The eDiscovery Cost Crisis
Electronic discovery has become one of the most expensive components of modern litigation. The volume of electronically stored information (ESI) continues to grow exponentially, with IDC estimating that global data creation reached 181 zettabytes in 2025. For litigation teams, this means that a single matter can involve reviewing millions of documents at costs that often exceed the amounts in controversy.
A 2025 RAND Institute for Civil Justice study found that eDiscovery costs represent 60-80% of total litigation expenses for complex commercial matters. Document review alone, the process of evaluating each document for relevance, privilege, and responsiveness, accounts for 70% of those eDiscovery costs. At traditional review rates of 40-60 documents per hour per reviewer, a collection of 2 million documents requires approximately 40,000-50,000 review hours.
At blended contract reviewer rates of $50-$75 per hour, the review alone costs $2 million to $3.75 million. AI-powered eDiscovery fundamentally disrupts this cost structure while simultaneously improving review quality and consistency.
Understanding Technology Assisted Review
TAR 1.0: The Foundation
Technology Assisted Review (TAR) was the first generation of AI applied to document review. TAR 1.0, also known as simple passive learning, follows a straightforward workflow. A senior attorney reviews a random seed set of documents, typically 1,000-2,000 documents, coding each as relevant or not relevant. The AI model trains on these coding decisions and then ranks the remaining documents by predicted relevance.
The ranked results allow the review team to focus on the most likely relevant documents first, reaching a defensible cutoff point beyond which the probability of relevance is statistically negligible. Courts have consistently upheld TAR 1.0 as a defensible review methodology since Judge Andrew Peck's landmark decision in Da Silva Moore v. Publicis Groupe in 2012.
TAR 1.0 typically reduces the volume of documents requiring human review by 60-75%. However, it has limitations. The seed set must be representative of the entire collection, the model does not adapt based on review team feedback after initial training, and it treats relevance as a binary classification.
TAR 2.0: Continuous Active Learning
TAR 2.0 introduced continuous active learning (CAL), a significant advancement over the seed-set approach. In CAL workflows, the AI model continuously updates based on every coding decision made by any reviewer throughout the review process. There is no separate training phase; learning and review happen simultaneously.
The AI prioritizes the most informative documents for human review, those where the model is least certain of the correct classification. This approach concentrates human judgment where it is most needed while allowing the AI to confidently classify clear-cut documents.
CAL workflows deliver superior results:
- **Higher recall rates**: CAL typically achieves 90-95% recall compared to 75-85% for TAR 1.0
- **Greater efficiency**: Documents requiring human review reduced by 75-85%
- **Faster convergence**: The model reaches defensible accuracy levels more quickly
- **Adaptability**: The model adjusts to evolving relevance criteria as the legal team refines its understanding of the case
TAR 3.0: Multi-Modal AI Review
The current generation of AI eDiscovery, which some vendors call TAR 3.0, leverages large language models and multi-modal AI to understand documents in ways that previous generations could not. These systems comprehend document context, identify conceptual relevance, analyze images and handwriting, and understand the significance of communication patterns.
TAR 3.0 capabilities include:
- **Conceptual understanding**: The AI understands the subject matter of documents, not just keyword matches, enabling identification of relevant documents that discuss key issues without using expected search terms
- **Communication pattern analysis**: AI maps communication networks to identify key custodians, unusual communication patterns, and potential evidence of coordination
- **Temporal analysis**: Automated timeline construction from document metadata and content, identifying critical time periods and event sequences
- **Multi-language processing**: Native understanding of documents in multiple languages without requiring separate translation workflows
- **Image and attachment analysis**: AI processing of embedded images, charts, and handwritten notes that previous systems could not evaluate
Predictive Coding: Deep Dive
How Predictive Coding Works
Predictive coding is the specific AI methodology within TAR that classifies documents based on machine learning models trained on human coding decisions. The process involves several key stages:
**Feature extraction**: The AI extracts features from each document, including text content, metadata, structural elements, and relationships to other documents. Modern systems use transformer-based embeddings that capture semantic meaning rather than simple bag-of-words representations.
