At $1.50 per document, a 250,000-document matter costs $375,000 in first-pass review alone — before a single privileged document is logged. Here's where that number comes from and what AI-augmented review actually replaces.
The per-document review fee is the labor cost of putting human eyes on each document to determine whether it is responsive to the discovery requests at issue. It is billed by document count, not by hour, because the staffing agencies and managed review vendors who run these projects have converted their underlying labor cost — contract attorney hours — into a per-unit metric that is easier to quote against a specific dataset size.
The market rates are well-established. Contract review through traditional staffing agencies sits at $1–$3 per document for first-pass responsiveness review and $4–$8 per document for privilege review. The rate variation reflects three things: the seniority of the reviewer (a junior contract attorney performing binary responsive/non-responsive calls versus a senior associate with privilege training performing a multi-factor analysis), the complexity of the responsiveness determination (a broad RFP covering all communications about a topic versus a narrow RFP keyed to specific contract terms), and the platform overhead built into the rate (the staffing agency's margin, the review platform license, and the project management layer all sit somewhere in the per-document number even if they are not broken out as separate line items).
On a 250,000-document matter — typical for a 100 GB commercial dispute after deduplication and near-duplicate suppression — that single line item at $1.50 per document runs to $375,000. At $3 per document, it is $750,000. Neither number includes privilege review. Neither number includes processing, hosting, production, or privilege log generation. The per-document review fee is the floor, not the ceiling.
Working through the numbers on a representative 100 GB commercial dispute illustrates how quickly the per-document fee dominates the invoice. After processing and deduplication, 100 GB typically yields roughly 250,000 reviewable documents. This is the universe that goes to first-pass responsiveness review.
First-pass review at $1.50 per document produces $375,000 in labor cost before the reviewer has finished their first week. Of those 250,000 documents, industry benchmarks suggest that roughly 6% — about 15,000 documents — will be tagged as responsive to the requests for production. Of those, approximately 2,000 will show privilege signals and require privilege review. Privilege review at $6 per document adds another $12,000. Total review labor: $387,000.
This single cost category — the work of reading and tagging documents — represents approximately 84% of a $460,000 all-in traditional vendor quote for the same matter. The other 16% covers processing ($75–$150 per GB), hosting ($25–$40 per GB per month over four to six months), privilege log generation ($10–$15 per entry), and production delivery ($0.05–$0.15 per page). The per-document fee is the dominant variable. It is also the category where the cost differential between traditional and AI-augmented platforms is most dramatic: on a DecoverAI matter, all of this is included in the $60 per gigabyte per month base rate.
| Review stage | Document count | Rate | Subtotal |
|---|---|---|---|
| First-pass responsiveness (contract review) | 250,000 | $1.50/doc | $375,000 |
| Privilege review (senior associate) | 2,000 | $6.00/doc | $12,000 |
| Total review labor | $387,000 | ||
| As % of $460K all-in quote | 84% |
The responsiveness pass that took a contract review team six weeks in 2015 can now be performed in an afternoon by a multi-model GenAI classifier trained on the matter's requests for production. The classifier reads each document, scores it for relevance to each RFP, flags it for privilege signals, and returns a tagged set with confidence scores — at a cost that is effectively zero per document at platform scale, because the marginal cost of running one more document through the model is fractions of a cent in API compute, not $1.50 in attorney labor.
What AI classification does not replace is equally important to understand clearly. It does not replace attorney judgment on close calls — the document that scores 0.52 on a 0–1 relevance scale needs a human to resolve the ambiguity. It does not replace privilege adjudication for documents that show attorney-client or work-product signals, which require an attorney to assess the communication in the context of the broader matter. And it does not replace the quality control layer that validates the AI's output: a QC sample review drawn from both the high-confidence responsive set and the high-confidence non-responsive set is a standard practice in any defensible TAR workflow, and that QC review is attorney time.
These are the legitimate remaining labor costs. They are not the 250,000-document first pass. The per-document fee that has historically consumed 84% of a matter's budget was always billing for the mechanical work of reading and binary-tagging at volume — work that AI does faster, cheaper, and with measurable recall and precision rates. The attorney cost that remains after AI handles the first pass is real, but it is scoped to a fraction of the document universe and it is doing genuinely discretionary work rather than volume processing.
Privilege review at $4–$8 per document is the second layer of per-document billing, and it compounds in a way that the first-pass fee does not. On a 2,000-document privilege set at $6 per document, the privilege review labor is $12,000 — manageable in isolation. The problem is what comes next.
Privilege log generation adds another $5–$15 per entry on top of the review fee. At $10 per entry on a 2,000-entry log, that is another $20,000. The total cost of identifying and logging 2,000 privileged documents — $12,000 in review plus $20,000 in log generation — is $32,000, for a set of documents that will never be produced and that opposing counsel will never see. The compounding of per-document review fees and per-entry log generation fees on privilege materials is the specific pattern that turns a $40,000 matter into a $120,000 invoice: the privilege set gets reviewed twice (once in first-pass to flag, once in privilege review to adjudicate), and then each entry gets drafted separately as a log entry.
AI-assisted privilege classification reduces the human review to QC and close-call adjudication of the flagged set. Auto-generated privilege log entries replace the $10–$15 per-entry drafting fee with a platform function that produces a defensible, formatted log as part of the standard workflow. The attorney still reviews the log before it is served — that step is not eliminated — but the per-entry drafting cost is. See Privilege Log Fees: Why You're Paying $10–$15 Per Entry for Something AI Drafts in Seconds for the detailed privilege log cost analysis.
The per-document fee persists for the same structural reason that hourly billing persists in legal services generally: it converts the vendor's labor cost into a client line item that scales with the volume of the matter. A vendor whose pricing model is built on per-document review rates has a structural incentive to not reduce the number of documents reviewed, because every document removed from the review universe is a document removed from the invoice. This is not a conspiracy; it is a consequence of how the pricing model is designed.
The incentive runs in the wrong direction for the client. A client's interest is to minimize the cost of production while maintaining a defensible review. The vendor's interest, under a per-document model, is to maximize the volume of documents that reach human review — which means investing less in upstream culling, deduplication, and AI pre-classification than the technology now makes possible. The client who asks only "what is your per-document rate?" is optimizing the wrong variable. The correct question is: what is your all-in cost for this matter if AI handles the first pass? See the all-in vendor pricing guide for the framework to evaluate vendor quotes on this basis.
The persistence of per-document billing also reflects the fact that it is a familiar and legible line item for clients who review invoices but do not have deep eDiscovery expertise. A per-document rate is easy to audit against a document count. An all-in per-GB rate requires the client to understand what "all-in" actually covers — which is a more sophisticated evaluation but ultimately a more useful one for controlling total spend.
92% of the cost differential between traditional and AI-augmented pricing on a 100 GB matter comes from this single line item. AI does not save money on the gigabytes — it saves money on the human hours that the per-document model exists to bill against.