eDiscovery AI Workflow Guide for Legal Teams: Document Review in Practice

EDiscovery document review is the task where AI has made the most measurable, documented impact in legal practice — and also where the failure modes are most consequential. A missed responsive document or a mislabeled privileged communication can result in sanctions, adverse inferences, or professional discipline. That makes it exactly the kind of workflow where understanding what AI does and does not do is not optional background knowledge.

This guide is scoped to document review within eDiscovery: the phase after collection and processing where a legal team must identify which documents are responsive, privileged, or relevant to specific issues. It is written for associates, paralegals, and legal ops staff who are responsible for running or supervising AI-assisted review projects. Partners and in-house counsel overseeing review teams will find the sections on human sign-off requirements and professional responsibility considerations most relevant.

The Underlying Task: What Document Review Actually Requires

Before placing AI anywhere in the workflow, it helps to be precise about what the task demands. A document review project requires a team to make a series of binary or categorical decisions about each document in a collection — typically: is it responsive to the requests for production? Does it contain privileged communications? Does it touch a specific issue or custodian? Is it a duplicate of another reviewed document?

The volume problem is what makes AI relevant. Large commercial litigations routinely involve collections of 500,000 to several million documents. Manual review at $50–$150 per hour per reviewer, even with contract attorneys, produces costs that can easily reach seven figures for a single matter. That cost pressure is what drove the adoption of technology-assisted review (TAR) over the past decade, and it is what is now driving the integration of generative AI tools alongside traditional predictive coding.

The quality problem runs in the opposite direction. Human reviewers working at speed have documented error rates. Studies of manual review accuracy have found error rates ranging from 20% to over 40% depending on fatigue, ambiguous instructions, and document complexity. AI does not eliminate error — it shifts where errors occur and, when properly configured, can make certain error patterns more detectable and correctable.

Where AI Is Currently Inserted in the Review Workflow

AI is not a single insertion point in eDiscovery review — it appears at multiple stages, with different tools and different levels of human oversight required at each. The following breakdown reflects how AI is currently deployed in documented practitioner workflows, not how any single vendor positions their product.

Stage 1: Pre-Review Culling and Prioritization

Before a reviewer sees a single document, AI-based tools can reduce the collection size substantially. Deduplication (exact and near-duplicate identification), email threading, and concept clustering are all well-established at this stage. Generative AI adds the ability to generate conceptual summaries of document clusters, helping project managers design review batches around issue categories rather than arbitrary date ranges or custodians.

Stage 2: Technology-Assisted Review (Predictive Coding)

Traditional TAR workflows — TAR 1.0 (simple active learning) and TAR 2.0 (continuous active learning) — use machine learning models trained on attorney-reviewed seed sets to predict responsiveness scores across the full collection. These are not generative AI tools; they are classification models. Most established eDiscovery platforms (Relativity, Everlaw, Reveal, Nuix) have offered this capability for years.

What has changed recently is the integration of large language model-based review assistance alongside predictive coding. Rather than replacing TAR, LLM-based tools are being layered in to handle tasks the classification models handle poorly: documents with ambiguous language, foreign-language content, and documents where context outside the document itself is needed to make a responsiveness determination.

Stage 3: Privilege Review and Logging

Privilege review is where AI assistance carries the highest professional responsibility stakes. AI tools are being used to flag potentially privileged documents for attorney review, generate draft privilege log entries, and identify common privilege patterns across a collection. None of these functions are appropriate for full automation without attorney sign-off.

The privilege log generation use case has seen the most active development. Tools like Relativity's AI-assisted privilege log features and comparable functionality in other platforms can draft log entries from document metadata and content, which a reviewing attorney then edits and certifies. The time savings are real — drafting privilege log entries is highly repetitive work — but the certification step cannot be delegated to the AI output.

Stage 4: Issue Tagging and Issue-Specific Review

Beyond responsiveness and privilege, complex litigations require documents to be tagged against specific factual issues, timeframes, or custodian-related categories. This is where generative AI's ability to understand context and classify documents against nuanced criteria shows the clearest advantage over keyword search or traditional predictive coding.

An LLM-based tool given a well-constructed prompt describing a specific issue — say, all documents discussing the negotiation of a particular contract term between identified parties — can classify documents against that issue with reasonable accuracy across large volumes. The key variable is prompt quality, which is discussed in the next section.

