AI for Document Review and eDiscovery: The Modern Workflow

Document Review Has Changed More Than Any Other Legal Workflow

For decades, document review was the labor-intensive bottleneck of litigation. Junior associates spent thousands of hours coding documents. Contract reviewers were rented by the hour to wade through millions of pages. Predictive coding (technology-assisted review, or TAR) emerged in the 2010s and helped, but still required heavy human input.

In 2026, document review has fundamentally changed. Modern AI can read, classify, and analyze documents with reliability that rivals senior associate review - at 1 to 2% of the cost.

This article walks through the modern AI document review workflow, the tools that power it, and the verification layer that keeps your firm out of trouble.

The Old Model vs. The New Model

Old model (2015 to 2022):

• TAR/predictive coding with seed sets
• Heavy human coding of training data
• Time to first useful classifier: weeks
• Cost: $30 to $50/document for traditional, $5 to $15/document for TAR
• Best-case accuracy: 70 to 85%

New model (2024 to 2026):

• Foundation model-based classification with prompt-defined criteria
• Minimal seed coding (often zero-shot)
• Time to first useful classifier: hours
• Cost: $0.10 to $2/document depending on size and tool
• Accuracy: 85 to 95% on most categories

The economics have shifted dramatically. So has the workflow.

The Five-Step Modern Document Review Workflow

Step 1: Define Review Categories in Plain English (30 minutes)

Forget complex coding schemes for now. Open Claude or ChatGPT and write out your review categories in plain English:

"For this matter, I need to classify each document as:

• Privileged (attorney-client)
• Privileged (work product)
• Relevant - to claim
• Relevant - to defense
• Hot document (smoking gun)
• Irrelevant
• Need legal review

Privileged means [your specific definition]. Hot document means [your specific criteria]. Relevant to claim means [criteria]."

The output of Step 1 is your classification rubric. It will be 1 to 2 pages. For the prompt structure, see the lawyer's guide to prompt engineering.

Step 2: Validate the Rubric Against 25 Documents (1-2 hours)

Run the rubric on 25 documents the senior partner has already personally reviewed. Compare AI classifications to human ones. Where do they diverge? Where is the AI consistently wrong? Refine the rubric.

This step is non-negotiable. It's how you build justified confidence in the AI's classifications.

Step 3: Run the Full Review (depends on volume)

Modern AI tools that handle this at scale:

• Everlaw with built-in AI: $150 to $400/user/month
• Reveal AI: enterprise pricing
• Logikcull (now part of Reveal): cost-effective for mid-market
• Disco Cecilia AI: top-tier eDiscovery integration
• Onna + Claude API: custom builds

For boutique firms doing volumes under 100K documents, you can sometimes build a workflow directly with Claude API ($3 to $15/million tokens) for under $5,000 total. We've built custom systems like this through our Legal AI Tools service.

Step 4: Privilege QC (always human)

Privilege determinations should always be human-confirmed. The AI can identify candidates - the lawyer must confirm. This is true even when AI accuracy is high. Privilege errors cause waiver, sanctions, and malpractice claims. Spend the time.

Step 5: Production and Audit Trail

Modern AI review tools produce audit trails: which model, which version, which prompts, which documents, which classifications. Save all of this. It's both protective for your firm and increasingly expected (and sometimes required) in production negotiations.

Privilege Detection: The Highest-Value Use Case

The single highest-value AI document review use case is privilege identification. The reason: privilege determinations are high-stakes, voluminous, and pattern-recognizable.

The modern workflow:

1. AI scans every document for privilege indicators (lawyer names, legal advice patterns, work product signals).
2. AI scores each document for privilege likelihood (0 to 100).
3. Human attorney reviews everything above a threshold (typically 30+).
4. Below threshold, spot-check 10 to 20% as QC.

This combination catches 95%+ of privileged documents at a fraction of the cost of traditional review.

Production-Ready Privilege Log Generation

Several modern tools (Everlaw, Disco, Reveal) can now generate first-draft privilege logs automatically:

• Document description
• Author / recipient identification
• Date
• Privilege basis (AC, WP, common interest)
• Description of withheld content

The logs require human review and signature - but the first draft is 80% complete.

The Hot Document Hunt

Beyond classification, AI can run "hot document" searches that go far beyond keyword search:

"Find all documents that suggest knowledge of [issue] before [date], including documents that don't use the obvious keywords but describe the situation in roundabout terms."

This is the kind of analysis that used to require senior associates reading every document. Now AI does the first pass.

Cost Comparison: Real Numbers

For a 200,000-document review in a mid-size commercial litigation:

The cost differential is real. So is the speed differential. And accuracy is generally higher with modern AI review than with rent-a-coder document review centers.

Risks and Verification

The verification standards are the same as for any AI workflow:

• Privilege determinations: always human-confirmed.
• Hot documents: always human-reviewed before production.
• Spot-check the relevance categorizations for at least 5 to 10% of documents.
• Maintain audit trails.

See our deeper guide on AI hallucinations in legal work for the verification framework that applies here too.

Ethics and Disclosure

Several jurisdictions now require disclosure of AI-assisted document review when negotiating production protocols. The trend is clear: opposing counsel and courts increasingly expect to know how AI was used.

Be prepared with:

• A written description of your AI tools and methodology
• Accuracy benchmarks from your validation
• Privilege QC procedures
• Audit trail capabilities

For the broader ethics framework, see AI ethics for lawyers.

Where to Start

If your firm has document review work but no AI workflow:

1. Identify your highest-volume document type (PI medical records, contracts, employment files, etc.).
2. Build the rubric using Claude or ChatGPT.
3. Validate against 25 partner-reviewed documents.
4. Pilot on one matter before scaling.

If you want help building this into a proper firm workflow, our Legal AI Tools service builds custom document review systems for firms. Or take the AI Readiness Assessment to see if your firm is positioned to capture these savings.

Document review went from a profit center for ALSPs to a competitive advantage for firms that adopted AI first. The window to lead this is still open - but not for much longer.