Law firms are using AI for document review and eDiscovery primarily in three ways: automated first-pass responsiveness classification, AI-powered privilege detection and log generation, and production and redaction automation. Platforms like DecoverAI run multiple language models simultaneously on each document to achieve consensus-based classifications — reaching F1 accuracy of 0.86 at $0.017 per document, compared to $1.50–$3.00 per document for traditional contract review.
This is not a description of where legal AI is heading. It is a description of how high-performing litigation practices are running matters today, in 2026. The question has shifted from whether to use AI to how to implement it defensibly — how to validate results, document the workflow, and satisfy FRCP obligations when opposing counsel or a judge asks how your review was conducted.
First-Pass Responsiveness Review: From Weeks to Hours
Traditional first-pass review worked like this: a project manager batched documents into sets of 500 and distributed them to contract reviewers billing at $35–$60 per hour. At an average throughput of 50–75 documents per hour, the all-in cost per document landed between $1.50 and $3.00 — before quality control, privilege screening, or any escalations. On a matter with 200,000 documents, first pass alone ran $300,000 to $600,000.
AI classifiers change that arithmetic entirely. A large language model reads each document against a written responsiveness definition and returns a binary classification — responsive or non-responsive — along with a reasoning trace explaining which facts in the document drove the decision. The attorney reviewing the output sees not just the label but the model’s reasoning, which makes QC tractable and the workflow auditable.
The more defensible approach is multi-model consensus review. Rather than relying on a single LLM, the platform runs multiple models simultaneously: Claude, GPT-4o, and others, each evaluating the same document against the same definition independently. Documents where all models agree are cleared with high confidence. Documents where models disagree are flagged for attorney review. This architecture surfaces genuinely ambiguous cases rather than hiding them inside a single model’s output — and it is the approach behind DecoverAI’s published F1 score of 0.86 at $0.017 per document.
Before AI review begins, the team manually reviews a seed set of documents — typically 200–500 — and uses them to calibrate the model. Recall targets (typically >75%, often higher for high-stakes matters) are agreed on with opposing counsel before review starts. A QC sampling protocol then tests a statistically significant random sample of non-responsive documents to catch false negatives before production. Courts require this documentation; it is not optional.
On the court acceptance side, the trajectory is clear. Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012) was the first US court to approve predictive coding — the forerunner of today’s LLM-based review. Hyles v. New York City, No. 10 Civ. 3119 (S.D.N.Y. 2016) went further: Judge Peck ordered TAR over keyword search when TAR was demonstrably more accurate, writing that courts should not “condone exhaustive manual review.” The trend is now settled: courts expect technology-assisted review to be used when it is more accurate than the alternatives, and question its absence.
Privilege Detection and Automated Privilege Log Generation
Privilege review is two separate AI tasks that operate in sequence.
The first task is classification: does this document contain attorney-client communication, attorney work product, or common interest privilege? AI classifiers look for signals attorneys would look for: attorney names appearing as authors or recipients, legal department email domains, “Privileged and Confidential” header language, and the character of the communication (legal advice versus business direction). The model returns a privilege flag and a confidence score; high-confidence privilege calls are routed to the privilege review queue, while borderline determinations get attorney attention first.
The second task is log entry generation: a privilege log entry must describe the subject matter of a withheld document without revealing the privileged content. This is exactly the kind of controlled summarization task LLMs handle well. The model reads the document, identifies the relevant privilege context, and produces a description such as “Email from in-house counsel to CEO re: regulatory compliance strategy” — informative enough to satisfy FRCP Rule 26(b)(5) without disclosing the advice itself.
The attorney role remains under FRCP Rule 26(g): the attorney certifies the privilege log. AI drafts; the attorney reviews and approves. What changes is the cost structure. Manual privilege log drafting has run $10–$15 per entry. AI drops the drafting cost to near zero; attorney time collapses to a structured QC pass over entries already drafted.
Production and Redaction Workflows
Production — formatting, numbering, and delivering reviewed documents to opposing counsel — has historically been manual and error-prone. AI handles two parts of this that are particularly high-value.
The first is AI-assisted redaction identification. Responsive documents often contain PII that must be redacted before production: Social Security numbers, credit card numbers, dates of birth, protected health information. AI classifiers identify these fields automatically, surfacing redaction candidates with far greater consistency than a reviewer scanning documents manually under time pressure.
