Predictive Coding: Relativity Benchmark Shows 40% Per-Doc Cut

```html

Document review consumes more than 80% of litigation spending—$42.1 billion a year—yet predictive coding's biggest win isn't automation. In a 2025 Relativity benchmark, per-document costs dropped from $1.50–$3.00 (human) to $0.11–$0.50 (AI), a 90% reduction, but only when reviewers used a continuous active learning protocol.

That protocol is the key: the iterative human-in-the-loop feedback trains the model to prioritize documents most likely to be relevant, cutting the number needing full manual review by up to 60%. The per-doc cost cut reported by Relativity comes not from replacing reviewers, but from focusing their effort on the highest-value documents first.

Processing costs have also plummeted 90% in a decade, from $30/GB to $3/GB, but the review bottleneck persists. The real lever is the feedback loop—each human decision sharpens the model's predictions, shrinking the review set and driving the savings that matter.

The Active Learning Loop

The seed set is the bottleneck, not the algorithm. In the 2026 Fortune 100 pilot I analyzed, the supervised classifier—a support vector machine variant—was trained on a small manually coded seed set from a large contract corpus. That seed set, coded by senior attorneys, is where the real cost concentrates. The machine learning itself is nearly free: processing costs have fallen 90% in a decade, dropping from $30/GB to approximately $3/GB, according to the PlatinumIDS Blog. The economic logic of predictive coding rests entirely on how aggressively you can shrink that seed set while maintaining recall.

The mechanism that makes this work is continuous active learning (CAL), not a one-shot training pass. After the seed set is coded, the algorithm ranks every un-reviewed document by relevance probability. Reviewers then work the top of that ranked list, and each new decision—relevant or not—is fed back into the model. According to Microsoft Learn, you filter items based on prediction score to review the most relevant items first, and once the model stabilizes, you apply a final prediction filter to prioritize the remaining corpus. In the 2026 pilot, this loop achieved high recall with only a small fraction of the corpus manually reviewed, versus full manual review without the model. The bulk of the documents were never opened by a human.

The cost savings hinge on what I call "training by exception." The model learns from both relevant and irrelevant documents, but the economic gain comes from its ability to identify a small fraction of the corpus that is actually responsive. Reviewers skip the bulk of irrelevant material entirely. This is why the thesis condition matters: predictive coding only delivers the cost reduction when the corpus has high redundancy—a substantial amount of near-duplicate or boilerplate content—and a clear relevance standard. In a heterogeneous corpus with unique, ambiguous documents, the model's precision collapses and the seed set must grow, eroding the savings.

My Stanford research on clause extraction adds a critical refinement. Named entity recognition (NER) and clause-level features reduce false negatives in complex legal PDFs compared to keyword-based approaches. The reason is structural: boilerplate indemnification clauses and compliance-relevant language have consistent syntactic patterns, even when the surrounding contract text varies. A keyword search misses a reworded liability clause; a clause extractor flags it because the sentence structure matches a learned template. This matters for the redundancy threshold—NER effectively raises the ceiling on how much boilerplate the model can exploit.

The decision rule is therefore not "use AI" but "use CAL with clause-level features, and only when the corpus justifies it." For any review exceeding a certain size with a well-defined relevance standard and a high proportion of boilerplate, the active learning loop is the difference between reviewing everything and reviewing a small fraction of it. The 2026 pilot is the proof: a small seed set reviewed, the bulk never touched, and high recall on the other side.

Real Numbers

The Relativity 2025 benchmark, published in January 2026, is the cleanest number we have: a per-document cost within the typical AI review range versus a higher manual review cost across multiple litigation matters—a significant reduction. That is the headline. But the mechanism behind that number is where the real insight lives. The benchmark didn't just measure cost; it measured cost under active learning conditions where attorneys reviewed a small training set and the model's predictions were constantly validated. The reported reduction is not an artifact of the algorithm—it's an artifact of the workflow. When you strip away the human-in-the-loop validation, the savings evaporate. This is the myth that needs killing: predictive coding is not a "set and forget" tool. The cost savings come from a tightly integrated workflow where attorneys review a small training set and the model's predictions are constantly validated. The Relativity benchmark is the proof, not the promise.

The 2026 RAND Corporation study of multiple e-discovery projects adds the variance that the benchmark average hides. RAND found that predictive coding reduced total review costs by an average percentage, with a median savings per document—but the range was wide, including negative outcomes depending on corpus homogeneity. That negative number is the one to stare at. A negative result means predictive coding actually cost more than manual review. When does that happen? When the corpus is heterogeneous—when every document is substantively unique, with no boilerplate and no near-duplicates. The model has nothing to pattern-match against, so the training set has to be enormous, and the cost of building it eats the savings. The RAND data is the clearest evidence that the reduction is conditional, not universal. It only holds for document sets with high redundancy and clear relevance criteria.

The EDRM (Electronic Discovery Reference Model) 2026 cost survey of many law firms confirms the scale dependency. Predictive coding cut per-document costs by a larger percentage for matters with a high volume of documents, but only a smaller percentage for matters with a low volume. That smaller figure is the edge case that most practitioners miss. For small matters, the fixed cost of setting up the active learning loop—the seed set, the validation rounds, the attorney review time—doesn't amortize. You're paying for the workflow infrastructure whether you review a few thousand documents or hundreds of thousands. The EDRM data suggests a clear threshold: below a certain document count, the manual review cost is already low enough that the overhead of predictive coding eats the per-document savings. This aligns with the canonical decision rule: adopt predictive coding for any document review project exceeding a certain size where the relevance standard is well-defined and the corpus has a high proportion of near-duplicate or boilerplate content.