Stage 5: Quality Control and Consistency Checking

AI is also being used at the back end of review for QC purposes: identifying inconsistently coded documents, flagging documents where the AI's responsiveness prediction diverges significantly from the reviewer's coding, and surfacing documents that may have been overlooked in initial review passes. This is one of the more defensible uses of AI in the workflow because it functions as a check on human review rather than a replacement for it.

AI Capabilities by Review Stage: A Summary

Prompt Engineering for Document Review Tasks

For LLM-based review tools, the quality of the classification instruction — the prompt or review protocol fed to the model — is the primary determinant of output quality. This is not a peripheral technical concern. A vague responsiveness definition produces vague classification results. A well-constructed prompt that mirrors the specificity of the RFP language and the meet-and-confer agreements produces materially better output.

Several patterns have emerged from documented practitioner experience with LLM-based review tools:

• Specificity over generality: Prompts that reference specific parties, date ranges, subject matter, and document types consistently outperform general responsiveness descriptions.
• Explicit exclusions: Including what is not responsive in the prompt reduces false positives, particularly for common business documents that touch the subject matter tangentially.
• Issue-by-issue instructions: Running separate classification passes for each distinct issue, rather than attempting to classify all issues in a single pass, produces more auditable results.
• Validation against seed sets: Before deploying a prompt across the full collection, test it against a human-reviewed seed set of at least 200 documents. A precision and recall measurement at this stage catches systematic errors before they scale.
• Prompt versioning: Document the exact prompt text used for each classification pass. If the protocol is challenged in discovery, you need to be able to produce it.

Documented Limitations and Known Failure Modes

AI-assisted review has documented failure modes that legal teams need to account for in their review protocols, not just be aware of abstractly.

Context Window and Document Length Constraints

LLM-based review tools have context window limits. Long documents — deposition transcripts, lengthy contracts, multi-hundred-page regulatory submissions — may be truncated or chunked in ways that cause the model to miss relevant passages that appear outside the processed portion. Review protocols should include a separate handling procedure for documents over a defined length threshold, typically 50–100 pages depending on the platform.

Implicit and Coded Language

In cases involving potential misconduct, relevant communications often use indirect or coded language. AI classification tools trained on standard business communication patterns can miss documents where the substance is conveyed through implication, industry jargon, or deliberate circumlocution. Human review of statistically sampled documents from the non-responsive population remains important precisely because of this failure mode.

Non-Text Content

Spreadsheets, images, audio files, video, and CAD drawings are not processed reliably by text-based LLM review tools. Most platforms handle this through OCR for images and metadata extraction for other file types, but the classification accuracy on these document types is substantially lower than for standard text documents. Collections with significant non-text content require separate handling protocols.

Foreign Language Documents

LLM-based tools vary significantly in their performance on non-English documents. English-language responsiveness prompts applied to foreign-language documents produce unreliable results. Matters with significant foreign-language document populations need platform-specific capability verification and, in most cases, separate review protocols for each language.

Human-in-the-Loop Requirements: What Attorneys Cannot Delegate

The practical question for most teams is not whether to use AI in document review — the cost and volume pressures make that decision straightforward — but where attorney judgment must remain in the loop. The following steps are not optional for professionally responsible AI-assisted review.

1. Designing the review protocol: The responsiveness definitions, privilege criteria, and issue tagging framework must be set by an attorney with knowledge of the case. AI can draft a starting point, but the attorney is responsible for the final protocol.
2. Approving the culling methodology: Before any documents are excluded from review on the basis of AI culling, an attorney must review and approve the methodology and the sample audit results.
3. Certifying privilege determinations: Every document withheld as privileged requires an attorney's judgment. AI-generated privilege flags and log entries must be reviewed and certified by a licensed attorney.
4. Statistical validation sign-off: At each stage where TAR or LLM classification is used, a supervising attorney must review the statistical validation results (precision, recall, elusion rate) and make a documented decision to proceed or adjust.
5. Production certification: The attorney signing the discovery responses and production certification is certifying the review process, including the AI components. That attorney needs to understand and be able to defend the methodology.
6. Responding to challenges: If opposing counsel challenges the review methodology, the attorney needs to be able to explain and defend each AI-assisted step. Reliance on "the platform did it" is not a sufficient explanation.

