Large-scale litigation and regulatory investigations routinely involve reviewing millions of documents to identify those that are relevant, privileged, or responsive to discovery requests. Manual review at this scale costs millions and takes months. Technology-assisted review (TAR) using AI reduces review populations by 80-95% while maintaining quality that meets or exceeds manual review standards.

The Problem

A typical corporate investigation or complex litigation matter may involve 5-10 million documents collected from custodians’ email, file shares, and messaging systems. Manual review at a rate of 50 documents per hour per reviewer requires an enormous investment in contract reviewers and time. At the same time, courts increasingly expect parties to use efficient review methods - proportionality requirements in procedural rules discourage brute-force manual review when technology alternatives exist.

The specific challenges include: identifying relevant documents across diverse topics and file types, detecting privileged communications that must be withheld, maintaining consistency across a review team of dozens of reviewers, and completing the review within court-imposed deadlines.

AI Approach

Continuous active learning (CAL) - SageMaker hosts the active learning model that iteratively refines document classifications. A senior reviewer codes an initial seed set. The model trains on these examples, scores the remaining population, and presents the most informative documents (those near the decision boundary) for the next round of human review. Each coding decision improves the model. CAL achieves recall rates of 85-95% while reviewing only 5-20% of the total population.

Concept clustering - OpenSearch clusters documents by topic using embedding-based similarity. Reviewers can assess entire clusters rather than individual documents, accepting or rejecting groups of similar documents. This is particularly effective for large populations where broad topic categories can be quickly dispositioned.

Privilege detection - Bedrock analyzes document content to identify communications likely protected by attorney-client privilege or work product doctrine. The model looks for legal advice content, attorney involvement, and privilege-indicating language. Given the consequences of inadvertent privilege waiver, privilege detection operates at high sensitivity with human review of all flagged documents.

Near-duplicate detection - Email threads and document versions create massive duplication. Near-duplicate detection groups documents that are substantially similar, allowing reviewers to code the group by reviewing a representative sample. This typically reduces the effective review population by 30-50%.

Architecture

Documents are ingested from the ediscovery processing platform into S3. Text extraction and metadata processing populate OpenSearch for full-text and semantic search. SageMaker hosts the active learning model with a review interface that presents documents for coding and captures reviewer decisions in real-time. Bedrock powers the privilege detection layer. DynamoDB tracks review progress, coding decisions, and quality control metrics.

Key Considerations

Defensibility - Courts scrutinize AI-assisted review methodologies. The process must be documented, the validation methodology (statistical sampling of the unreviewd population) must be sound, and the recall metrics must be demonstrable. Well-implemented TAR is legally defensible and increasingly expected.

Quality control - Random sampling of both the reviewed and unreviewed populations validates model performance. Reviewer consistency checks ensure that human coding decisions are reliable training data for the model.

Subject matter expertise - The seed set reviewers must be experienced lawyers who understand the substantive legal issues. Poor seed set quality produces poor model performance regardless of the algorithm.

Multilingual review - Cross-border matters involve documents in multiple languages. The embedding models and classification pipeline must handle multilingual content, either through multilingual models or language-specific pipelines.

Next Steps

Engage with ediscovery vendors who provide TAR platforms integrated with AWS infrastructure. For a pilot matter, run TAR in parallel with planned manual review and compare recall, precision, and total cost. Use the results to establish an organizational protocol for when TAR is appropriate and how to validate its outputs for court submission.