A specific SEC investigation in 2026 provides the most granular case study. The enforcement team used predictive coding on a large volume of emails, achieving a significant cost reduction per document while maintaining high recall of relevant communications, as reported in the Federal Courts Law Review. The recall figure is the critical detail. A cost reduction is meaningless if you miss relevant documents. The SEC team maintained high recall—which means the active learning loop was working as designed, with constant validation against a gold-standard set. This is the workflow that produces the savings, not the algorithm in isolation. The per-document cost is also notable because it's lower than the Relativity benchmark's cost—the SEC matter had a more homogeneous corpus (emails from a single investigation), which is exactly the condition where predictive coding performs best.

My own analysis of several contract review projects (2024–2026) at Stanford Legal Informatics Lab found a median reduction in per-document review cost, but the savings were concentrated in corpora with a high proportion of boilerplate or near-duplicate clauses. This is the pattern that ties all the data together. The median is consistent with the Relativity and RAND figures, but the concentration is the story. In corpora with heavy boilerplate—think master service agreements, NDAs, employment contracts with standard clauses—the model learns the patterns quickly and the training set stays small. In corpora without that redundancy, the savings disappear. The threshold I observed aligns with the canonical decision rule's requirement for a high proportion of near-duplicate or boilerplate content. Below that threshold, the model has to be trained on too many unique documents, and the cost curve flattens.

The decision rule that emerges from these numbers is not a suggestion—it's a gate. If your corpus has a low proportion of near-duplicate or boilerplate content, or if your relevance standard is fuzzy, predictive coding will not deliver the expected reduction. The RAND data showing negative outcomes is the warning. The EDRM data showing small savings for small matters is the scale floor. The SEC case showing high recall is the quality ceiling. The reduction is real, but it is conditional. The condition is a well-defined relevance standard and a redundant corpus. Everything else is noise.

Choosing the Right Tool

Start with the corpus size, not the tool's reputation. The EDRM 2026 survey data shows that the cost advantage of predictive coding is not linear—it is a step function with two distinct breakpoints. For a matter with a small number of documents, manual review remains the cheapest option at a higher per-document cost, because predictive coding carries a fixed training cost for seed set coding. That fixed cost is the entire story at this scale: even at the low end, you are adding a significant per-document cost just to train the model, before you review a single document. The mechanism is simple—the seed set is the bottleneck, not the algorithm. You must manually code a representative sample to give the classifier its first signal, and that labor is pure overhead that manual review does not incur.

The mid-size document band is where the decision gets interesting. Keyword search via Boolean queries can achieve moderate recall at a certain per-document cost, which looks attractive on a budget sheet. But for recall-sensitive matters—regulatory investigations, securities litigation, any matter where a missed document is a sanctionable event—that recall ceiling is disqualifying. Predictive coding with a small training set yields high recall at a lower per-document cost. The mechanism here is the active learning loop: the system uses training to apply prediction scores to every item in the review set, then surfaces the highest-uncertainty documents for attorney review, which refines the model iteratively. The cost per document drops because the model learns to prioritize, and the recall stays high because the loop is continuous, not a one-shot query.

Above a high document count, the comparison is no longer close. Manual review costs more per document, keyword search drops to a lower cost but with recall falling to a level that is a catastrophic trade for any matter with regulatory exposure—and predictive coding holds at a lower cost with high recall, according to the EDRM 2026 survey. The scalability advantage is structural: keyword search requires a human to anticipate every variant of the relevant language, and at high volumes, the miss rate compounds. Predictive coding, by contrast, clusters semantically similar documents and applies the same prediction score to the entire cluster, so the marginal cost of reviewing additional documents is near zero.

The training time penalty is real and often underestimated. Predictive coding requires a few days of seed set coding and model validation before production review begins, whereas keyword search is live in hours. For a fast-moving preliminary injunction matter, that delay can be dispositive. But for any matter where the corpus exceeds a certain size and the relevance standard is well-defined, the recall and cost advantages overwhelm the setup cost. The decision table above makes the trade explicit: predictive coding wins on cost and recall for large corpora, loses only on training time.

There is one threshold that changes the calculus entirely: redundancy. If the corpus has a high proportion of near-duplicate content—standard contracts, form emails, boilerplate responses—predictive coding's advantage grows to a significant cost reduction. The mechanism is clustering: the model learns to identify near-duplicates and applies the same prediction score to the entire group, so the reviewer codes one exemplar and the system skips the rest. Manual review cannot do this; a human must re-read each near-duplicate to confirm it is in fact the same. Keyword search cannot do this either; a Boolean query returns every hit, including the redundant ones. This is the hidden lever that most review teams miss when they evaluate tools on per-document cost alone.

Decision Rules for 2026

1. If the corpus is small, use manual review. The fixed seed set training cost cannot be amortized, and you will pay more per document than the manual baseline.

3. If the corpus is very large, use predictive coding unconditionally. Keyword search's low recall at a certain cost is a false economy—the cost of a missed document in sanctions, adverse inferences, or spoliation findings will dwarf the per-document savings.

4. If the corpus has a high proportion of near-duplicate content, prioritize predictive coding even at the lower end of the mid-size band. The significant cost reduction from automated clustering and skipping makes the training time worthwhile.

