Industry newsPrivilege review ai
AI Privilege Review Can Cut Costs 75% — If You Can Afford Two Weeks of Attorney Time First
In an April 8, 2026 post on Arnold & Porter's eData Edge blog, attorney Bryan M. Marra described using Relativity's aiR for Privilege — a generative-AI privilege review tool — on a large antitrust document review. Against a traditional linear review staffed by contract attorneys, the tool produced roughly 75% in cost savings on that matter. The catch: getting there required close to two weeks of focused work by a firm attorney mapping privilege relationships across every law firm, lawyer and third party in a dataset of about 500,000 documents, before the model could make a single prediction.
How much money does the AI actually save?
Roughly 75% versus a linear review staffed by contract attorneys, according to one attorney's account of a single antitrust matter — not an audited benchmark.
Marra is explicit that the number is case-specific and depends on dataset size and how much human review still gets layered on top. That's a useful caveat trade press coverage of "AI cuts review costs" claims routinely drops: the baseline being beaten (contract-attorney linear review) is the most expensive option available, not the cheapest defensible one. Firms already running lean TAR-based workflows should ask what the comparison looks like against their own baseline, not against the priciest alternative a vendor can find.
What's the hidden cost before you see any savings?
Close to two weeks of a senior, privilege-savvy attorney's time spent classifying every entity in the document set as aligned, adverse or neutral.
Marra says that effort was "similarly time-intensive" to the subject-matter-expert training a traditional TAR model requires. That's a meaningful data point for buyers: generative AI privilege review does not remove the up-front human investment that predictive coding already demanded — it relocates it, from labeling sample documents to mapping an entity graph. For a firm evaluating a new tool mid-litigation, that two-week clock needs to be built into the deadline, not treated as a rounding error.
Does the AI replace human review of privileged documents?
No — Relativity tunes aiR for Privilege to favor recall over precision, so a broad first pass still needs substantial human QC.
Marra confirms this design choice held up in practice: the tool flagged documents broadly to avoid missing anything privileged, which meant meaningful review and quality control remained downstream. Buyers should press any vendor on where precision and recall were measured, on what document population, and by whom — Marra's account is candid about the tool's categorization being more useful than its written rationales, which is exactly the kind of granular finding a headline metric obscures.
Is this cheaper than standard predictive coding?
Not necessarily — savings scale with dataset size, and the author is explicit that smaller or lower-volume matters may not clear the cost bar.
This lines up with a point this site has raised before: as every legal AI vendor converges on a "good enough" underlying model, the differentiator shifts to workflow economics — setup time, QC burden, output structure — not raw accuracy claims.
Frequently asked questions
Is aiR for Privilege the only generative AI tool doing this?
Marra's account covers Relativity's product specifically; the post doesn't compare it against other vendors' generative privilege-review offerings.
Does the 75% figure include the two weeks of setup time?
Marra doesn't say. That ambiguity alone is worth asking any vendor to itemize before you accept a headline savings number.
Sourced from Arnold & Porter's eData Edge blog.
The original report
Industry headlines from other publications. Each links to the original reporting on the publisher's own site.