Counsel Ops HQLitigation Matter & Spend Tech
eDiscovery

Document Review Cost Compared: Linear, TAR and Generative AI on 200,000 Documents

We costed one hypothetical 200,000-document review three ways. Generative AI came out cheapest, but only about 12% below TAR, and its per-document fee decides whether it wins at all.

By the Counsel Ops HQ editorial team · November 2026 · 9 min read

Three Quotes, One Matter: What 200,000 Documents Really Cost

Say you're holding three proposals for the same production. One is a linear review with contract attorneys. One uses technology assisted review (TAR) with continuous active learning. One uses a generative AI tool that codes every document and sends the uncertain ones to people. The totals look far apart, but the line items underneath are hard to compare. You need to know which number is real before you sign a statement of work.

This piece builds one hypothetical 200,000-document review and costs it all three ways, line by line. No real client or vendor quote sits behind it. The inputs are typical 2026 market ranges, and they vary by contract, region and data type. The short version: totals fall from about $365,000 for linear review to about $194,000 for TAR and about $171,000 for generative review. Most of the saving comes from first-pass reviewer hours. Privilege review, redaction and validation barely move. To forecast document review cost honestly, model those fixed human layers on their own lines.

Advertisement

Assumptions Behind the Hypothetical Review

Every number below comes from these assumptions. Change one and the totals shift, so swap in your own figures before relying on the comparison.

  • Review set: 200,000 documents after processing, de-duplication and date filtering, about 60 GB hosted, mostly email and office files.
  • Richness: 20% responsive, so about 40,000 documents.
  • Potentially privileged: 6,000 documents flagged by screens, with about 2,000 ending up on the privilege log.
  • Redactions: 4,000 documents need redaction for personal data or confidentiality.
  • Reviewer pace: 50 documents per hour for first-pass review, 30 for privilege review, 20 for redaction.
  • Blended rates: first-level reviewer $50/hour, senior or QC reviewer $75/hour, associate for privilege calls $275/hour.
  • Platform: hosting at $12/GB/month and review seats at $75/user/month.
  • Generative AI fee: $0.25 per document, plus a small budget for prompt-testing re-runs.
  • Recall target: roughly 75–80% for both TAR and generative review, validated by sampling.

The Line-by-Line Cost Comparison

Line itemLinear reviewTAR (CAL)Generative AI review
Hosting (60 GB at $12/GB/month)$2,880 (4 months)$2,160 (3 months)$1,440 (2 months)
Review seats ($75/user/month, 2 months)$3,750 (25 users)$1,500 (10 users)$900 (6 users)
Analytics or generative AI fees$0$3,000$52,500 ($50,000 + $2,500 test runs)
Setup: protocol, training, prompt development$8,500$9,000$14,250
First-pass human review$200,000 (4,000 h)$60,000 (1,200 h)$12,000 (240 h)
QC and validation sampling$24,750 (330 h)$9,750 (130 h)$6,750 (90 h)
Privilege review$55,000 (200 h)$55,000 (200 h)$41,250 (150 h)
Privilege log$15,000 (200 h)$15,000 (200 h)$6,000 (80 h)
Redaction$10,000 (200 h)$10,000 (200 h)$10,000 (200 h)
Project management and supervision$45,000$29,000$26,000
Total$364,880$194,410$171,090
Cost per document$1.82$0.97$0.86
Elapsed time (approx.)5–6 weeks3–4 weeks1.5–2.5 weeks

Two things stand out. Linear review costs nearly twice as much as either alternative. And generative review beats TAR by only about $23,000, because the per-document fee absorbs most of the labor it saves.

Where the Linear Budget Goes, and Contract Reviewer Rates by Region

In the linear scenario, first-pass review is about 55% of the bill. At 50 documents an hour, 200,000 documents means 4,000 reviewer hours, or roughly 25 people full time for four weeks. Every dollar of hourly rate moves the total by $4,000, so reviewer rates matter more here than any platform fee.

Rates differ widely by market. These are rough, indicative 2026 ranges that vary by vendor and by the language skills a matter needs:

  • US and Canada: roughly $35–$75/hour for first-level contract attorneys, more for foreign-language or jurisdiction-licensed reviewers.
  • UK, Ireland and Western Europe: roughly $40–$80/hour equivalent through managed review providers.
  • India, the Philippines and South Africa (managed offshore review): often $15–$35/hour.
  • Gulf states and Singapore: often $45–$90/hour, because the pool of qualified local-language reviewers is smaller.

Offshore rates can cut first-pass cost sharply. Cross-border data rules often take that option away, though. The GDPR's transfer restrictions, India's Digital Personal Data Protection Act, and data-localization rules in parts of the Middle East and Southeast Asia can limit where data is reviewed. Have counsel confirm where review can lawfully happen before you price an offshore team.

Technology Assisted Review Cost: Cheap Software, Real Validation Work

Software is rarely the expensive part of a TAR budget. Many major platforms, including Relativity, Everlaw, Reveal and DISCO, offer continuous active learning, and some include it in standard licensing while others add an analytics fee. Ask which applies. We used $3,000. The saving comes from reviewing only about 30% of the set (60,000 documents) before the model stabilizes and the low-ranked remainder can be set aside.

Don't skimp on validation. The $9,750 QC line covers a 10% QC sample of reviewed documents and an elusion test of about 1,500 documents drawn from the discard pile. If opposing counsel or a court asks how you know recall reached 75–80%, that sample is your answer. Courts in the US, England and Wales, and Ireland have accepted TAR when the process was transparent and validated. Many protocols now fix sample sizes in advance. Agree on them early, because re-sampling after a dispute costs more than doing it right once.

