Verdictbase Tech.
eDiscovery

Technology-Assisted Review Explained: When TAR Works, and When Courts Accept It

TAR can cut review cost sharply, or start a fight you didn't budget for. Here's how to tell which before you sign: TAR 1.0 vs continuous active learning, recall targets, and the court rulings that matter.

Verdictbase Tech editorial team · November 2026 · 10 min read

Four questions that decide whether TAR fits this matter

It usually arrives as a budget problem. The review vendor quotes a linear review of several hundred thousand documents, and someone suggests technology assisted review instead. The partner wants to know two things: will it cut cost, and will it start a fight over the production? Both are largely predictable before anyone signs a statement of work.

Technology assisted review, often called predictive coding, uses a machine-learning model trained on lawyers' relevance decisions to rank or classify the rest of a document set. Reviewers code documents, the model learns which features predict relevance, and the team either reviews the highest-ranked material first or stops once statistical testing shows it has found enough. Four questions decide whether it's worth it:

  • Volume after culling. Once deduplication, date filters and email threading are applied, a collection below roughly 30,000 to 50,000 documents often loses its savings to setup and validation overhead. Above about 100,000, TAR is almost always worth pricing. These are rules of thumb, not thresholds.
  • Content type. The models learn from text. Email, memos, contracts and chat exports work well. Spreadsheets with little text, images, engineering drawings, audio and poorly OCR'd scans don't, and they need a separate workflow.
  • Prevalence. If only 1% of the collection is relevant, the model needs more examples before it finds the relevant material. Older workflows struggle with that. Continuous active learning copes better.
  • Opponent and forum. Will you have to disclose the method, agree a protocol or defend it before a judge? The answer sets how much validation work to budget for.

TAR 1.0 vs TAR 2.0: seed sets versus continuous active learning

Most confusion in vendor demos comes from two generations of the technology being sold under one name.

Advertisement

TAR 1.0 front-loads the training. A senior lawyer who knows the case codes a seed set and several training rounds, often a few thousand documents in total. Progress is measured against a randomly drawn control set until the model stabilizes. The model then scores every document, the team picks a cutoff score, and documents above it are reviewed or produced while those below are set aside. Training ends before the main review starts.

TAR 2.0, or continuous active learning (CAL), keeps learning throughout. Every coding decision by every reviewer feeds back into the model, which re-ranks the queue so the likeliest relevant documents come up next. Review continues until the share of relevant documents in each batch drops sharply. The team then validates what was left unreviewed.

FactorTAR 1.0 (seed set / control set)TAR 2.0 (continuous active learning)
TrainingUp front, then frozenContinuous, from every reviewer decision
Who trainsOne or two senior subject-matter expertsThe whole review team, with QC
New data arriving mid-matterOften requires retraining and new control setsAdded to the ranking as it arrives
Low prevalence collectionsWeak; needs large training setsStronger; keeps hunting for rare documents
Main cost leverProducing above a cutoff with limited eyes-on reviewStopping early once relevant documents run out
Typical fitSingle, static collection; fast classificationMost litigation, investigations and rolling productions

For most matters, CAL is the better default. Collections rarely arrive all at once, CAL doesn't depend on a partner coding thousands of training documents, and the major review platforms generally offer it as a standard workflow. TAR 1.0 still makes sense where one static collection must be classified quickly and the producing party is comfortable producing above a cutoff after a separate privilege screen.

Recall, precision and the TAR validation numbers you'll have to defend

Recall is the share of all relevant documents in the collection that your process found. Precision is the share of documents you marked relevant that really are relevant. Recall protects the requesting party, because it measures what they might be missing. Precision protects your budget and keeps irrelevant confidential material out of the production.

Negotiated protocols mostly argue about recall. No fixed legal threshold is generally recognized. Recall estimates somewhere around 70% to 80% are often cited as reasonable and proportionate, and parties sometimes push higher in high-stakes matters. Linear manual review is a weak benchmark anyway: human reviewers regularly disagree with each other on the same documents.