Professional Responsibility Considerations

The professional responsibility framework for AI-assisted review has developed substantially over the past two years. State bar ethics opinions and the ABA's formal guidance have converged on several consistent principles, though the specific obligations vary by jurisdiction.

Competence (ABA Model Rule 1.1)

Competent representation now includes understanding the AI tools being used in a matter well enough to supervise them. This does not require technical expertise in machine learning, but it does require understanding what the tool does, what its error patterns are, and what validation steps are appropriate. An attorney who deploys AI-assisted review without understanding the methodology is not meeting the competence standard as most bar ethics opinions have interpreted it.

Supervision (ABA Model Rule 5.3)

AI tools are not licensed attorneys, but the supervision obligations under Rule 5.3 — which govern the supervision of non-attorney staff — have been applied by analogy to AI tool use in multiple bar ethics opinions. The supervising attorney is responsible for the work product, including work product generated with AI assistance. Supervision means reviewing outputs, not just approving them.

Confidentiality (ABA Model Rule 1.6)

Sending client documents to a third-party AI platform triggers Rule 1.6 confidentiality obligations. Before using any cloud-based AI review tool, the supervising attorney must verify the vendor's data handling practices: whether documents are retained after processing, whether they are used for model training, and whether the vendor's security practices meet the firm's obligations to the client. Many eDiscovery platforms offer zero-retention or isolated processing environments specifically to address this concern, but the obligation to verify rests with the attorney, not the vendor.

Court Disclosure Requirements

A growing number of federal district courts have issued standing orders requiring disclosure of AI use in filed documents and, in some cases, in discovery processes. The disclosure landscape is changing rapidly — as of mid-2026, dozens of federal judges have issued orders, and several district courts have adopted court-wide AI disclosure rules.

For eDiscovery specifically, the key question is whether your jurisdiction requires disclosure of AI-assisted review methodology in discovery responses or meet-and-confer discussions. Some judges have indicated that TAR and AI-assisted review methodologies should be disclosed proactively; others have taken the position that the existing discovery rules are sufficient. Check the standing orders for the specific judge and court before finalizing your review protocol.

Choosing the Right Tool Configuration for Your Matter

Not every matter requires the same AI configuration. The appropriate tool setup depends on collection size, document type mix, matter complexity, and budget. A few practical distinctions:

Documenting the Review Protocol for Defensibility

One aspect of AI-assisted review that is easy to overlook during a fast-moving matter is documentation. The review protocol — including the AI tools used, the prompts or training protocols applied, the validation methodology, and the results of each validation check — needs to be documented contemporaneously. After the fact reconstruction of a review methodology is not credible and will not hold up under challenge.

A defensible review protocol document should include:

• The platform(s) used and the version or configuration at the time of review
• The responsiveness definitions and issue coding criteria, including any amendments made during review
• The exact prompt text or training protocol used for each AI-assisted classification pass
• The seed set composition and size for any TAR workflow
• Statistical validation results at each checkpoint (precision, recall, elusion rate where applicable)
• The identity and qualifications of the supervising attorney who approved each stage
• Any deviations from the original protocol and the reason for each deviation

Some firms are now treating the review protocol document as a matter file record that is retained alongside the production log. That practice is reasonable and increasingly expected in large commercial litigation contexts.

Common Mistakes in AI-Assisted Review

• Skipping the culling audit: Assuming the AI culling correctly excluded only non-responsive documents without running a sample audit of the culled population. This is where responsive documents most commonly disappear.
• Treating AI privilege flags as determinations: Using AI-flagged privilege documents to populate the privilege log without attorney review of each entry. Privilege determinations require attorney judgment.
• Insufficient seed sets: Running TAR with seed sets under 200 documents, particularly in complex matters. Underpowered seed sets produce unreliable models that appear to validate but miss systematic categories of responsive documents.
• Single-pass validation: Validating the model once at the beginning of review and not re-validating as the collection is processed. Active learning models drift as they are exposed to more documents; re-validation at defined checkpoints is necessary.
• Ignoring the non-responsive population: Treating all AI-classified non-responsive documents as reviewed. A statistically valid elusion sample from the non-responsive population is required to support a defensible completeness representation.
• Not disclosing when required: Failing to check whether the presiding judge has a standing order requiring disclosure of AI review methodology before certifying the production.