Technology-assisted review
Technology-assisted review (TAR) is the use of a machine-learning classifier, trained by attorney coding decisions on documents from the collection itself, to rank or classify the remaining documents by their likelihood of being responsive. Review is then worked in rank order and stopped at an agreed point, rather than run linearly through every document in the set.
Also called: TAR · predictive coding · continuous active learning · CAL · TAR 1.0 · TAR 2.0 · computer-assisted review
Updated September 6, 2026
How does technology-assisted review work?
A classifier is trained on documents from the matter, coded responsive or not by attorneys who know the case. It learns which textual features separate the two, then scores every unreviewed document. Reviewers work the highest-scoring documents first. Because responsive documents are concentrated at the top of the ranking, the rate at which new responsive documents appear falls as review proceeds, and that curve, measured against a statistically sampled estimate of total responsive volume, is what justifies stopping.
Continuous active learning collapses training and review into one loop. Every coding decision retrains the model, and the next batch served is the one the updated model rates highest. There is no separate seed set to argue about, which removed one of the main sources of TAR disputes.
TAR is not a substitute for culling; it is what happens to the set that culling leaves. It also sits alongside email threading and near-duplicate grouping, which reduce the number of documents a reviewer has to read to make the same decisions.
Defensibility rests on the validation, not the algorithm. Parties agree on a recall target, a sampling method for estimating it, and what is disclosed, and that agreement belongs in the ESI protocol.
What does technology-assisted review look like in practice?
The case for spending on review technology rather than review hours is arithmetic. RAND’s Where the Money Goes found review accounted for about 73 percent of eDiscovery production cost, against 19 percent for processing and 8 percent for collection. Grossman and Cormack’s 2011 study in the Richmond Journal of Law and Technology found technology-assisted review achieved recall and precision at least as good as exhaustive manual review while reviewing a small fraction of the documents.
Against that, the classic benchmark for the alternative: Blair and Maron’s 1985 study found that lawyers who believed their search terms were retrieving about 75 percent of the relevant documents were in fact retrieving about 20 percent.
How does technology-assisted review show up on an eDiscovery invoice?
TAR usually appears as an AI and analytics surcharge (charged per document, per gigabyte, or as a premium module) which is the line the Fee Field Guide flags as the defining irony of legacy pricing: paying a premium for the technology sold as the way to reduce cost. That is changing; several major platforms now include AI review in the base price at no charge, and published per-document pricing has been falling.
TAR also interacts with per-seat user licenses. A workflow that lets a small team review a ranked set costs less in seats than one that needs a large contract team, and the seat line is where that saving shows up.
Ask which analytics are included in the base price and which are extra, whether AI is charged per document, per GB or flat, and, bluntly, whether the other line items actually fall when review hours do.
Sources
- Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012).
- Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015).
- Maura R. Grossman & Gordon V. Cormack, “Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review,” 17 Rich. J.L. & Tech. 11 (2011).
- David C. Blair & M.E. Maron, “An Evaluation of Retrieval Effectiveness for a Full-Text Document-Retrieval System,” 28 Communications of the ACM 289 (1985).
- RAND Institute for Civil Justice, Where the Money Goes (Pace & Zakaras, 2012).
- The Sedona Conference, TAR Case Law Primer.
Related terms
Early case assessment · Culling · Near-duplicate · Email threading · Processing · Hosting
eDiscovery at one flat monthly rate
$3,250 a month on a 12-month agreement. No per-user fees and no per-GB hosting charges.
Common questions
Is TAR accepted by courts?
Yes. Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012), was the first written opinion approving its use. By Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015), the same court wrote that "the case law has developed to the point that it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it." Courts have been markedly less willing to force TAR on a party that does not want it.
What is the difference between TAR 1.0 and TAR 2.0?
TAR 1.0 trains a classifier on a control set and a seed set, validates it, then applies it once to the remaining documents: training and review are separate phases. TAR 2.0, usually called continuous active learning, retrains constantly as reviewers code, always feeding them the documents the model currently rates highest. CAL needs no stable seed set and copes better with collections that arrive in stages.
Does TAR replace attorney review?
No. Every training decision is an attorney coding decision, and the documents the model promotes are still read by people. What TAR changes is the order and the stopping point: instead of reviewing 100 percent of a set to find the responsive documents, a team reviews the top of a ranked list until recall estimates justify stopping. Privilege review in particular is normally still run over everything produced.