The standard validation method is an elusion test. You draw a random sample from the documents you did not review or produce, have an experienced reviewer code it blind, and use the result to estimate how many relevant documents you left behind. A sample of roughly 385 documents gives a margin of error of about ±5% at 95% confidence. About 1,500 narrows that to roughly ±2.5%.

Worked example (hypothetical): Say a CAL review of a 400,000-document collection reviews 70,000 documents and finds 36,000 relevant. That leaves 330,000 unreviewed. An elusion sample of 1,500 turns up 30 relevant documents, or 2%. That implies about 6,600 relevant documents missed, so estimated recall is 36,000 ÷ 42,600, or about 85%. Precision in the reviewed set is about 51%. Report recall as a range reflecting the margin of error, not as a single number.

Settle these in the protocol before review starts:

  • The stopping criteria, and who decides when they've been met
  • Sample sizes, who codes the validation sample, and whether that coding is blind
  • What happens if estimated recall comes in below target (usually more review, not a restart)
  • How document families, attachments and non-text files are handled
  • How much the other side sees: metrics only, or samples of non-privileged documents coded not relevant

The validation protocol entered in In re Broiler Chicken Antitrust Litigation (N.D. Ill., 2018) is widely used as a reference for this level of detail.

How courts from New York to Melbourne have treated TAR

The leading decisions in common-law jurisdictions agree on more than they differ. Courts don't demand perfection. They want a process that is reasonable, proportionate, transparent enough to test, and validated.

CaseCourtWhat buyers should take from it
Da Silva Moore v Publicis Groupe (2012)S.D.N.Y., United StatesEarly, widely cited judicial approval of predictive coding in an appropriate case, with emphasis on a transparent, cooperative process
Rio Tinto v Vale (2015)S.D.N.Y., United StatesApproved an agreed protocol and treated TAR as an accepted method where parties agree to it
Hyles v New York City (2016)S.D.N.Y., United StatesA requesting party generally can't force the producing party to use TAR; the producing party chooses its method
Irish Bank Resolution Corporation v Quinn [2015] IEHC 175High Court, IrelandApproved TAR for a large discovery exercise, provided the process is transparent and validated
Pyrrho Investments v MWB Property [2016] EWHC 256 (Ch)High Court, England and WalesApproved TAR; cited volume, cost proportionality and the lack of evidence that manual review is more accurate
Brown v BCA Trading [2016] EWHC 1464 (Ch)High Court, England and WalesApproved TAR despite objection, mainly on cost and proportionality grounds
McConnell Dowell v Santam [2016] VSC 734Supreme Court of Victoria, AustraliaThe court directed TAR for a very large collection and expected the parties to agree a workable protocol

Court rules have since caught up. In England's Business and Property Courts, Practice Direction 57AD expects parties to consider technology for large disclosure exercises and to address their approach in the Disclosure Review Document. The Supreme Court of Victoria has a technology practice note that addresses TAR. In the United States, the proportionality test in Federal Rule 26(b)(1) and the Sedona Principles shape most disputes. Other common-law systems, including Canada, Singapore, Hong Kong and India, apply similar proportionality reasoning, though the volume of reported TAR rulings varies. Check local authority.

In civil-law systems across much of Europe, Latin America and the Middle East, broad documentary disclosure is rare. There, TAR mostly earns its keep in internal investigations, regulatory and antitrust requests, international arbitration document production, and large data-subject access requests. Everywhere, data-protection rules such as the GDPR, India's DPDP Act and Gulf-state data laws can restrict where documents are hosted and processed. Check the vendor's hosting regions before you choose a platform.

Where generative review fits alongside TAR

Generative review works differently. You don't train a model on coded examples. You write review instructions, essentially a careful restatement of the requests and issues, and a large language model reads each document and predicts relevance, usually with a written rationale. Several major platforms, including Relativity, Everlaw and DISCO, offer versions of this.

Its strengths are real. It needs no training set, it handles small matters where TAR isn't viable, and the rationales speed up QC. The weaknesses are real too. It typically costs more per document, small wording changes in the instructions can shift results, very long documents can exceed what the model reads in one pass, and rationales can sound confident while being wrong.

