Generative AI Document Review Cuts Human Review Hours 98% While Improving Recall at Redgrave LLP

Redgrave LLP wanted to move past theoretical debate about generative AI in legal review and test it directly against a traditional managed review workflow. The firm chose a deliberately difficult population of roughly 45,000 documents from a real-world public data set, governed by a nuanced responsiveness standard tied to pharmaceutical marketing, controlled substances, and federal compliance obligations, where documents were not responsive merely for mentioning opioids or sales activity.

The firm ran a head-to-head study pitting Relativity's aiR for Review, a generative AI review workflow, against a traditional first-pass managed review using active learning performed by a 24-person team over seven business days. A subject-matter expert conducted a blind review of a random sample to establish ground truth, then reassessed a set of documents where his coding disagreed with aiR for Review's predictions, this time with visibility into the tool's rationale and citations.

The account is written by Redgrave's co-managing partner. In the study, the aiR for Review workflow required approximately 18 hours of attorney time, versus approximately 1,123 hours from a 24-person team over seven business days for the active learning workflow, a roughly 98 percent reduction in cumulative human hours. aiR for Review also achieved higher recall (88% vs. 64%) and lower elusion (1% vs. 3%). Of the 151 documents the expert re-reviewed after seeing aiR for Review's rationale and citations, he changed 10 calls to responsive. aiR for Review flagged more documents for potential review and had lower precision than the active learning workflow (29% vs. 39%); the author notes the population had low richness, which tends to lower precision in any workflow. The author also says the study should not be read as a universal benchmark for every matter.

• In this study, aiR for Review required far fewer human hours and also delivered higher recall and lower elusion than active-learning managed review.
• Generative AI can act as a second lens on human judgment, surfacing specific documents worth a second look rather than replacing expert review outright.
• A wider net brought lower precision (more documents flagged for review), a tradeoff the author says may be reasonable where missing key documents is the greater risk.