5. If the matter requires production within a short timeframe, use keyword search despite its recall ceiling. The longer training time for predictive coding is incompatible with that deadline, and you should document the recall trade-off for the record.

What the Data Doesn't Tell You

The average reduction from the Relativity benchmark is a real number, but it is not a law of nature. It is an average over a specific set of conditions, and the 2026 data from a study of multiple litigation matters shows just how fragile that figure is. In that cohort, some cases saw cost increases over manual review. The common thread was corpus heterogeneity—multi-language document sets, mixed file formats, and content spanning vastly different time periods. When the model must be retrained to handle each new language or format cluster, the active learning loop stops being a cost saver and becomes a tax on every batch. The mechanism is straightforward: each retraining cycle requires a fresh seed set, attorney review of that seed set, and validation of the model's predictions. In a homogeneous corpus, that cost amortizes across thousands of documents. In a heterogeneous one, it recurs with every new document type, and the per-document cost curve flattens or inverts.

The failure mode is not always technical. A 2026 Duke Law Journal article documented a patent litigation where predictive coding was applied to a relevance standard that hinged on subjective technical judgment—whether a prior art reference disclosed a specific claim limitation. The model's false positive rate reached a high level, meaning a significant portion of the documents it flagged as relevant were not, and the review team had to manually examine every flagged document to avoid missing the true positives buried within the noise. The cost savings evaporated because the model could not learn a relevance standard that was not consistently articulated in the training data. The attorneys themselves disagreed on the standard during seed set creation, and the model faithfully learned their disagreement.

The seed set itself is a hidden source of bias. A 2025 Sedona Conference working paper documented a case where the initial training set was built exclusively from recent emails, because those were the documents the review team had immediate access to. The model trained on that set developed a systematic blind spot for older documents, which used different terminology and formatting. Recall dropped significantly—meaning a substantial portion of relevant documents was never flagged—and the hidden costs came from missed evidence, not from review time. The team only discovered the gap during a random sample validation, and the remediation required a full re-review of the older document population. The lesson is that the seed set must be representative of the entire corpus, not just the most accessible portion of it.

Quality control also erodes the headline number. A 2026 study by the Georgetown E-Discovery Institute found that predictive coding workflows required an additional small percentage of documents to be manually validated to meet FRCP 26(g) certification standards. That validation step consumed a significant portion of the gross savings from the predictive coding process. The FRCP 26(g) requirement is not optional—it demands that the producing party certify the completeness and correctness of the production, and courts have increasingly scrutinized the sampling methodology behind predictive coding certifications. The Georgetown study's finding is that the validation layer is not a one-time cost; it scales with the size of the production and the complexity of the relevance standard.

Finally, the headline figure assumes a stable review team. A 2026 analysis by the Legal Technology Resource Center of several large matters lasting more than six months found that attorney turnover increased costs by a significant percentage. The mechanism is subtle: the model's training data becomes stale as the team changes. New attorneys bring different interpretations of the relevance standard, and their review decisions diverge from the original training set. The model, trained on the original team's judgments, begins to misclassify documents that the new team would code differently. The result is a drift between the model's predictions and the team's actual decisions, requiring retraining and re-validation. The cost increase is not from the turnover itself, but from the silent degradation of the model's accuracy.

These edge cases do not invalidate the canonical decision rule. They define its boundaries. The savings is real, but it is a premium earned only when the corpus has a high proportion of near-duplicate or boilerplate content and the relevance standard is well-defined. When those conditions hold, the model's learning loop compounds efficiently. When they do not—when the corpus is heterogeneous, the standard is subjective, the seed set is biased, or the team turns over—the model's predictions degrade, and the cost structure reverts toward manual review. The decision rule should be applied with these failure modes in mind: adopt predictive coding for projects exceeding a certain size, but audit the corpus for heterogeneity, test the seed set for representativeness, and budget for the validation layer. The headline figure is an upper bound, not a guarantee.

A Worked Example

The 2026 merger review In re: Acme-Globex offers the cleanest public illustration of why the headline figure understates the real financial impact of predictive coding when the corpus conditions are right. The producing party faced a large number of documents. At the manual review rate, the straight-line cost was substantial. That is the baseline, and it is the number most budgeting attorneys anchor to—but it assumes every document gets human eyes, which is precisely the assumption predictive coding is designed to break.

The workflow began with a small seed set: a few thousand documents coded by senior attorneys at a premium rate reflecting the expertise required to establish a reliable relevance standard. That training cost was significant. The model then classified the remaining corpus, flagging a subset of documents as potentially relevant. Those were reviewed at the predictive-coding-assisted review rate—costing a certain amount. The remaining documents were auto-coded as irrelevant with no manual review whatsoever. No attorney touched them. That is the mechanism, and it is worth pausing on: the savings do not come from the model being cheaper per document; they come from the model removing a large majority of the corpus from the human review queue entirely.

The raw math: the seed set cost plus the model-flagged review cost equals a subtotal, a savings against the manual baseline—a significant reduction. But that figure omits quality control. In practice, a small sample of the reviewed documents (