**Model training**: Human coding decisions teach the model what constitutes a relevant, privileged, or responsive document for the specific matter. The model learns not just from the words present but from the overall patterns that distinguish relevant from irrelevant documents.
**Classification and scoring**: The trained model assigns relevance scores to unreviewed documents. These scores represent the model's confidence that a document would be coded as relevant by the human reviewers.
**Validation**: Statistical validation confirms that the model's classifications meet defensibility standards, typically measured through precision, recall, and F1 scores on held-out validation sets.
Defensibility Standards
Courts across jurisdictions have established clear standards for defensible predictive coding workflows. Key requirements include:
- **Transparency**: Parties must disclose their use of predictive coding and the methodology employed
- **Validation metrics**: Statistical measures must demonstrate that the review achieved acceptable recall and precision rates
- **Quality control**: Documented QC protocols must verify ongoing review accuracy
- **Senior attorney involvement**: Subject matter experts must direct the training process and validate results
- **Documentation**: Complete records of the training process, validation results, and quality control measures
The 2025 Sedona Conference best practices recommend a minimum recall rate of 80% for standard matters and 90% for high-stakes matters, with precision rates of at least 60%. AI-powered reviews routinely exceed these thresholds.
AI-Powered Privilege Review
The Privilege Review Challenge
Privilege review is the most sensitive and expensive component of document review. Inadvertent production of privileged communications can waive attorney-client privilege, potentially exposing an organization to catastrophic consequences. The stakes demand accuracy, but the volume of potentially privileged documents makes manual review prohibitively expensive.
Traditional privilege review requires senior attorneys to evaluate each potentially privileged document, assess the communication participants, determine whether the dominant purpose was legal advice, and create detailed privilege log entries. At large-firm partner rates, privilege review costs can exceed $100 per document.
AI Privilege Classification
AI privilege review systems transform this process through intelligent classification and automation:
**Attorney identification**: AI identifies all attorneys (in-house and external) across the document collection, building comprehensive attorney rosters that improve classification accuracy. The system resolves name variations, nicknames, and role changes over time.
**Communication analysis**: The AI evaluates communication patterns, participant roles, and content to assess privilege likelihood. Documents involving identified attorneys discussing legal matters receive high privilege scores, while purely business communications between non-attorneys receive low scores.
**Purpose analysis**: Modern AI systems can assess whether the dominant purpose of a communication was seeking or providing legal advice, even in mixed-purpose communications. The system identifies legal analysis, legal recommendations, and discussions of litigation strategy.
**Privilege log generation**: For documents classified as privileged, AI automatically generates privilege log entries including document description, participants, date, and privilege basis. Senior attorneys review and approve these entries rather than creating them from scratch.
Organizations implementing AI privilege review report 70-80% reduction in privilege review costs while achieving higher consistency than manual review. The AI eliminates the variability that occurs when multiple reviewers apply privilege standards differently.
Document Classification Beyond Relevance
Issue Coding
Modern eDiscovery AI classifies documents across multiple dimensions simultaneously. Beyond simple relevance determinations, AI assigns issue codes that categorize documents by the specific legal or factual issues they address. This multi-label classification enables the legal team to quickly assemble document sets for specific deposition topics, motion arguments, or trial themes.
Hot Document Identification
AI systems identify "hot" documents, those with the highest potential impact on the case, by analyzing content significance, communication context, and relevance to key issues. Rather than discovering these critical documents randomly during review, the legal team receives prioritized access to the most important evidence early in the review process.
Confidentiality and Sensitivity Classification
AI classifies documents by confidentiality level, identifying trade secrets, personally identifiable information, protected health information, and other sensitive content. This classification supports appropriate handling designations and redaction workflows. For teams managing data privacy alongside litigation, our article on [AI privacy management platforms](/blog/ai-privacy-management-platform) covers the intersection of these disciplines.
Implementing AI eDiscovery Workflows
Pre-Collection Analytics
AI analysis should begin before collection, not after. Pre-collection analytics use AI to analyze custodian data sources and estimate the volume, relevance, and cost of collecting from each source. This enables targeted collection strategies that reduce downstream volume and cost.
