Per-document AI review pricing is falling fast. The Winter 2026 eDiscovery Pricing Survey places GenAl-assisted review in the $0.11-$0.50 range. Major review platforms are folding AI into standard workflows.
That is good news.
But it does not answer the buyer's question: why is my matter still so expensive?
The answer is usually not the review unit price. It is the population entering the pipeline.
If collection sends 100,000 document versions downstream when only a small fraction corresponds to actual custodian interactions, every downstream stage gets taxed: hosting, processing, culling, review, QC, privilege, and verification. This is the matter-level math your team is not being shown.
RAND's 2012 study - still the benchmark - placed document review at 73% of total production cost in large matters. ComplexDiscovery's 2024 market analysis put review at 64% of a $16.9 billion global market, or roughly $10.8 billion. Collection's share grew from 8% to 16% and is projected to double to $6.03 billion by 2029, the fastest-growing segment in the industry.
Review is compressing, collection is expanding, and neither trend tells you anything useful about what a specific matter will cost, because unit cost is a denominator metric. What determines a matter's cost is the numerator - the number of documents in your pipeline - and the numerator is set at collection, not at review.
A $0.30 unit cost on 1,000,000 collected documents is $300,000. A $0.30 unit cost on 50,000 collected documents is $15,000. Same unit price. Twentyfold cost swing. Nothing about the review tool changed.
That math is what we analyzed in detail on the RGR blog, the upstream-fidelity argument from the standard's perspective. This piece is the matter-level math from the buyer's side.
Consider a common fact pattern: 12 custodians in scope, collaborating through a shared SharePoint library containing roughly 5,000 documents, each with an average of 20 versions accumulated through normal co-authoring activity.
Traditional "collect the source" methodology: 5,000 × 20 = 100,000 document versions. Every one of them is hosted, processed, culled, and - for whatever survives - reviewed. At typical per-GB hosting and processing fees and AI-assisted review at $0.30/document post-culling, a modest matter configured this way easily clears six figures in platform costs before attorney time is factored. Deduplication helps, but catches exact copies only - near duplicate versions of the same document each carry a different hash, and each one survives dedup.
Context-aware collection: targets only the document versions the custodians actually interacted with. If those same twelve custodians substantively interacted with twelve files from the library, the collected population is twelve documents at their correct versions, plus documented exceptions where the correct version is no longer retrievable. Hosting, processing, culling, and review costs all drop because the downstream population is smaller, often by orders of magnitude.
In that simplified example, the savings are in never sending 99,988 document versions through the pipeline in the first place. That math is the reason we built Cloudficient's Context-Aware eDiscovery: a collection approach designed to target custodian interactions, version correspondence, and exception reporting before data enters review.
Your processing vendor will tell you about deduplication, email threading, and keyword culling - and those tools are real. Deduplication can cut volumes 30–40%, sometimes up to 90%. In a Perkins Coie case study, email threading reduced review costs by up to 65% where threads were intact.
But these tools were built to filter well-formed collections. They cannot fix what collection got wrong. Near-duplicate versions of the same SharePoint document survive dedup. Keyword filters hit any version containing the term, regardless of whether that version corresponds to a custodian's actual interaction. Threading can't reconstruct a conversation that was split across compliance records at the source.
When collection produces version sprawl, culling reduces volume but not noise. The reviewer, human or AI, still has to decide which version of each document mattered, or whether any of them did. That decision time shows up somewhere. Usually on the bill.
We learned this early. Downstream filtering tools assume the collection entering them is structurally sound. When it isn't, no amount of dedup, threading, or keyword culling recovers the fidelity the collection workflow didn't preserve. That is why we moved the investment upstream, to the collection layer, where the decisions that actually drive downstream volume get made.
A recent ComplexDiscovery analysis cited a METR study finding that knowledge workers using AI completed tasks 19% slower than without it, while believing they were faster, with 37% of supposedly saved time consumed by reviewing, correcting, and verifying AI-generated output. In legal review, that verification tax is not theoretical: privilege calls made by AI on ambiguous documents, classification on documents whose version doesn't match the communication they came from, and disposition decisions on populations whose completeness was never validated.
The $0.30/document AI review price does not include the cost of verifying whether the AI's call was right, and in most matters, it is not even being measured. An AI classifier handed a version-sprawled collection produces confident outputs on shaky inputs. The verification overhead then lands on attorneys, litigation support, or outside counsel, not on the review vendor's Al-price line item.
Cloudficient's approach is to reduce that tax upstream by making the collection cleaner and by flagging the documents where fidelity is known to be imperfect.
Cloudficient's Context-Aware eDiscovery is an implementation aligned to the Reconstruction-Grade eDiscovery Standard. RGR is vendor-neutral and open; any vendor can build against it, declare against it, or be evaluated against the same criteria.
Cloudficient had already been investing in this architecture before RGR was published. The value of the standard is that it turns those architectural choices into measurable criteria that buyers can ask any vendor to declare against, including us.
In practice, that translated into four capabilities we designed the workflow around, each mapped directly to the cost drivers in sections 2 through 4:
Per-custodian scoping of shared sources. Rather than collecting an entire SharePoint library because twelve custodians touched some of it, we collect only the documents those custodians actually interacted with - at the versions that correspond to their interactions.
Version correspondence, not version sprawl. Where retention configuration and source history permit, we target the version of a hyperlinked document that existed at the time of the custodian's interaction, not the current version, and not every version that ever existed.
The important point for buyers is that these claims should be testable. RGR defines public conformance criteria and measurable tiers. Any vendor, including Cloudficient, can be evaluated against the same rubric.
If you are a general counsel, head of litigation support, or on the procurement side of an eDiscovery evaluation, here are five questions that will separate vendors with measurable collection-fidelity capabilities from vendors still relying on broad collection and downstream cleanup:
Per-custodian scoping. Can you collect only the documents a specific custodian interacted with from a shared library, or only the full library?
Version targeting. Does your workflow collect the version of a hyperlinked document that existed at the custodian's interaction, the current version, or all versions?
Exception reporting. When the correct version is no longer retrievable, do you produce a documented exception or silently substitute the current version?
Auditability. Can you export the relationship between parent communication, linked file, targeted version, collected version, and exception reason in a defensible report?
RGR conformance. At which tier of the Reconstruction-Grade eDiscovery Standard does your collection methodology conform? If you do not claim conformance, can you map your workflow to the RG-Aware disclosure catalogue and explain the gaps?
The fifth question is the shortest way to ask the first four. RGR is the vendor-neutral, open standard; any vendor can build against it. Requiring tier declaration or, where that is not yet honestly available, a structured gap explanation, is how a buyer makes collection quality measurable without endorsing any specific product.
The review-cost compression story is real. Per-document AI pricing will keep falling. But matter-level costs are not determined by the price of the fastest step in the pipeline; they are determined by how much of the pipeline should not have been running at all.
The most impactful cost reduction available in eDiscovery today is not a cheaper review platform. It is a collection methodology that declines to send the wrong documents to review in the first place, and documents the gaps where it couldn't capture the right ones.
That discipline is what makes AI review economics work as advertised. Without it, you are paying unit-cost-compressed prices on a population that grew at every stage upstream.
If your current workflow is pushing entire shared sources, all versions, or current-version cloud attachments into review, the math above is not theoretical. It is probably already on your invoice.
Cloudficient helps buyers move the cost decision upstream: collect the right document, at the right version, for the right custodian, and document the exception when the right version is gone.