E-discovery has been an AI use case longer than almost any other in law. Technology-assisted review — using machine learning to predict document relevance in large productions — has been court-approved and ethically sanctioned since at least 2012, when a prominent federal court decision (Da Silva Moore) endorsed its use for the first time. AI in document review is not an emerging practice; it is an established one.
What has changed is access. The enterprise e-discovery platforms that large firms use (Relativity, Everlaw, DISCO) were priced and built for large firms. The emergence of AI document review capabilities at lower price points, and the availability of general AI tools for smaller-scale review tasks, has made this workflow accessible to solos and small litigation firms in a way it was not five years ago.
This guide covers the e-discovery AI landscape for attorneys who are not inside a large firm's discovery department — what the technology does, where it applies, and how to use it ethically.
What AI does in document review
Relevance prediction (technology-assisted review). The classical AI document review application: a human reviewer codes a seed set of documents as relevant or not; the model learns from those decisions and predicts the relevance of the remaining population. Used for large productions (typically hundreds of thousands of documents or more), this dramatically compresses review time by allowing reviewers to focus on predicted-relevant documents first and by reaching a defensible completeness threshold faster than linear review.
Conceptual search and issue identification. Rather than searching for specific keywords, conceptual search identifies documents related to a concept, regardless of the exact words used. Useful for finding documents about a topic when you do not know the exact vocabulary the producing party uses.
Near-duplicate detection and email thread analysis. Identifies duplicate and near-duplicate documents to reduce the volume that requires individual review. Email threading groups related messages so they can be reviewed in context rather than as isolated documents.
First-pass privilege review assistance. AI can flag documents that are likely to contain privileged communications based on patterns — lawyer names in the to/from fields, legal vocabulary, claim-privilege language — to accelerate the privilege review layer. The attorney still makes every privilege determination; AI identifies the candidates.
Summarization and extraction for smaller productions. For small firms dealing with document sets in the hundreds or low thousands — too small for an enterprise platform to be economical — a general AI assistant can produce summaries, extract key facts, and identify themes from individual documents or small batches.
The scale question
The tools appropriate for document review depend on the volume involved.
Large-scale discovery (100,000+ documents): this is the domain of dedicated e-discovery platforms — Relativity, Everlaw, DISCO, Reveal, and others. These are not self-service tools you spin up yourself; they are managed platforms, often through an e-discovery vendor relationship. For a small firm handling a large case that requires this scale, the practical path is to engage an e-discovery vendor who manages the platform and provides the review workflow. The per-document economics at this scale make the investment rational.
Mid-scale discovery (10,000–100,000 documents): cloud-based e-discovery platforms have come down in price, and some offer self-service tiers that a small firm can use directly. Everlaw and DISCO in particular have moved toward more accessible pricing for smaller matters. If you regularly handle mid-scale discovery, evaluating one of these platforms directly is worth the time.
Small-scale discovery (under a few thousand documents): for productions that arrive as a set of PDF files, emails, or documents in the hundreds, the enterprise platform may not be cost-justified. Here, a general AI assistant or a document-review-specific tool handles the task: upload or paste documents, ask targeted questions, extract key facts, flag specific types of content. This is not technology-assisted review in the formal sense — it is AI-assisted document analysis, and for small sets it works well.
| Scale | Approach | Access model |
|---|---|---|
| 100k+ documents | TAR/CAL on a dedicated platform | E-discovery vendor engagement |
| 10k–100k documents | Cloud TAR platform (Everlaw, DISCO, etc.) | Self-service or vendor-managed |
| Under a few thousand | AI document analysis (general assistant or targeted tool) | Self-service |
| Individual contracts / agreements | Contract AI (Spellbook, CoCounsel) | Self-service subscription |
The competence obligation
ABA Formal Opinion 512 makes competence in AI tools an explicit requirement — and e-discovery is where the competence obligation has the most established case law behind it. Courts have consistently held that attorneys supervising document review are responsible for the adequacy of the review process, including the technology used to conduct it. Using AI for document review does not create a safe harbor; it creates an obligation to understand what the tool does, what its error rates are, and whether the process was defensible.
The practical implications:
Quality control sampling is not optional. Any AI-assisted review process requires a quality control step: a human reviewer samples predicted-not-relevant documents to verify that the model is not misclassifying relevant material. The size and structure of the QC sample depends on the platform and the matter, but it must happen. An AI review with no QC is not a review; it is a delegation.
Document the process. If opposing counsel challenges the adequacy of your review, your ability to defend it depends on your ability to describe it: the tool used, the seed set approach, the training method, the QC steps taken. Maintain a process log contemporaneously.
Privilege determinations remain with the attorney. AI can flag privilege candidates; it cannot make privilege decisions. Every document withheld or redacted as privileged reflects an attorney's judgment, and the attorney who made that judgment is responsible for it.
Warning
The confidentiality framework for discovery documents
Discovery documents contain opposing-party information, third-party information, and often some client information mixed in. Before any document is uploaded to an AI tool:
Check the tool's data handling terms. Enterprise e-discovery platforms have data handling terms designed for litigation use. General AI assistants vary by tier — consumer tiers train on inputs; business tiers generally do not. Never upload a production to a free AI tool.
Review the protective order. If a protective order governs the production, it may restrict the tools you can use to review it. A protective order that limits use of "cloud computing services" or "external processing" may constrain AI options. Read the order before you configure a workflow.
Data residency. For matters with international discovery or cross-border data transfers, data residency requirements may restrict which platforms can process documents. This is an issue primarily for larger matters with international parties.
How small firms access e-discovery AI
For the solo or small litigation firm that handles discovery as part of the practice rather than as a specialty:
Build a vendor relationship before you need it. Find an e-discovery vendor in your market who handles small-to-mid-scale matters and establish a relationship. When a large production arrives, you want to be able to call someone who already knows your practice, not spend three days evaluating vendors under deadline pressure.
Evaluate whether a cloud platform is worth a direct subscription. If discovery comes up in multiple matters per year, a self-service tier on a platform like Everlaw or DISCO may be more economical than case-by-case vendor engagement. Run the per-page economics on a typical matter.
Use general AI for small-scale analysis. For the smaller document sets that do not justify a platform, a business-tier general AI assistant handles document summarization, fact extraction, and issue spotting effectively. The workflow is described in the AI for document review prompts section.
The line between AI-assisted and AI-conducted review
The ethical touchstone is consistent across every AI application: AI assists; the attorney decides. In document review, that means:
- AI identifies candidates; attorneys make responsiveness calls
- AI flags privilege; attorneys make privilege determinations
- AI builds the training set; attorneys code the seed documents
- AI produces a QC sample; attorneys review it
The attorney is not optional — not for the process design, not for the judgment calls, and not for the final production decisions. AI is the force multiplier for a review that still requires professional judgment at every critical decision point.
For the full landscape of AI tools relevant to litigators, see the best AI tools for lawyers. For the responsible-use protocol on any AI tool that touches legal authority, see AI for legal research.
Related reading: AI for legal research | The best AI tools for lawyers | AI contract review for lawyers