TAR's weak spot is low richness. If only 2–3% of the set is responsive, training takes longer and the share you must review can climb. Ask each vendor what review percentage they expect at your richness level.

Generative AI Document Review Pricing: The Fee Decides Who Wins

Generative tools, such as Relativity aiR for Review and comparable products from Everlaw, DISCO and others, code every document against a written prompt. In our example that cuts first-pass human review to 240 hours, spent on borderline and low-confidence calls. The per-document fee then becomes the largest single line. Pricing models include per document, per GB and consumption credits. We assumed a range of about $0.10 to $0.60 per document, which varies with volume commitments and document length. Confirm current pricing with each vendor.

Generative fee per documentFee line (incl. $2,500 test runs)Scenario totalVersus TAR ($194,410)
$0.10$22,500$143,590About $51,000 cheaper
$0.25$52,500$171,090About $23,000 cheaper
$0.40$82,500$201,590About $7,000 more expensive
$0.60$122,500$241,590About $47,000 more expensive

In this model, break-even against TAR is about $0.37 per document. Above that, generative review mostly buys speed. Many teams lower the fee by culling with TAR or search terms first, then running generative review on a smaller set. Setup also takes more hours: a 2,000-document human-coded control set plus about 30 attorney hours of prompt iteration. Budget for it. A weak prompt can cost you more in recall than you save on the fee.

The Human Hours That Don't Go Away

Look at hours instead of dollars. Counting only the lines above (excluding project management), human time drops from about 4,930 hours for linear review to about 1,930 for TAR and about 760 for generative. Senior legal time barely moves. Associate privilege hours run 200, 200 and 150. Privilege calls, protocol decisions and validation sign-off still need qualified lawyers. That holds whether the jurisdiction frames the issue as attorney-client privilege, legal advice privilege or professional secrecy under civil-law rules.

Redaction stays flat at 200 hours in every scenario, because someone still has to mark and check each redaction. Generative tools help with privilege mainly by cutting false-positive flags and drafting log descriptions. Our model trims privilege review by 25% and log time by 60%. Treat those as hypothetical inputs, not benchmarks, and test them on a sample of your own data.

Document review cost per document falls a lot. The cost of judgment doesn't. Privilege review and the log make up roughly a fifth of the linear total but more than a third of the TAR total. If your matters are privilege-heavy, such as internal investigations or regulatory responses, expect a smaller saving than this example shows.

Red Flags in Vendor Quotes

  • A per-document price with no definition of "document." Ask whether family members, attachments, embedded images and re-runs each count as billable units.
  • No line for validation. A quote without elusion testing or precision and recall sampling is pricing a review you may not be able to defend.
  • Privilege review buried in "attorney time as needed." It ran from roughly a fifth to over a third of the totals above. Get an estimate up front.
  • "Unlimited" prompt iterations that are billed per run. Get the re-run price in writing.
  • Unclear hosting terms after the review. At $12/GB/month, 60 GB left on the platform for 18 months of litigation costs nearly $13,000.
  • No answer on where the model runs. Data residency and whether customer data trains models are contract terms, not marketing claims. Confirm both, especially under the GDPR or other transfer rules.

A Decision Rule for Your Next Matter

Use linear review only for small sets (under roughly 20,000 documents) or when a protocol requires eyes on every document. For most matters between 50,000 and 500,000 documents, TAR is the safer default: the software is often bundled or cheap, courts in several jurisdictions have accepted it, and validation methods are well established. Choose generative review when the per-document fee lands below your own break-even (about $0.37 here) or when the deadline matters more than the last $20,000.

Before you sign, put your own volume, richness, privilege rate and local reviewer rates into the ten lines above, then ask each vendor to price all ten. A vendor that won't is showing you where the overruns will appear. Have litigation counsel review the validation protocol and any data-transfer terms.

Questions buyers ask

What is a realistic document review cost per document in 2026?

In our hypothetical, it ranged from about $1.82 per document for linear review to $0.97 for TAR and $0.86 for generative AI review. Real figures depend on reviewer rates, richness and privilege volume. Offshore reviewers can lower the first-pass line sharply, while privilege-heavy matters can run much higher. Treat these as typical 2026 market ranges and confirm pricing with vendors.

Is generative AI review always cheaper than TAR?

No. In our model, generative review beats TAR only while the fee stays below about $0.37 per document. Above that it costs more and mainly buys speed. Running generative review on a set already culled by TAR or search terms often improves the economics.

Does technology assisted review cost extra on major platforms?

Sometimes little, sometimes a modest analytics fee. Some platforms include continuous active learning in standard licensing and others charge per GB or per matter, so ask. Most of the cost is human: the documents reviewed before the model stabilizes, plus validation sampling.

Can generative AI handle privilege review on its own?

It can flag likely privileged documents and draft log descriptions, which reduces hours. Final privilege calls still need qualified lawyers who know the privilege rules of the jurisdiction concerned. Have counsel decide how far model output can be relied on for a given matter.

How much should we budget for QC and validation?

In our example, QC and validation ran from about $6,750 to $24,750, roughly 4–7% of each total. For TAR and generative review, include an elusion sample from the discard set and a precision and recall check on the coded set. Agree on sample sizes in the review protocol before work starts.

Do data-protection rules affect which review method we can use?

Yes. Rules such as the GDPR, India's DPDP Act and localization requirements in parts of the Middle East and Southeast Asia can restrict offshore reviewers and where a generative model processes data. Confirm hosting and processing locations in the contract, and get privacy counsel's sign-off before choosing an approach.

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