AI for e-discovery — the identification, collection, review, and production of electronically stored information (ESI) in litigation — has transformed a process that once required hundreds of attorney hours and millions in legal spend into a far more efficient, accurate, and cost-effective workflow. Technology-Assisted Review (TAR), predictive coding, and AI-powered document review have received judicial approval in US federal courts, UK courts, and EU jurisdictions. In 2026, AI e-discovery tools achieve 80–90% recall on relevant documents at a fraction of manual review cost. This guide covers the technology, the judicial standards, and the workflow for enterprise legal teams.
The E-Discovery Workflow
EDRM: Electronic Discovery Reference Model
The EDRM (Electronic Discovery Reference Model) defines the standard e-discovery workflow: Information Governance → Identification → Preservation → Collection → Processing → Review → Analysis → Production → Presentation. AI touches every stage but delivers the largest value in Review — the most labour-intensive step where attorneys historically read every document to determine relevance, privilege, and responsiveness. AI-assisted review (Technology-Assisted Review / TAR) trains a predictive model on attorney-reviewed seed documents, then scores the entire document population for relevance — reducing the number of documents requiring attorney review by 60–80%.
TAR Approaches
| Approach | How It Works | Best For |
|---|---|---|
| TAR 1.0 (Predictive Coding) | Attorney reviews seed set → model trained → entire collection scored | Large collections; standard litigation |
| TAR 2.0 (Continuous Active Learning) | Model improves iteratively as attorneys review documents it selects | Faster; more efficient; current best practice |
| Generative AI Review | LLM reads each document; determines relevance with reasoning | Complex fact patterns; nuanced relevance determinations |
| Concept Clustering | Groups documents by semantic similarity without keyword searching | Issue identification; early case assessment |
80%
Typical TAR recall — 80% of relevant documents identified at a review rate of 20–30% of the total collection. Manual review rarely achieves better than 60–70% recall due to reviewer fatigue and inconsistency. Courts have accepted 75–80% recall as reasonable in multiple rulings
Da Silva Moore
The landmark 2012 US federal court ruling (Da Silva Moore v. Publicis Groupe) that first approved TAR for e-discovery — establishing that predictive coding is an accepted method for e-discovery review. Now routine in SDNY, NDCA, and UK courts
Relativity
The dominant enterprise e-discovery platform — Relativity's AI features (Active Learning, Assisted Review, Analytics) are the industry standard tools for TAR implementation in large litigation matters. Used by Am Law 200 firms and Fortune 500 legal departments
Early Case Assessment
Before full review, AI concept clustering and email threading enables early case assessment: identify the key custodians (whose documents are most relevant), the time period of maximum activity, the communication networks around key events, and early relevance indications. This informs preservation strategy, custodian interviews, and settlement valuation before investing in full review. Relativity Analytics and Reveal AI both provide early case assessment dashboards. Use concept search (not keyword search) for ECA — keywords miss conceptually relevant documents that don't use the search term.
Continuous Active Learning (TAR 2.0)
TAR 2.0 workflow: (1) Attorney reviews 200–500 seed documents (mixture of relevant and non-relevant); (2) AI model trained on seed set scores entire collection; (3) AI presents highest-scoring (most likely relevant) documents for attorney review; (4) Attorney reviews and codes — model updates continuously; (5) Repeat until elusion rate (relevant documents in non-reviewed set) drops below threshold. Result: 70–80% of the collection never requires attorney review. Defensibility requires: logging all attorney review decisions, model performance metrics, and a validation set to confirm recall before closing review.
Privilege Review
AI privilege review identifies attorney-client communications and work product for privilege log preparation — the most error-prone manual review task. AI classifies documents as potentially privileged based on: attorney names in to/from/cc fields (against a firm roster), legal advice language patterns, litigation context signals. AI generates privilege log entries (document date, author, recipient, privilege claimed, description) for attorney verification. This does not replace attorney privilege review — it prioritises and assists it. Relativity's Privilege Detect and Reveal's privilege AI handle this workflow.
Cross-Border and Multi-Language Review
LLM-based e-discovery tools enable multi-language document review without translation: review French, German, Spanish, and Mandarin documents in English using AI translation integrated into the review workflow. This is critical for international litigation and regulatory investigations where documents span multiple languages. Relativity Translate and Reveal AI's multilingual support handle 50+ languages. Caveat: AI translation for legal review requires quality control — translation errors in key documents must be caught before production. Sample 10–15% of AI-translated documents against a professional translator for quality assurance.
E-Discovery Technology Implementation
Our AI consulting and software development teams help legal departments evaluate and implement AI e-discovery platforms with appropriate workflow and defensibility controls. Book a free advisory session.