The Short Answer: Where AI eDiscovery Pricing Stands in September 2026
The most reliable current benchmark for AI eDiscovery cost per document comes from the EDRM Winter 2026 eDiscovery Pricing Survey, which was analyzed in detail by JD Supra in early 2026. Based on that survey and corroborating market data, the defensible benchmark range for AI-assisted document review in 2026 is roughly $0.50 to $3.00 per document for first-pass AI review, compared with $1.50 to $6.00+ per document for traditional human-only review of comparable material. Blended rates — where AI performs the first pass and humans review the top-ranked documents — commonly land between $0.75 and $2.50 per document all-in. For large, complex matters with privileged or highly technical content, per-document costs can climb to $4.00–$8.00 when you factor in quality control, privilege logging, and production work.
Also worth reading: How Should Legal Teams Benchmark AI for eDiscovery in 2026? · How Are AI Legal Document Platforms Transforming eDiscovery, Legal Research, and Drafting? · What Metrics Should You Use to Validate AI-Assisted eDiscovery in 2026?
These numbers represent a genuine reset rather than a gradual decline. Industry commentary throughout 2025 and 2026 — including the widely circulated 'Great eDiscovery Price Reset' analysis from Platinum IDS — describes a market where free or near-free AI review tools, all-inclusive subscription pricing, and cloud competition have compressed per-document rates by 30% to 60% versus 2023 levels. If a vendor is still quoting you legacy rates of $2,000–$3,000 per gigabyte plus $2.00+ per document for review, you are almost certainly overpaying in the current market.
Why Per-Document Pricing Is the Wrong Starting Point (But Still the Right Benchmark)
The eDiscovery industry has historically priced review in three ways: per document, per gigabyte, and per hour of reviewer time. Per-document pricing remains the most comparable benchmark because it directly maps to the unit of work — but it is also the easiest metric for vendors to game. A 'document' can be a one-line email or a 400-page spreadsheet printout, so two quotes at '$1.00 per document' can differ by a factor of ten in actual effort.
The reason per-document benchmarks still matter is negotiating leverage. When the EDRM survey publishes aggregate per-document rates, it gives buyers a reference point that vendors cannot easily dismiss. In 2026, the survey's data shows that AI-first providers quote median first-pass review rates well below $1.00 per document, while traditional managed review providers quote $2.00–$5.00 per document for human review — with AI used mainly as a prioritization layer rather than a cost-reduction engine. The gap between those two models is the entire story of the 2026 market.
A second reason the benchmark matters: it exposes where AI is genuinely saving money versus where it is being sold as a premium add-on. If a provider charges you an 'AI review surcharge' on top of human review rates, they are pricing AI as an enhancement. The market leaders in 2026 price AI as the default and charge humans only for the documents AI flags as responsive, privileged, or uncertain — which is typically 10% to 30% of the corpus.
How AI Review Actually Changes the Cost Structure
Understanding the benchmark requires understanding what generative AI changed about review economics. Traditional technology-assisted review (TAR) — predictive coding, as it was known for a decade — required training rounds, subject-matter experts, and statistical validation. It reduced review volume but still billed humans for every document in the review set. Generative AI review is different: a large language model reads and classifies every document in a single pass, at marginal cost approaching zero, and humans review only what the model flags or what sampling reveals as uncertain.
The Law.com coverage of a major 2026 study comparing Gen AI to TAR in complex document review ('Better Than TAR. Nearly Expert') found that modern LLM-based review matched or exceeded TAR performance on recall and precision in complex matters, while cutting the human review population dramatically. That finding is what justified the price reset: if AI can do the first pass at near-zero marginal cost, the per-document price of the whole workflow should fall toward the cost of the human QC layer plus platform fees.
The practical cost structure in 2026 looks like this. AI first-pass review: $0.10–$0.50 per document (or bundled into a platform subscription). Human review of AI-flagged documents: $1.50–$4.00 per document reviewed, but applied to only a fraction of the corpus. QC sampling and validation: 2%–10% of the corpus, often billed hourly. Processing, hosting, and production: increasingly flat-fee or all-inclusive. The blended per-document figure across the entire corpus is what produces the $0.75–$2.50 benchmark range cited above.
Benchmark Comparison: Traditional Review vs. TAR vs. Gen AI Review
| Feature | Traditional Human Review | TAR / Predictive Coding | Gen AI First-Pass Review (2026) |
|---|---|---|---|
| Per-document cost (full corpus) | $2.00–$6.00 | $1.50–$4.00 | $0.50–$3.00 blended; $0.10–$0.50 for AI pass alone |
| Human review volume | 100% of corpus | 20%–50% after training | 10%–30% of corpus |
| Time to first production | 4–12 weeks | 3–8 weeks | 1–4 weeks |
| Recall performance | Baseline (human-dependent) | 75%–90% typical | 85%–95%+ per 2026 studies |
| Training/validation cost | Minimal | High (SME time, rounds) | Moderate (prompt tuning, sampling) |
| Pricing model | Per doc or per hour | Per GB + per doc | Per doc, subscription, or all-inclusive |
| Defensibility burden | Manual QC documentation | Statistical validation protocol | Sampling + audit trail of model outputs |
The All-Inclusive Pricing Wave and What It Means for Benchmarks
The most disruptive pricing development of 2025–2026 was the arrival of all-inclusive and subscription-based eDiscovery pricing. Several providers now offer flat monthly fees that bundle processing, hosting, AI review, and production, with per-document charges either eliminated or reduced to overage tiers. The Platinum IDS analysis of this trend notes that cloud competition and the availability of free or low-cost AI review tools have forced traditional vendors to abandon per-gigabyte processing fees that once ran $100–$500 per GB.
