Redgrave LLP Study Compares Relativity aiR for Review With Active Learning Managed Review on 45,000 Documents

In a head-to-head study at Redgrave LLP, the generative AI review workflow Relativity aiR for Review required approximately 18 hours of attorney time, while a traditional first-pass managed review workflow using active learning required approximately 1,123 hours from a 24-person team over seven business days, a roughly 98 percent reduction in cumulative human hours. aiR for Review achieved 88 percent recall compared with 64 percent for active learning, and an elusion rate of 1 percent compared with 3 percent. The Redgrave LLP article states: "Faster review is valuable, but faster review that also improves completeness is far more valuable." The team chose a deliberately difficult population of approximately 45,000 documents from a real-world public data set, with a nuanced responsiveness standard tied to pharmaceutical marketing, controlled substances, and federal compliance obligations. Documents were not responsive simply because they mentioned opioids, sales activity, or drug promotion; they had to contain evidence related to compliance with, violation of, or reckless disregard of federal requirements. A subject-matter expert conducted a blind review of a random sample to establish the ground truth against which both workflows were measured. The document population had low richness, meaning responsive documents were relatively rare. The study also tested human plus AI. The expert first reviewed the validation sample blind, without seeing aiR for Review's scores, rationales, or citations. He then revisited documents he had coded not responsive but that aiR for Review had predicted responsive, this time with its rationale and citations visible. Of the 151 documents re-reviewed, he changed 10 calls to responsive, roughly one out of every 15, and in most instances he stood by his original decision. aiR for Review also flagged more documents for potential review and had lower precision than the active learning workflow, 29 percent compared with 39 percent. Redgrave describes this as a wider net: more documents would be routed for potential human review, but the net also caught more responsive documents and missed fewer important ones. Redgrave states that the study is not the final word on generative AI review and should not be read as a universal benchmark for every matter, since review quality depends on the data, protocol, workflow, validation design, and the judgment of the people overseeing the process.