Precedent for generative review is still thin compared with Da Silva Moore or Pyrrho. The safest assumption is that courts will apply the same yardstick as for TAR: was the process reasonable and proportionate, and was recall validated by sampling? Use the same elusion tests and the same disclosure commitments.

In practice, the two combine like this:

  • Use CAL to rank a large collection, then run generative review on the top tiers for first-pass coding and issue tags.
  • Use generative review on collections under roughly 30,000 documents, where TAR setup costs don't pay back.
  • Use it for QC on TAR decisions and for drafting privilege-log descriptions, always with a lawyer signing off.

In a contested production, don't swap a validated CAL workflow for a generative one you haven't validated. Run both on a sample first and measure them against the same blind-coded set.

Cost drivers, red flags and the protocol decision

These are indicative 2026 market ranges. They vary widely by contract, volume and region, so confirm them with vendors:

  • Hosting: roughly $5 to $20 per GB per month. Many platforms include TAR and CAL in hosting; some charge a separate analytics fee.
  • Human review: roughly $35 to $80 an hour for contract reviewers in the US, UK and Western Europe, and often $10 to $30 an hour through managed review centers in markets such as India, the Philippines or South Africa. This remains the biggest line, and it's what TAR reduces.
  • Generative review: usually priced per document or per GB, with a wide spread between vendors and volume tiers. Get a written quote on your own data.
  • Protocol and validation expertise: consultants or vendor specialists commonly charge a few hundred dollars an hour. Expect more when protocols are contested.

Red flags in a vendor pitch or an opponent's proposal:

  • An accuracy claim like "99% accurate" with no recall estimate, sample size or margin of error
  • A stopping rule based on a fixed document count or a budget instead of a validation result
  • Validation samples coded by the same reviewers who did the review, without blind coding
  • No stated plan for spreadsheets, images, short messages or foreign-language documents
  • A protocol that gives the other side approval over every training decision, or hands over the whole training set without good reason

Here's the decision rule. If the post-cull collection is above about 100,000 documents, mostly text, and you can afford a blind elusion sample of around 1,500 documents, price a CAL workflow first and write the validation terms before review begins. Between roughly 30,000 and 100,000 documents, quote both CAL and generative review and validate whichever you pick. Below that, a well-run manual or generative review is usually cheaper than building a TAR protocol. In every case, ask litigation counsel in the relevant jurisdiction what the court or the other side is likely to expect.

Common questions

Do I need the other side's agreement to use technology assisted review?

Often not. In the US, courts have generally let the producing party choose its method, and English and Irish courts have approved TAR on proportionality grounds. Practice varies by forum, though, and England's disclosure rules expect the approach to be discussed. Disclosing the method and agreeing validation terms early reduces the risk of a later challenge. Ask your litigation counsel what the forum expects.

What recall rate will a court accept?

No jurisdiction sets a fixed number. Estimates in roughly the 70% to 80% range are often cited as reasonable, and stakes and proportionality can push the target up or down. Courts and opponents look hardest at whether the estimate comes from a properly designed random sample.

Is TAR 2.0 always better than TAR 1.0?

For most litigation, it's the better default. Continuous active learning copes with rolling collections and low prevalence and doesn't need a senior lawyer to train it. TAR 1.0 can still suit a single static collection that has to be classified quickly above a cutoff score.

Does TAR work on non-English documents?

Yes, if the documents contain extractable text, but you need reviewers fluent in each language doing the training and validation coding. Mixed-language collections often work better when separated by language. Check how the vendor's platform handles languages other than English before you sign.

Has any court approved generative review the way courts approved TAR?

Precedent is still thin compared with Da Silva Moore or Pyrrho. Expect courts to apply the same tests they use for TAR: reasonableness, proportionality and validated recall. Validate generative review with blind elusion sampling, just as you would TAR.

How much does TAR validation cost?

Mainly reviewer time to code the validation samples. An elusion sample of about 1,500 documents is typically a few days of senior reviewer work, depending on document length. Add consultant time if the protocol is negotiated or contested. Get quotes from vendors, because rates vary widely by region and contract.

Written by the Verdictbase Tech editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.