AI-assisted discovery refers to the use of artificial intelligence technologies throughout the civil discovery process — the pretrial procedure by which parties exchange information and documents relevant to the disputed claims and defenses. Discovery encompasses the collection, processing, review, and production of electronically stored information (ESI) — emails, documents, spreadsheets, databases, text messages, and other digital information — as well as traditional paper documents.
Modern litigation involving organizations generates discovery volumes that make traditional human-only document review impractical. A large commercial dispute, regulatory investigation, or class action may involve tens of millions of documents across email systems, shared drives, enterprise applications, and collaboration tools. At historical attorney review rates (roughly 50–60 documents per hour), reviewing 10 million documents would require thousands of attorney-hours and cost millions of dollars.
AI-assisted discovery addresses this volume problem through several complementary technologies that reduce the number of documents requiring individual attorney review, prioritize the review of likely-relevant documents, and automate coding decisions on clearly irrelevant materials.
The term encompasses a spectrum from well-established technology-assisted review (TAR/predictive coding) to emerging generative AI overlays that can summarize, categorize, and analyze documents at scale.
The economics of eDiscovery have been a major driver of litigation cost for two decades. In major commercial litigation, eDiscovery costs can exceed legal fees — the cost of processing, reviewing, and producing millions of documents often dwarfs the cost of the underlying legal work. This has created intense pressure to reduce per-document review costs while maintaining accuracy and defensibility.
AI-assisted discovery is the profession's primary response to this cost pressure. Technology-assisted review, when properly validated, can reduce per-document review costs by 60–80% compared to linear human review. For a matter with 5 million documents, this can represent savings of millions of dollars.
Beyond cost, AI-assisted discovery offers accuracy advantages. Human reviewers are inconsistent — different reviewers make different coding decisions on the same document, and reviewer fatigue increases coding error rates over time. AI-assisted review applies consistent criteria across the full document population, reducing inter-reviewer inconsistency.
For in-house legal departments managing litigation spend, AI-assisted discovery is a critical cost management tool. For outside counsel, it is an expected capability in complex litigation — clients and courts increasingly expect the efficiency AI tools provide.
How It Works
Technology-Assisted Review (TAR 1.0 — Predictive Coding). The original TAR methodology, also called predictive coding, uses a supervised machine learning approach:
- A senior attorney "seeds" the system with a training set — coding a representative sample of documents as relevant or not relevant 2. The model trains on the attorney's coding decisions, learning to distinguish relevant from not-relevant documents based on linguistic features 3. The model applies relevance predictions to the full document population 4. Attorneys review model predictions, correct errors, and retrain the model iteratively until performance stabilizes 5. High-predicted-relevance documents receive priority human review; low-predicted-relevance documents are batch-coded as not relevant
Continuous Active Learning (CAL / TAR 2.0). An evolution of the predictive coding approach in which the model learns continuously as review proceeds rather than through discrete training rounds. CAL has largely replaced TAR 1.0 in modern platforms because it requires less upfront training effort and adapts more effectively as reviewers surface new responsive document types.
Generative AI Overlays (TAR 3.0 — Emerging). The newest generation of eDiscovery AI uses large language models to understand document content semantically rather than statistically — generating document summaries, categorizing documents by issue or topic, identifying key document themes, and answering specific questions across the document population. Relativity AI's aiR and DISCO's generative AI tools represent this emerging capability layer.
Concept Clustering and Near-Duplicate Identification. AI tools group documents by conceptual similarity, enabling reviewers to make coding decisions across conceptually similar document clusters rather than one at a time. Near-duplicate identification ensures that minor variations of the same document are coded consistently.
Email Threading. AI identifies complete email conversation threads, enabling reviewers to read conversations in context and make more accurate relevance and privilege determinations.
Key Considerations for Law Firms
Validation is required for defensibility. AI-assisted review must be validated to demonstrate that it meets discovery obligations. Validation methodology varies — some practitioners use statistical sampling to estimate recall (the percentage of relevant documents actually identified), others use elusion testing (sampling predicted-irrelevant documents to confirm the AI is not missing significant relevant materials). Document your validation methodology and results in case opposing counsel or the court challenges your review approach.
Negotiate TAR protocols with opposing counsel early. Discovery orders and case management orders should address AI-assisted review methodology, including training data, validation approach, and recall targets. Early agreement prevents disputes during or after review completion. Courts have consistently upheld TAR when parties agreed to or courts approved the methodology.
Privilege review requires attorney oversight. AI privilege identification is a first-pass tool, not a final determination. Every document identified as potentially privileged requires attorney review before a privilege log entry is created or a clawback claim is made. AI can identify documents that contain attorney names or "legal advice" language — the privilege determination requires legal analysis.
Data security in cloud eDiscovery platforms. ESI in litigation often includes highly confidential client information, trade secrets, and personal data subject to privacy regulations. Evaluate the security architecture of any cloud-based eDiscovery AI platform before uploading sensitive litigation materials. Consider zero-data-retention provisions that require the vendor to delete your data after the matter closes.
Limitations and Risks
Recall is never 100%. No AI-assisted review methodology achieves 100% recall of all relevant documents. Some relevant documents will not be identified by any AI system. Courts have accepted less-than-perfect recall as meeting discovery obligations — perfect recall is not the standard — but parties should understand that AI-assisted review optimizes cost and efficiency against a recall target, not perfect completeness.
Seed set quality determines model quality. For traditional predictive coding (TAR 1.0), the quality of the attorney's training decisions determines the quality of the AI's relevance predictions. Inconsistent or incorrect coding in the training set produces a model with systematic biases. The senior attorney responsible for creating the training set must apply careful, consistent judgment.
New document types may underperform. AI models trained on email and standard office documents may perform poorly on unusual document types — database exports, structured data files, proprietary application formats, multimedia files. Assess whether your document population includes unusual document types before selecting an AI review platform.
Cross-border discovery complications. International discovery involves data privacy regulations — GDPR, China's Personal Information Protection Law, other national frameworks — that restrict transfer of personal data outside the country of origin. AI eDiscovery platforms hosted in the US may not comply with international data transfer restrictions without additional safeguards. Cross-border discovery requires jurisdiction-specific privacy analysis before data collection and processing.