Unnamed Am Law 100 Firm

E-discovery document review sits at the intersection of legal risk and operational cost for Am Law 100 firms. In large commercial litigation and regulatory matters, document populations routinely reach hundreds of thousands or millions of records, requiring attorneys to make consistent relevance determinations at scale. Manual review is not only expensive — linear review by contract attorneys can cost millions of dollars per matter — but introduces inconsistency that opposing counsel and courts can challenge. The core tension is between recall (finding everything relevant) and precision (not over-producing privileged or irrelevant documents). Achieving both simultaneously through human review alone is effectively impossible at enterprise litigation scale.

The firm deployed DISCO Auto Review, a machine learning-based predictive coding platform, to automate relevance classification across large document sets. The system uses supervised learning: a seed set of attorney-coded documents trains a model that propagates those judgments across the full population, continuously refining as reviewers code additional examples. Rather than replacing attorney review, the tool functions as a force multiplier — prioritizing the highest-confidence relevant documents for human review while suppressing clearly non-responsive records. DISCO's platform integrates directly into the review workflow, enabling a continuous active learning loop where model confidence scores guide review sequencing. This approach allowed the firm to apply consistent, documentable decision logic across the entire document universe rather than relying on keyword search or linear batching.

The firm achieved 89% recall and 99.5% precision — metrics that satisfy both the practical and legal requirements of defensible e-discovery production.

• 99.5% precision means fewer than 1 in 200 produced documents was non-responsive, dramatically reducing the risk of inadvertent disclosure
• 89% recall demonstrates the model captured the substantial majority of relevant documents, meeting the proportionality standard courts expect

Together, these figures represent a recall-precision balance that is difficult to achieve through manual review, where precision typically degrades as reviewers rush to meet deadlines. The results establish a documented, reproducible process that can withstand judicial scrutiny and adversarial challenge.

• Train on quality, not quantity: the model's performance depends on the consistency of the seed set — invest attorney time in early coding decisions rather than maximizing throughput.
• Precision and recall are not mutually exclusive when the model is properly calibrated; firms should not accept trade-offs between them without testing.
• Defensibility requires documentation: maintain audit trails of model training, coding decisions, and recall validation to support Rule 26 disclosures.
• Continuous active learning outperforms one-shot training — re-train as reviewers code deeper into the population.
• Early-stage AI review adoption in Am Law 100 matters signals that predictive coding has cleared the bar for complex commercial litigation.