative artifacts" (key custodians, key threads, board decks, investigation reports, deposition transcripts)
- Embedding retrieval + targeted context windows for the rest (because cost-per-token and latency still matter at million-document scale)
Agentic document review: review becomes "policy execution with audit logs"
The most disruptive change is that first-pass review stops being "hours of human eyes" and becomes "machine execution of a review protocol," with humans shifting to quality control, exception handling, and strategy.
DISCO's Auto Review is essentially this model: legal teams define tag descriptions (protocol logic), the system applies them at scale, and outputs include explanations to support QA and defensibility.[11] Relativity's aiR family is framed similarly: accelerate review and privilege decisions inside a controlled environment.[10] Everlaw's Deep Dive -- answer with citations back to the corpus -- illustrates the same direction on the strategy side: move senior attorney thinking earlier, backed by traceable links.[12]
Figure 4: The staffing model shift. The expensive middle layer (contract review armies) is where AI delivers the most impact.
The technical "perfect" here is not magical accuracy. It is measurable, monitorable behavior: calibrated confidence thresholds, consistent sampling, adjudication workflows, and audit trails that can be narrated to a court. That emphasis tracks bar guidance insisting lawyers remain responsible for work product and must verify outputs.[8][15]
The business model collision: how AI could reprice litigation for everyone
If review truly dominates production cost -- and outside counsel captures much of that spend -- then automating first-pass review is not a marginal efficiency. It is a structural repricing of how business litigation is staffed and billed.
RAND's findings (review ~73% of production costs; outside counsel ~70% of eDiscovery production costs) imply that reducing review hours hits the part of the system that has historically been monetized through leverage: armies of reviewers supervised by a smaller number of senior lawyers.[3]
The industry already expects billing pressure. The 2025 eDiscovery Innovation Report shows 90% of respondents expecting conventional billing practices to be significantly altered within two years.[9]
This has two near-term consequences that can coexist.
Downward pressure on "review as revenue," upward pressure on "strategy as product"
When GenAI systems can perform large portions of first-pass relevance and issue coding faster than humans, the billing center shifts from time spent classifying documents to time spent deciding what matters and how to use it. Firms that adapt will capture value through fixed fees, phased budgets, or premium pricing for faster strategic outcomes -- not for the mere existence of review labor.
A real opening for smaller players -- if they can be defensible
If the "minimum viable discovery" for many disputes becomes (a) automated ingestion from core systems, (b) embedding-based retrieval, and (c) agentic review with auditable QC, then smaller firms and smaller organizations can compete on speed-to-insight without needing the same headcount. Relativity's decision to include key GenAI tools in standard packaging is explicitly framed as removing adoption barriers.[10]
This is the part I care most about. I've spent two decades watching smaller parties get crushed -- not because they were wrong, but because they couldn't afford to prove they were right.
The access to justice gap in this country is staggering. According to the Legal Services Corporation, approximately 80% of low-income Americans cannot afford legal assistance, and 92% of those surveyed didn't get adequate legal help for problems that substantially impacted their lives. Seventy-four percent of low-income households experienced at least one civil legal problem in the past year; 39% experienced five or more.[17] And this isn't just a poverty problem -- 40-60% of the legal needs of middle-class Americans go unmet, caught between making too much for legal aid and too little for private representation.[25]
This isn't just a consumer legal problem -- it extends to small and mid-size businesses that can't afford the six- and seven-figure eDiscovery bills that come with serious commercial disputes. When the median organizational spend on outside legal services is $1.8 million (and the top 25% spend at least $11.2 million annually), smaller companies facing a single large dispute are playing a game they can't afford to enter.[19]
If AI can compress the cost of competent document review by an order of magnitude, it doesn't just make BigLaw more profitable. It makes the justice system more accessible. That's not a technology story. That's a civil rights story.
But the danger is real: "democratization" can also mean "cheap mistakes at scale." Ethics guidance and real-world sanctions demonstrate the downside of unaudited AI outputs -- especially when hallucinated citations reach a tribunal.[8][15] And on the research side, copyright friction can limit how far a model-only entrant can go without licensing agreements; the Thomson Reuters v. Ross ruling signals an incumbent-friendly legal environment for protecting proprietary legal content.[16]
What "winning" likely looks like in the next few years
The most plausible equilibrium is not "incumbents die" or "LLMs replace lawyers." It is a re-bundling:
Figure 5: The re-bundling landscape. Incumbents with deep content moats (right) remain strong but face pressure to integrate AI (upward). Foundation models (top-left) need content partnerships to move right.
- Incumbents keep advantages where proprietary content, security controls, and defensibility are the product (legal research corpora, audit logging, standardized workflows).[5][16]
- Model providers and agent frameworks pressure margins by making high-quality reasoning and long-context analysis broadly available, forcing platforms to justify their pricing through end-to-end workflow value rather than "access to search."[6][14]
- Litigation teams reorganize around fewer reviewers and more "review architects": lawyers and litigation support specialists who translate legal theories into machine-executable protocols and defend the results with sampling and reporting.[10][11]
In that world, the "perfect case" is not just cheaper. It is faster to truth: faster identification of the real story, faster proportionality arguments grounded in actual corpus structure, and faster settlement decisions based on merits rather than review backlog -- precisely the kind of efficiency the rules framework has been trying to encourage for years.[1][2]
The question isn't whether AI will transform litigation technology. It's whether the transformation will be captured by the same incumbents who've profited from inaccessibility for decades -- or whether it will finally open the courthouse door a little wider.
From where I sit, after 20 years of watching this industry resist change, I'm cautiously optimistic that this time is different. The technology is too good, the cost pressure is too real, and the access gap is too wide. The incumbents will adapt and survive -- but they won't be able to keep the gates locked anymore.