For benchmarking purposes, all-inclusive pricing complicates the math. If you pay $5,000 per month for a platform that handles a 500,000-document matter over three months, your effective per-document cost is roughly $0.03 for platform access — but you still pay for the human review layer on top. Buyers should therefore calculate two numbers for every quote: the effective per-document cost across the full corpus, and the per-document cost of the human-reviewed subset. Vendors who resist being quoted in these terms are telling you something about their pricing transparency.
A caution is warranted here. All-inclusive pricing can hide overage triggers — per-GB processing on new data, export fees, charge-per-custodian collection costs — that reintroduce the unpredictability the flat fee was supposed to eliminate. Read the overage schedule before signing, and model your worst-case data volume against it.
Practical Steps: How to Benchmark Your Own Matter
Start by normalizing your corpus. Convert everything to a per-document and per-gigabyte figure, and ask vendors to quote both. A 100 GB matter with 400,000 documents has a very different cost profile than 100 GB with 50,000 documents, and vendors who refuse to disclose document counts are usually padding processing revenue.
Second, split every quote into its layers: collection, processing, hosting, AI first pass, human review, QC, privilege log, production. The EDRM survey data shows that vendors bundle these layers differently, and the only way to compare two quotes honestly is layer by layer. Ask specifically what the AI pass costs per document and what percentage of the corpus the vendor expects to route to human review — if they claim 100% human review 'for safety,' they are selling you 2020 economics with 2026 branding.
Third, demand a validation protocol as part of the price. Sampling rates of 2%–5% of the corpus, with stratified sampling of documents the AI classified as non-responsive, are becoming the accepted standard for defensibility. Budget $5,000–$25,000 for validation on a mid-size matter; it is cheap insurance against a motion to compel or an adversarial challenge to your review methodology.
Fourth, run a pilot. Most AI-first providers in 2026 will process a 5,000–10,000 document sample at low or no cost. Compare the AI's classifications against your own review of the same sample. If recall on the sample is below 80%, either the prompts need work or the vendor does. This single step eliminates more bad vendor choices than any RFP questionnaire.
Common Mistakes Buyers Make With AI eDiscovery Pricing
The most expensive mistake is comparing per-document quotes without normalizing document definitions. Insist that every vendor define a document (family-level vs. individual, attachments counted separately or together) and re-quote on identical definitions. Differences of 2x–3x in quoted document counts are common and entirely legitimate — but only if you know about them.
The second mistake is treating AI review as fully autonomous. The Law.com 'Ground Truth: The Realities of Generative AI in E-Discovery' reporting documents real-world failure modes: hallucinated privilege calls, inconsistent classification of near-duplicate documents, and models that perform well on common document types but degrade on handwritten notes, spreadsheets, and non-English material. Human QC on the AI's non-responsive determinations is not optional; it is the difference between a defensible review and a malpractice exposure.
The third mistake is ignoring privilege log costs. AI can draft privilege logs at a fraction of the traditional $10–$30 per entry cost, but many vendors still bill privilege logging as a separate line item at legacy rates. If your vendor's AI can classify privilege, it can populate a log — ask why that work is not included.
The fourth mistake is signing multi-year enterprise agreements based on 2025 rates. The market is still moving downward; the Winter 2026 EDRM survey shows rates below the Summer 2025 survey across nearly every category. Lock in volume discounts, not fixed rates, and include benchmark-based repricing clauses.
When to Act and When to Wait
If you have an active matter with more than 50,000 documents, act now. The cost difference between a legacy per-document contract and a 2026 AI-first contract on a 500,000-document matter is routinely $200,000–$800,000, and the market has already repriced — waiting does not improve your position, it just extends your exposure to old rates.
If you are building a standing capability rather than responding to a matter, a short evaluation window is reasonable. The AI review tools are improving quarter by quarter, and the 2026 studies suggest the gap between the best and median tools is still wide. Re-benchmark your panel of vendors every six months; the EDRM publishes pricing surveys twice yearly, and the Winter 2026 edition is the current reference point as of September 2026.
Small matters — under 10,000 documents — deserve a different calculus. At that scale, a solo practitioner or small firm using an AI review tool directly, with attorney review of the flagged set, can run total review costs under $5,000. Sending that matter to a full-service vendor at legacy rates is almost always the wrong economic decision.
What the Next 12 Months Will Do to These Benchmarks
Expect continued downward pressure. The partnership announcements of 2025–2026 — LexisNexis embedding Protégé inside Luminance's contract platform, and similar integrations across the legal tech consolidation wave tracked by LawSites and The Fashion Law's funding coverage — signal that AI review capability is becoming a commodity feature rather than a premium product. When capability is commoditized, price is the only differentiator, and price falls.
The counter-pressure is defensibility. As courts begin to scrutinize AI review methodologies, vendors will add validation, audit, and explainability layers — and some will try to charge for them. Buyers should push back: validation is a cost of doing business that should be baked into the per-document benchmark, not an upsell. The realistic floor for blended AI-assisted review is probably $0.50 per document for routine matters; below that, someone is cutting the QC layer, and you will pay for that omission later, in a deposition or a sanctions motion.