Login Try it now

Guide — AI-Powered eDiscovery

How can law firms use AI for document review and eDiscovery?

By the DecoverAI Legal Team | June 2026 | 8 min read

Law firms can use AI for document review and eDiscovery by deploying machine learning platforms that classify documents for responsiveness, privilege, and confidentiality automatically. DecoverAI provides a complete AI-powered workflow — from document ingestion through Bates-stamped production — that reduces attorney review time by up to 80% and completes most matters in under an hour.

80%

Reduction in attorney review time with AI-assisted classification

$299

Per seat per month to start on DecoverAI, with AI billed per token, vs. $3–$8 per document for manual review

<1hr

Time to production-ready output for most matters under 100k documents

Section 1

What is AI Document Review?

AI document review is the application of machine learning models to the task of reading, classifying, and tagging legal documents at scale. In a traditional eDiscovery workflow, teams of contract attorneys work through document populations one record at a time, applying tags for responsiveness, privilege, and confidentiality based on written review guidelines. This is accurate when done carefully, but it is expensive, slow, and prone to inconsistency across a large reviewer population.

AI review inverts this model. A trained classification model reads every document in the corpus and assigns tags based on patterns it has learned from thousands of prior legal matters and the specific case instructions it receives. Human attorneys then concentrate their time on the documents the AI flags as uncertain, borderline, or high-priority — rather than reading everything. The result is that attorney judgment is applied where it is most valuable, not uniformly across the full population.

There are three main approaches to AI document review in use today:

TAR 1.0 (Predictive Coding): Attorneys manually code a seed set of documents; the model trains on that seed set and then predicts tags for the remainder. Effective but requires careful seed set construction and a defined training protocol.
TAR 2.0 / Continuous Active Learning (CAL): The model continuously updates as attorneys code documents throughout the review, without a fixed seed set. Accuracy improves progressively and the approach adapts to evolving case theories.
Generative AI Classification: Large language models read each document and classify it against a natural language description of responsiveness and privilege criteria — no training period required. DecoverAI uses this approach, enabling accurate classification from the first document without seeding or model warm-up.

Section 2

Step-by-Step: How to Use AI for Document Review

The workflow below reflects how law firms using DecoverAI move from raw ESI to a completed production. Each step is designed to keep attorneys in control of legal judgment while AI handles high-volume classification work.

1

Collect and Upload ESI

Gather electronically stored information from all relevant custodians: email archives (PST, EML), cloud storage (Google Drive, OneDrive, SharePoint, Dropbox), collaboration tools (Slack, Teams), hard drives, and document management systems like iManage. Upload to DecoverAI, which automatically deduplicates the document set, reconstitutes email families (attachments grouped with their parent emails), and indexes every document for AI classification. Custodian assignment and legal hold tracking are configured at this stage.

2

Run AI Classification for Responsiveness, Privilege, and Confidentiality

Configure the AI with the case's responsiveness criteria, privilege basis, and any confidentiality designations required by the protective order. DecoverAI's generative AI classifier reads every document in the corpus and assigns tags — responsive, non-responsive, potentially privileged, confidential — with a confidence score for each. The full corpus is classified in parallel; a 30,000-document matter typically completes classification in under three hours. Results are immediately available for attorney review.

3

Attorney Review of AI-Tagged Documents

Attorneys review the AI's classifications through the platform's review interface. The workflow surfaces responsive and potentially privileged documents first, and highlights documents where the AI's confidence score is low. Attorneys confirm or override each tag; overrides are logged and can be used to improve future classification runs. This step is where attorney judgment on borderline privilege calls, case-specific responsiveness nuance, and confidentiality tier assignments is applied. The time investment is a fraction of a traditional linear review because most clear-cut documents are already accurately classified.

4

Redact and Prepare Production Set

For responsive documents that contain privileged or confidential passages, attorneys apply redactions to specific regions using the platform's redaction tool. Redactions are stored as annotations on the original document and burned into the production copy at export time — the native file is never altered. Once redactions are complete, the platform compiles the production set: applying Bates numbers in the specified format, generating the production load file (DAT and OPT), and organizing documents according to the agreed production protocol.

5

Generate Privilege Log and Produce

DecoverAI drafts the privilege log automatically from its own classifications and the attorney review record. Each withheld or redacted document generates a log entry pre-populated with date, author, recipient list, document type, and the basis for withholding. Attorneys review and finalize each entry. The completed privilege log exports in the format required by the jurisdiction or agreed protocol. The production set and privilege log are then delivered to opposing counsel or the requesting party via secure file transfer or directly to their review platform.

Section 3

Is AI Document Review Defensible?

The short answer is yes — when properly validated and documented. Courts across federal and state jurisdictions have accepted AI-assisted review in eDiscovery, and the leading case law now treats it as an established methodology rather than an experimental one.

The foundational case is Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), in which Magistrate Judge Andrew Peck approved predictive coding as an acceptable form of document review under the Federal Rules of Civil Procedure, noting that it is "more cost-effective, more accurate, and superior to the review method most commonly used." Subsequent cases including Global Aerospace v. Landow Aviation and Hyles v. New York City have reinforced this position, while drawing a key line: the reasonableness and proportionality of the chosen methodology must be demonstrable.