The second is pre-production QC. Before a production goes out, the platform runs a check that flags missing metadata, corrupt or incomplete files, unredacted fields that should have been redacted, and format inconsistencies. Catching these before delivery is far cheaper than dealing with a clawback request after the fact.
Bates numbering, production format selection (TIFF vs. native vs. PDF), and load file generation are automated. The result is that the complete workflow — ingestion, classification, privilege review, production packaging — runs in a single platform rather than across multiple tools with hand-offs between them.
What Courts Have Said About AI in Document Review
The legal foundation for AI-assisted review is well-established. The key cases:
The first US judicial approval of predictive coding. Judge Peck allowed computer-assisted review over plaintiff’s objection, establishing that TAR is an acceptable method of searching for responsive ESI when properly validated and disclosed. The court’s approval was conditional on a transparent seed set protocol and iterative refinement — not the technology itself.
Established the validation protocol requirements that remain the standard today. Parties using TAR must disclose their validation methodology to opposing counsel. The opinion made clear that TAR is not self-validating — the producing party must demonstrate accuracy, document the seed set, and show iterative refinement.
Judge Peck’s most influential opinion: he ordered TAR over keyword search, finding that courts should not “condone exhaustive manual review” when TAR is demonstrably more accurate. The case shifted the frame from “is TAR acceptable?” to “why aren’t you using it?”
Reinforced that AI review is defensible when properly validated and documented. The court’s analysis focused on the transparency of the process, not the technology itself — confirming that documentation and the validation workflow matter as much as the tool.
Courts now expect AI review to be used when it is more accurate than alternative approaches. The question is not whether to use AI — it is how to document it. A producing party that defaults to manual keyword review without justification faces the same scrutiny that TAR skeptics once did.
The Defensibility Checklist for AI-Assisted Review
The following protocol reflects what courts have required and what sophisticated opposing counsel will ask about at a meet-and-confer:
- Document your seed set. How many documents were manually reviewed, how they were selected (random, judgmental, or stratified), and how many are responsive versus non-responsive. The seed set is the foundation of your validation argument and the first thing opposing counsel will ask to examine.
- Agree on recall and precision targets before review begins. Negotiate these thresholds with opposing counsel before the review runs, or raise them at the ESI protocol conference. Typical targets are >75% recall, though high-stakes matters often go higher. Agreements made in advance prevent disputes made after.
- Run validation batches. Test the model against a held-out set of manually reviewed documents and report precision, recall, and F1 against the gold standard. Repeat validation as the model is refined.
- Document the AI model(s) used, version, and configuration. This includes the prompt or written responsiveness definition the model evaluated against, any configuration changes between runs, and the date of each pass.
- QC sampling over non-responsive documents. Review a statistically significant random sample of documents classified non-responsive to check for false negatives before production. This is the step most often skipped and the one opposing counsel focuses on most intensely.
- Attorney certification under FRCP 26(g). The attorney signing the discovery responses must understand the AI workflow well enough to certify it. “The vendor handled it” is not a defensible answer at a meet-and-confer.
- Privilege log generation documentation. How AI descriptions were generated, what instructions the model was given, and what attorney review was applied to each entry before the log was finalized and certified.
How DecoverAI Implements AI Document Review
DecoverAI is built around the multi-model architecture that the case law points toward. The platform runs three or more LLMs simultaneously on each document — consensus decisions are cleared, disagreements are surfaced for attorney review. The attorney’s attention goes to documents that actually need it, not to every document in the collection.
Every classification comes with a reasoning trace: the specific passages and metadata the model relied on to reach its decision. Attorneys reviewing a flagged document see what the model was reacting to, which makes QC faster and privilege review more accurate.
The privilege log workflow is built into the platform, not bolted on. The classifier identifies the privilege population; the LLM generates a first-draft log description for each document; the attorney reviews the drafts in the same interface where the underlying document is rendered. Edits are logged. The finalized log exports in standard FRCP format.
Validation is built in as well. The platform includes a seed set workflow with recall reporting, a QC sampling protocol over the non-responsive population, and exportable audit trails for every classification decision. At $60/GB/month all-in, AI classification, privilege log generation, QC, and production are all included — no per-document charges, no per-user fees, no add-on modules for features that should be standard.