By analyzing email metadata, file system structures, and communication patterns before full collection, legal teams can prioritize the most productive custodians and data sources. Pre-collection analytics typically reduce collection volumes by 30-50% compared to broad custodial collections.
Processing and Analytics
After collection, AI-powered processing goes beyond traditional deduplication and file extraction. Modern processing platforms apply:
- **Near-duplicate detection**: Identifying documents that are substantially similar, grouping them for efficient review
- **Email threading**: Reconstructing complete email conversations and enabling review of the most inclusive thread rather than every individual message
- **Concept clustering**: Grouping documents by topic to enable focused, efficient review workflows
- **Foreign language identification**: Automatically detecting document language and routing to appropriate reviewers or translation workflows
Review Workflow Design
Effective AI eDiscovery requires thoughtful workflow design that integrates human judgment with AI capabilities. Best-practice workflows include:
**First-pass AI classification**: The AI scores and classifies all documents, creating batches organized by relevance score, issue, and complexity.
**Prioritized human review**: Reviewers focus on documents where human judgment adds the most value: borderline relevance scores, privilege determinations, and hot document verification.
**Continuous quality control**: AI monitors reviewer consistency, flagging outlier coding decisions for supervisory review. This real-time QC is far more effective than traditional post-hoc sampling.
**Iterative refinement**: As reviewers code documents, the AI model updates continuously, improving its classifications for remaining documents.
Quality Control and Validation
AI eDiscovery quality control operates at multiple levels:
- **Reviewer consistency monitoring**: AI tracks each reviewer's coding patterns and flags statistical outliers that may indicate errors or misunderstanding of coding criteria
- **Model performance metrics**: Continuous monitoring of precision, recall, and F1 scores ensures the AI model maintains defensible accuracy levels
- **Elusion testing**: Random sampling of documents classified as non-relevant to verify that the model is not missing relevant documents
- **Overturn rate tracking**: Monitoring the rate at which QC reviewers overturn first-level coding decisions to identify training needs
Cost and Timeline Impact
Quantifying the Savings
The economic impact of AI eDiscovery is dramatic. For a matter involving 2 million documents:
| Approach | Review Hours | Cost | Timeline | |----------|-------------|------|----------| | Linear manual review | 40,000-50,000 | $2.5M-$3.75M | 6-9 months | | TAR 1.0 | 10,000-15,000 | $625K-$1.1M | 2-3 months | | TAR 2.0/CAL | 5,000-8,000 | $312K-$600K | 4-8 weeks | | TAR 3.0 multi-modal | 3,000-5,000 | $187K-$375K | 2-4 weeks |
These figures represent typical ranges; actual results vary based on document complexity, coding complexity, and the specific AI platform used. But the trend is clear: each generation of AI eDiscovery delivers step-function improvements in cost and timeline.
Beyond Direct Cost Savings
AI eDiscovery also delivers indirect benefits that are harder to quantify but equally important:
- **Earlier case assessment**: Faster access to key documents enables earlier and more informed case strategy decisions
- **Better settlement outcomes**: Comprehensive document understanding supports stronger negotiation positions
- **Reduced attorney burnout**: Eliminating tedious manual review improves attorney satisfaction and retention
- **Consistent quality**: AI review eliminates the quality variability inherent in large manual review teams
For legal teams exploring broader automation of their practice, our guide on [AI automation for legal firms](/blog/ai-automation-legal-firms) provides additional context on how eDiscovery fits within a comprehensive legal technology strategy.
Accelerate Your eDiscovery with AI
The trajectory of eDiscovery AI is clear: each year brings more powerful capabilities, lower costs, and broader judicial acceptance. Organizations that invest in AI eDiscovery capabilities now build institutional expertise that compounds over time, making every subsequent matter faster and more cost-effective.
Whether you are managing a single complex litigation or building an enterprise eDiscovery program, AI-powered tools should be at the center of your strategy.
[Contact our team](/contact-sales) to discuss how the Girard AI platform can transform your eDiscovery operations, or [sign up](/sign-up) to explore our document review automation capabilities.