E-discovery AI: a federal ruling and four measured outcomes

Will Drewes

Founder, Fern Strategy · LinkedIn · 2 min read

The court applied the old test

A federal judge recently allowed LinkedIn to use generative AI for document review in discovery, under the same reasonableness analysis courts already apply to technology-assisted review and predictive coding. GenAI review is now a method that has to be defensible, documented and proportionate, the same bar TAR has cleared for a decade.

That is the useful part of the ruling. Courts are still asking the questions they always asked: how was it validated, who supervised it, can you show your work. A process that can't answer those gets no credit for speed.

The outcomes

Privilege review, Fortune 100 telecom. A case study cited by Relativity describes AI-assisted privilege review that ran 80% faster than manual review and cut costs by 75%. Privilege review is the highest-stakes stretch of a large matter, and getting it wrong costs more than any time sheet shows. An 80% speed gain on that task specifically, rather than on first-pass culling, is the number I'd read twice.

Project compression, Morae. On a matter with a one-week deadline, Morae used AI-assisted review, cut costs by 75% and saved more than 25 days of project time. Fitting 25 days of work into a week without cutting review quality changes how legal ops plans staffing.

First-pass throughput, DISCO. DISCO's Cecilia Auto Review is rated at up to 32,000 documents per hour. That is a throughput ceiling rather than an outcome on a named matter, and it starts to matter on matters with millions of documents.

Review time, Am Law 100 firm. Everlaw published a case study in which an Am Law 100 firm cut document review time by two-thirds with generative AI. The firm is unnamed beyond its ranking, but review time is usually the largest line item in discovery, so two-thirds is a large number.

What this means if you run legal or compliance

Every result above came with a documented, supervised process behind it, and the LinkedIn ruling approved GenAI on exactly that condition. Build-versus-buy is a distraction at this stage. The harder question is whether your current process, with whatever tools, produces an audit trail that survives a reasonableness challenge. The speed and cost figures are the output. The audit log is what you hand opposing counsel when they ask how you got there.

Start with the workflow you can document.