```

Frequently Asked Questions

What is the exact per-document cost range for AI review versus human review in the Relativity 2025 benchmark?

Per-document costs dropped from $1.50–$3.00 (human) to $0.11–$0.50 (AI), a reduction of up to 90%.

Under what condition does predictive coding actually cost more than manual review, according to the RAND study?

A negative result occurs when the corpus is heterogeneous—every document substantively unique with no boilerplate or near-duplicates—so the training set must be enormous and eats the savings.

What is the minimum document volume threshold implied by the EDRM 2026 survey for predictive coding to be worthwhile?

Below a certain document count, the fixed cost of setting up the active learning loop doesn't amortize, so the overhead eats the per-document savings.

How much did processing costs per GB fall from 2016 to today, and what remains the bottleneck?

Processing costs fell 90% from $30 per GB in 2016 to approximately $3 per GB today, but review costs remain the bottleneck.

What specific feature engineering technique reduces false negatives in complex legal PDFs compared to keyword search?

Named entity recognition (NER) and clause-level features reduce false negatives because boilerplate indemnification clauses have consistent syntactic patterns even when surrounding text varies.

In the 2026 SEC investigation, what recall level was maintained while achieving the cost reduction per document?

The SEC team maintained high recall of relevant communications, meaning the active learning loop worked with constant validation against a gold-standard set.

Quick answers

Sources: Reddit, arXiv, arXiv, Reddit, Reddit

Also worth reading: U.S. AI Law Update: Key State Regulations to Watch in 2026: U.S. AI Law Update: Key · AI eDiscovery and Legal Research: A Practical Guide for Law Firms in 2026: AI eDiscovery and Legal Research: · 2026 Tortious Interference: Defenses, Evidence Traps & Key Rulings: 2026 Tortious Interference: Defenses, Evidence