AI Is Rewriting eDiscovery Economics — But the Invoice Hasn't Caught Up

The demonstration took eleven minutes. On a screen in a vendor's conference room, Relativity's aiR for Review ingested a privilege-review set, reasoned through tens of thousands of documents, and produced its work — the kind of job that once occupied a team of contract attorneys for weeks. The audience did the same mental math: the review labor had just fallen by three-quarters or more.

Then someone asked what the software charged for the trick. The room went quieter than it had been for the demo.

This is the central tension in eDiscovery right now. Artificial intelligence has collapsed the actual work of review — the task that the RAND Institute found consumes 73 percent of eDiscovery cost — while the invoices still reflect the economics of an era when humans read every page. The technology moved. The pricing didn't.

The collapse is real

Start with what the machines can do. In a HaystackID case study, a hybrid workflow pairing Relativity's aiR with classic technology-assisted review tackled 891,527 documents — a corpus only 4 percent rich, the needle-in-haystack set that makes review partners wince. TAR ranking shrank the volume sent for generative-AI processing by more than 90 percent, from 748,612 eligible documents to just 54,380 processed. Final precision: 90.00 percent. Recall: 95.58 percent. A defensible, statistically validated production of 27,976 documents.

The Redgrave study of aiR on a 45,004-document set: roughly 18 hours of prompt work and one hour of processing, against 1,123 hours a managed-review team would have spent in active learning. One practitioner writing on Mondaq reported 75 percent cost savings using aiR for privilege review versus a full linear review. Forrester's Total Economic Impact analysis of aiR found review throughput up 30 percent — users expect 40 to 50 percent gains — and per-document review time falling from 30 seconds to 20.

Step back and the arc is unmistakable. ComplexDiscovery's 18-year tracking shows review's share of eDiscovery spend falling from 73 percent in the RAND 2012 baseline to a projected 52 percent by 2030 — while global data volume explodes from roughly 181 zettabytes in 2025 to 812 zettabytes in 2030. The market reconciles at $28.08 billion by 2030 precisely because AI is expected to compress per-document cost faster than data growth inflates it: a 3.13-times productivity-per-dollar mandate the machines will meet whether the invoices admit it or not.

The price of the trick

Here is where the story gets uncomfortable. Generative AI charges by the document, and the vendors are in no hurry to say what that costs.

Relativity bills aiR consumption in "aiR Units" — one unit per document analyzed — but when researchers at the University of Waterloo and Elevate asked for the allowance included in a subscription and the overage rate, the company would not say. The only independent frame: the Winter 2026 ComplexDiscovery/EDRM Pricing Survey, where 19 of 53 respondents placed per-document generative-AI review pricing between $0.11 and $0.50.

Compare that to the old machine: traditional technology-assisted review — TAR 1 and TAR 2 — costs, in industry-standard terms, nothing per document. As Jeremy Pickens wrote in a widely cited white paper covered by eDiscovery Today last August, a supervised TAR approach on a low-richness matter could save almost 70 percent over linear review — while the headline generative-AI case in the same analysis saved only 50 percent. Pickens' point: the vendors' favorite comparison is generative AI versus full human review, "a dreadfully weak baseline that is trivial to beat." Against the actual best alternative — the free, mature, court-blessed TAR the industry has had since Judge Andrew Peck approved it in Da Silva Moore — generative AI has to earn its per-document fee every time, and sometimes it doesn't.

So the pricing model is shifting from one meter to another. The old meter ran on gigabytes: vendors billed for processing, hosting, and review whether the data mattered or not. The new meter runs on documents touched by the AI — and at $0.11 to $0.50 a pop on a million-document matter, the invoice can land between $110,000 and $500,000 before a single human reads anything. Neil Cameron of Legal IT Insider has been pressing this point about the aiR studies: faster reasoning is not the same as cheaper review, and nobody can check the arithmetic because the vendors won't publish the rates.

The vendors see it too

Give the industry credit where it's due: at least one major vendor has read the writing on the wall. DISCO — whose CEO Eric Friedrichsen reported first-quarter 2026 revenue of $41.9 million, up 14 percent — announced a new commercial model bundling its eDiscovery product, case-building tools, and AI platform into a single per-gigabyte rate. The logic is transparent: if AI collapses the cost of the work, stop billing the work. Bill the data.

Everlaw has been pushing the same direction with its AI Assistant, betting that the platform that makes review feel close to free keeps the subscription. Emergen Research's sizing notes that services — the old metered middle — still held 61.4 percent of the market in 2025, which means most of the industry's revenue still depends on billing the very labor AI is eliminating.

And there is the counterargument the vendors make in private, and it deserves air. AI review is not actually free at scale — the compute is real, the QC time is real, and a badly prompted model can cost more than the reviewers it replaced. A GC who demands the 75-percent savings on every matter will find some where the set is small or the richness high, and the fixed cost of training, prompting, and validation eats the gain. Pickens makes this case rigorously: the economical approach depends on the shape of the matter, not on which technology has the better demo. The fair objection to the AI pricing critique is that someone has to pay for the compute — the only question is who, and how visibly.

Bill the value, not the tokens

The fair objection, though, doesn't save the current model. It argues for a better one. When a vendor won't disclose the unit allowance or the overage rate for its flagship AI product, the market cannot price the technology — and what can't be priced gets billed on trust. That was the old game, with gigabytes. It is the new game, with tokens.

What comes next is visible. The per-gigabyte meter is dying; DISCO has already read it the last rites. The per-document AI fee will get arbitraged down by competition — Elevate, HaystackID, and the ALSP tier will run hybrid TAR-plus-GenAI workflows at thin margins and force the platforms to justify their premiums. The winners will sell outcomes, not units: a review done, defensible, at a fixed price.

The demo that ran eleven minutes was real. The savings are real. The only fiction is the invoice — and it has about five years before the data growth curve makes it indefensible. A market headed for a 3.13x productivity mandate cannot keep billing like the humans are still reading every page. They aren't. The only question left is when the bill says so too.