AI eDiscovery Review Workflow for Law Firms

eDiscovery review is one of the largest hour sinks in litigation practice. Document populations of 100,000 to millions are routine. Without AI, attorney hours scale linearly with document volume — and matters become economically unviable at scale.

AI-assisted review changes the math. Volume is no longer the bottleneck. The bottleneck is workflow design, attorney supervision, and verification discipline.

Here's the operator workflow.

What AI handles in eDiscovery

• Predictive coding / Technology-Assisted Review (TAR): Categorize documents as relevant, non-relevant, or privileged based on attorney training set
• Concept clustering: Group similar documents for batch review
• Communication analysis: Identify key custodians and communication patterns
• Privilege detection: Flag potentially privileged documents
• Key fact identification: Surface documents addressing specific issues
• Production preparation: Format and redact for production

What attorneys handle

• Initial classification training (seed set)
• Verification of AI categorization
• Privilege review confirmation
• Strategic decisions on production
• Discovery dispute resolution
• Final attorney sign-off

The attorney remains accountable for the production. AI accelerates; attorneys decide.

The tools

Relativity with aiR: Industry standard. Has been the dominant eDiscovery platform for years. The aiR AI layer adds modern AI capabilities.

DISCO: Modern challenger. Strong AI features, cloud-native architecture. Growing share at firms switching from Relativity.

Reveal (acquired Logikcull): AI-powered eDiscovery, growing capability.

Everlaw: Modern eDiscovery with strong AI features.

Brainspace (acquired by Veritone): Analytics-focused.

Zylab: International strength.

Choice depends on firm size, matter type, and existing infrastructure. Most AmLaw firms run Relativity. DISCO is gaining share at firms valuing modern architecture.

The standard workflow

Phase 1: Setup and processing (days to weeks)

• Collect documents from custodians and sources
• Process documents (deduplication, near-duplication, threading)
• Prepare for review platform ingestion
• Set up review universe

Phase 2: First-pass AI categorization (days)

• Train AI on seed set of attorney-coded documents (~500-2000 documents)
• Run AI prediction across full document population
• AI returns predicted relevance categorization with confidence scoring
• Attorney reviews AI output and refines training

Phase 3: Iterative training (days to weeks)

• Review batches of AI-categorized documents
• Code variances (where attorney disagrees with AI)
• Refine training set
• Re-run AI prediction
• Repeat until model converges to acceptable accuracy

Phase 4: Production preparation (days)

• Apply final AI categorization
• Privilege review of relevant population
• Redaction of privileged or protected information
• Production formatting

Phase 5: Sign-off and production (days)

• Lead attorney reviews production set
• Confirms relevance, privilege, and production accuracy
• Produces to opposing counsel
• Maintains production log

Where AI compresses time most

For a typical 500,000-document matter:

Without AI:

• Linear attorney review at typical pace (50-100 docs/hour) = 5,000-10,000 attorney hours
• Cost at $300-500/hour blended associate rate = $1.5M-5M

With AI:

• AI categorizes full population
• Attorneys review and verify ~30-40% of documents (relevance, privilege)
• Total attorney review: 1,500-3,000 hours
• Cost: $450k-1.5M

Net savings: 60-70% of review hours and cost. The savings scale with document population — large matters benefit most.

Defensibility of AI-assisted review

Courts have generally accepted predictive coding / TAR as defensible since Da Silva Moore v. Publicis Groupe (2012). Current state:

• AI-assisted review is defensible when properly executed
• Documentation of the process is critical
• Attorney supervision is required throughout
• Both parties may use AI; transparency about AI use is increasingly expected

The defensibility framework:

• Clear protocol established at outset
• Sound seed set training
• Iterative refinement documented
• Reasonable validation testing
• Attorney sign-off on final categorization

Privilege protection

AI assists but doesn't replace privilege review. Standard practice:

• AI flags potentially privileged documents (typically based on attorney-client communications, work product indicators)
• Attorneys verify each flagged document
• Privilege log generated for produced documents that reference privileged matter
• Quality control sampling of "non-privileged" categorization

Privilege errors are the highest-stakes errors in AI-assisted review. Build redundancy into the workflow.

Compliance and ethics

eDiscovery AI touches:

• Rule 26 cooperation obligations
• Rule 34 production scope
• Privilege and work product protection
• Sanctions exposure (Rule 37)
• State bar AI guidance

ABA Formal Opinion 512 (2024) discusses competent AI use generally. State-specific eDiscovery guidance exists in some jurisdictions.

The attorney remains accountable. AI accelerates the work; the attorney signs the production.

What can go wrong

Pattern 1: Inadequate seed set. Poor training leads to poor AI prediction. Train carefully on representative documents.

Pattern 2: Insufficient iteration. First-pass AI accuracy is rarely good enough. Refine through multiple rounds.

Pattern 3: Privilege miss. AI categorizes a privileged document as non-privileged, gets produced. Build redundant privilege review.

Pattern 4: Production scope errors. AI categorizes too broadly or narrowly. Validate against Rule 26 scope.

Pattern 5: Lack of documentation. Process isn't documented; defensibility is weakened. Document everything.

Economics by matter size

Small matter (under 10,000 docs): AI may not justify setup. Manual review competitive.

Mid-size matter (10,000-100,000 docs): AI compresses 40-60% of review time. Clear ROI.

Large matter (100,000-1M docs): AI essential. Without it, matter becomes economically unviable.

Mega matter (1M+ docs): AI mandatory. The only way to manage at scale.

What we deploy

For firms working with us on eDiscovery AI:

• Standard platform selection (Relativity, DISCO, or Reveal based on firm needs)
• Workflow design for AI-assisted review
• Attorney training on TAR protocols
• Compliance documentation framework
• Quality control sampling process
• Defensibility documentation

Setup at firm level: $50-200k initial + per-matter costs scaling with document volume.

Bottom line

AI-assisted eDiscovery review in 2026 is essential at any meaningful scale. The economics don't work without it for matters above 100,000 documents. The technology is mature, the defensibility framework is established, and the workflow is well-understood.

Done right, AI eDiscovery delivers 60-70% review hour reduction with maintained or improved accuracy. Done wrong — without proper training, without supervision, without verification — it creates production errors, privilege breaches, and sanctions exposure.

The discipline is what makes it work. Train the AI carefully. Iterate the model. Supervise the output. Verify privilege. Document the process. The defensibility comes from the discipline.