AI in Legal Discovery: How eDiscovery Is Changing in 2026
Legal discovery has always been a data problem. Modern litigation, regulatory investigations, and merger reviews can involve reviewing millions of documents. Before AI, law firms and legal departments hired contract review attorneys to read those documents, find the relevant ones, identify privileged communications, and categorize the rest. The work was expensive, slow, and relentless.
Technology-assisted review (TAR), also called predictive coding, arrived in the 2010s and changed the economics. Instead of reviewing every document, attorneys would review a sample, train a model on those reviews, and use the model to prioritize or classify the rest. Courts gradually accepted TAR as defensible, and by 2020 it was standard practice for large document sets.
In 2026, AI has gone further. The tools are more capable, the workflows have changed, and the adoption curve has accelerated. Here's what's actually in use and what it means for anyone managing litigation or regulatory matters.
From TAR to generative AI review
Traditional TAR systems worked by learning from human coding decisions and then predicting how reviewers would code unseen documents. The model was a classifier: relevant or not, responsive or not, privileged or not. This works well but has limits. The model can predict document categories, but it can't explain why a document is relevant to a particular legal theory or summarize what a document says.
Generative AI changes this. Modern eDiscovery platforms now include generative models that can read a document and answer questions about it, summarize what it says, explain why it might be relevant to specific claims, and flag potential issues (potential privilege, hot documents, key custodians). This is qualitatively different from classification and significantly changes how review workflows are structured.
The practical change: instead of attorneys reviewing documents one by one in a linear queue, they increasingly work by querying: "show me documents where the CFO discussed the revenue recognition policy," or "find communications between these custodians in the 30 days before the audit." The AI surfaces documents relevant to a query; the attorney evaluates what it found.
This is faster in many cases, but it introduces a new risk: the review is no longer exhaustive by default. If you didn't ask the right question, documents responsive to that question may not get reviewed. Protocol design matters more than before.
Privilege review: the highest-stakes application
Privilege review, identifying documents protected by attorney-client privilege or work product doctrine and withholding them from production, is one of the most consequential tasks in discovery. Inadvertently producing a privileged document can waive privilege. Missing a privileged document in the review and producing it by mistake has caused significant problems in high-profile cases.
AI privilege review works by identifying features associated with privilege: presence of attorneys in the communication, subject matter that suggests legal advice is being sought or given, specific phrases common in attorney communications, and metadata patterns (emails with outside counsel copied, documents created by attorneys).
The major platforms (Reveal, DISCO, Relativity with its AI modules, Logikcull) all offer some form of AI-assisted privilege identification. The typical workflow: AI pre-codes a privilege probability score for every document. Reviewers focus their attention on documents with high privilege probability and sample documents in the middle range for quality control.
A few things matter here:
Attorney review is still required. No court has accepted fully automated privilege determinations. The AI identifies candidates; a licensed attorney makes the final call. The efficiency gain is in triage: instead of an attorney reading every document looking for privilege, they read the ones the AI flagged, which is a small fraction of the total.
Model transparency is a practical issue. In some jurisdictions, producing parties have been asked to explain their review methodology, including any AI used. Knowing your platform's technical approach to privilege identification matters for being able to defend your process.
Cross-border complexity. Legal professional privilege varies significantly between jurisdictions. What's privileged in the US may not be privileged in the UK or Germany and vice versa. AI models trained primarily on US legal patterns may underperform on non-US privilege analysis.
Reveal: deep AI integration
Reveal (formerly Brainspace) has built one of the more deeply integrated AI review platforms. Their document analysis capabilities include conceptual clustering (grouping documents by topic without keyword searches), relationship visualization (mapping communication networks between custodians), and sentiment analysis that can flag emotionally charged communications that may be significant to the matter.
Their generative AI layer, added in the 2024-2025 timeframe, allows attorneys to have natural language conversations with the document set. "What did the sales team know about the product defect before the recall announcement?" is a query that would have required a complex Boolean keyword search before; now it's a direct question to the document corpus.
Reveal is typically used by AmLaw 200 firms and corporate legal departments handling large matters. Their pricing is primarily consumption-based (per gigabyte of data processed), which aligns costs with matter size but makes budgeting complex on unpredictably sized document sets.
