Unnamed Am Law 100 Firm

In fall 2024, a three-attorney team at a leading Am Law 100 firm faced a defining e-discovery challenge: reviewing 126,000 documents for production in a large-scale government investigation tied to a potential civil litigation matter, under a compressed timeline and constrained budget. The document set consisted primarily of emails, email attachments, and Microsoft Teams messages requiring coding for responsiveness across nearly two dozen distinct issue codes. The traditional path — a managed review staffed by approximately 20 contract attorneys working for four weeks — was neither economical nor fast enough. E-discovery costs already consume nearly 80% of total litigation spend, making manual review at this scale a significant client-cost liability.

The firm deployed EverlawAI Assistant Coding Suggestions, a large language model (LLM)-powered e-discovery tool that automates document coding through natural-language prompts. Working alongside managed services provider Right Discovery, the team developed a structured three-stage workflow: drafting initial code criteria with case background and context, iterating prompts against three representative document subsets, and finally running the model at scale across the full 126,000-document corpus once precision, recall, and F1 benchmarks were met. Right Discovery validated results using the same statistical metrics applied to traditional technology-assisted review (TAR), ensuring the AI-assisted workflow met defensibility standards before full deployment. The model produced a four-tier output — Yes, Soft Yes, Soft No, No — with an explanatory rationale for each document, enabling targeted human oversight of borderline cases.

After approximately 15 hours of prompt iteration, the team of five (three attorneys and two support staff) coded the full 126,000-document set in under 24 hours — a task that would have required 20 contract attorneys over four weeks using traditional managed review.

• 50–67% reduction in overall document review time
• 90%+ accuracy on recall and precision, at or above first-level attorney benchmarks
• Precision: 0.63 / Recall: 0.85 / F1: 0.72 on overall responsiveness in the final validation set
• One quarter of the personnel required versus a comparable managed review
• Greater coding consistency than human reviewers on identical documents

The lead attorney noted the solution delivered clear, demonstrable value to the client without sacrificing work quality.

• Prompt iteration (~15 hours) is more accessible than constructing complex Boolean search strings, lowering the technical barrier for legal teams adopting LLM-assisted review.
• A four-tier classification system (Yes / Soft Yes / Soft No / No) concentrates human review effort on genuinely ambiguous documents, maximizing attorney time where judgment is irreplaceable.
• Validating AI outputs with standard TAR metrics (precision, recall, F1) before full deployment is essential for defensibility in government investigations and adversarial proceedings.
• Prompts and AI rationales generated during initial review create a durable, reusable record that simplifies future production requests in long-running matters.