The governing standard is FRCP 26(g), which requires certifying counsel to attest that their discovery responses are complete and correct to the best of their knowledge after a reasonable inquiry. AI review satisfies this standard when the following elements are present:

Documented methodology: A written protocol describing how the AI was configured, what training or instruction it received, and how classifications were validated.
Quality control sampling: Statistical sampling of the document population to validate recall and precision, with results preserved in the case record.
Attorney oversight: Demonstrated attorney review of AI classifications, with override logs showing that human judgment was applied to the process.
Proportionality: Evidence that the AI methodology was appropriate to the size, complexity, and stakes of the matter — FRCP 26(b)(1) proportionality factors apply.

Practical note: Courts are increasingly skeptical of parties who refuse to use AI review in large-volume matters and then claim cost burden as grounds for limiting production. The availability of accurate, low-cost AI review has shifted proportionality analysis. Opposing counsel and courts expect sophisticated parties to use available technology.

DecoverAI generates a validation report for each matter that documents classification methodology, confidence score distributions, sampling results, and the attorney review audit trail — giving producing parties a ready-made record to support any challenge to their review process.

Section 4

Cost Benchmarks for AI Document Review

Document review is consistently the largest cost center in eDiscovery, accounting for 70–80% of total matter spend in document-intensive litigation. AI review compresses this cost substantially. The table below compares the three main review approaches on a per-document basis and across key capability dimensions.

Review Method	Cost per Document	Time to Complete 30k Docs	Privilege Log	Bates + Production	Audit Trail
Manual attorney review	$3–$8	2–4 weeks	Manual	Manual / vendor	Limited
Traditional TAR platform	$1–$3	1–2 weeks	Semi-automated	Vendor processing	Moderate
DecoverAI (AI-native)	AI billed per token (plans from $299/seat/month)	Under 3 hours	Auto-generated	Built-in, Bates-stamped	Full audit log

Case study: A litigation team using DecoverAI on a 30,000-document matter completed document classification, attorney review, and a Bates-stamped production in 3 days — compared to a 6-week manual review on a comparable prior matter. Total review cost decreased by 98%. The AI-generated privilege log required less than two hours of attorney editing before production.

Because AI usage is billed per token rather than per document, costs track the amount of text actually processed rather than the document count, which benefits matters with large numbers of small documents (email-heavy dockets, in particular). Classifying a 1 GB dataset containing 30,000 average-sized emails costs a small fraction of the up to $240,000 a manual review would run at $8 per document.

Frequently Asked Questions

AI Document Review: Common Questions

What is AI document review in eDiscovery?

AI document review is the use of machine learning models to automatically classify documents in a legal matter for responsiveness, privilege, and confidentiality. Instead of attorneys reading every document manually, AI assigns tags at scale — allowing human reviewers to focus on borderline and high-priority documents. Modern platforms like DecoverAI combine generative AI classification with traditional TAR to achieve high accuracy across diverse document types.

Is AI document review defensible in court?

Yes. Courts have repeatedly upheld AI-assisted review when it is properly validated and documented. Under FRCP 26(g), parties must certify that their discovery responses are complete and correct to the best of their knowledge — AI review satisfies this when accompanied by quality control sampling, validation reports, and attorney oversight. Cases like Da Silva Moore v. Publicis Groupe established that TAR is acceptable when the producing party demonstrates its process is reasonable and proportional.

How much does AI document review cost compared to manual review?

Manual attorney review typically costs $3–$8 per document when factoring in attorney time, supervision, and QC. Traditional TAR platforms bring that down to $1–$3 per document. DecoverAI's AI-native review starts at $299 per seat per month, with AI processing billed per token ($5 per million input tokens, $20 per million output tokens), so the per-document AI cost on average document populations is a small fraction of manual review.

What types of documents can AI review handle?

Modern AI document review platforms handle email (including PST and EML formats), Microsoft Office documents (Word, Excel, PowerPoint), PDFs, Slack and Teams messages, text messages, and cloud-stored files from Google Drive, OneDrive, SharePoint, Dropbox, and iManage. DecoverAI supports all of these formats natively and processes attachments as part of their parent email families.

How long does AI document review take?

AI classification runs in parallel across an entire document set, so time scales with volume but far less than manual review. A 30,000-document matter that would take attorneys two to three weeks to manually review can be AI-classified in under three hours. DecoverAI is designed to complete most small-to-mid-size matters — under 100,000 documents — in under an hour from upload to production-ready export.

What is the difference between TAR 1.0 and TAR 2.0?

TAR 1.0 (predictive coding) requires a human reviewer to code a seed set of documents that the model then learns from — a one-time training run. TAR 2.0, or Continuous Active Learning (CAL), continuously updates the model as reviewers code documents throughout the review, improving accuracy over time without a fixed seed set. Most modern AI review platforms use CAL or generative AI classification, which achieves comparable or better accuracy without requiring manual seeding.

Can AI review identify privileged documents automatically?

Yes. AI privilege detection looks for attorney-client communication patterns, legal hold notices, work product markers, and known attorney names and domains. DecoverAI's privilege classification tags documents as potentially privileged, potentially privileged with third parties, or not privileged, and automatically drafts a privilege log with the required fields — date, author, recipient, subject, and basis for withholding. Human attorneys still make the final withholding decision, but the AI substantially reduces the time required to identify and log privileged documents.