DISCO: the cloud-native challenger
DISCO built its eDiscovery platform natively in the cloud at a time when most competitors were adapting software originally built for on-premise deployment. The cloud-native architecture means faster processing, more predictable scaling, and a simpler data security model (all data in their cloud rather than complex on-premise installations).
DISCO's AI capabilities include DISCO Cecilia, their generative AI review assistant. Cecilia can summarize documents, answer questions about specific records, and help attorneys understand a document's significance to the matter.
For midsized law firms without large litigation technology infrastructure, DISCO's cloud-native model reduces the operational overhead of running an eDiscovery platform. They've captured significant market share among mid-market firms that previously relied on managed review services from outside vendors.
DISCO went public in 2021 and has had a volatile trajectory since, including significant cost-cutting in 2023. The product has continued to develop, but organizations signing multi-year contracts should be aware of that financial history.
Document classification at scale
Beyond privilege review, AI classification in eDiscovery covers several other categories:
Relevance: Is this document relevant to the matter? AI classifiers trained on initial attorney review can process millions of documents and produce ranked relevance scores. This is the core TAR function that has been in use longest.
Issue coding: Tagging documents to specific legal theories or issues (breach of contract, fraud, ERISA compliance, etc.) based on content. Useful for organizing review output and building a factual narrative.
Hot documents: Flagging documents that appear particularly significant, either because they directly support or undermine a key legal position, or because they contain alarming content. The "hot" flag is contextual; what matters in one case may not in another. Platforms implement this through a combination of keyword matching and semantic similarity to attorney-supplied examples of significant documents.
Deposition preparation: AI tools that can answer "given these documents, what would be the key topics to explore with this witness?" are in use at several large firms. This doesn't replace attorney preparation but can surface angles that might be missed in a manual review.
What courts have said
Judicial acceptance of AI in discovery has evolved substantially. The 2012 Da Silva Moore decision in the Southern District of New York was an early acceptance of TAR as a defensible review methodology. Since then, courts have generally accepted AI-assisted review when:
- The methodology is disclosed to opposing counsel
- The parties agree on validation procedures (sampling to verify recall rates)
- An attorney is responsible for and signs off on production decisions
- The approach is documented in a review protocol
The emergence of generative AI in review has prompted updated guidance. Several courts have issued standing orders in 2024-2025 requiring disclosure when generative AI is used in preparing legal documents or conducting review. The specific requirements vary by judge and jurisdiction.
The most cautious approach: treat any AI-assisted review workflow as subject to disclosure requirements, be prepared to explain the methodology in plain terms, and have a licensed attorney accountable for every production decision.
Cost implications
For organizations with significant litigation exposure, AI-powered eDiscovery has substantially changed the economics of large document review.
Before AI-assisted review became standard, a 1-million-document review at a major law firm might require 30-50 contract reviewers working for 4-8 weeks, at a total cost of $800,000 to $2 million just for first-pass review.
With AI-assisted workflows, the same review might involve 8-12 reviewers focusing on AI-prioritized documents, completing in 2-3 weeks, at a total cost of $200,000-$500,000, plus platform fees. The savings are real and substantial, and they've made large-scale litigation more financially accessible for smaller organizations.
The catch: the AI and platform costs are not trivial. A 1-million-document review on a platform like Reveal or DISCO might run $15,000-$40,000 in platform fees before attorney time. For small matters with limited documents, traditional review may still be cheaper.
What litigators should actually know
If you're a litigator or in-house counsel who doesn't work directly with eDiscovery technology, the practical takeaways:
Ask your eDiscovery vendor or litigation support team what AI tools are being used in your matter and how. Be prepared to describe the methodology if opposing counsel asks. Consider whether your jurisdiction's judges have issued any guidance on AI disclosure.
Review validation is still required. AI reduces the documents you review but doesn't eliminate the attorney review obligation. A fully AI-automated production without attorney review would be professionally problematic and technically defensible only in narrow circumstances.
Query-based review with generative AI is powerful but not exhaustive. If you use it, supplement with a traditional random-sample quality check to ensure you're not missing responsive material you didn't think to query for.
The tools are better than they've ever been. The fundamental legal obligations, attorney responsibility for production decisions, completeness, and privilege protection, haven't changed.