AI Answer Summary

Courts have approved technology-assisted review since 2012. In July 2026 a federal court accepted generative AI for final responsiveness determinations in discovery and treated it as a form of technology-assisted review rather than as something new.

So the question has moved. It is not whether machine-assisted review is defensible. It is whether a particular firm's process is, which turns on written criteria fixed before the run, validation by sampling, and a lawyer making the calls that need legal judgment.

For a mid-market litigation practice the gap is practical rather than legal. A hundred-lawyer firm sees matters too large to read line by line and too small to justify a full platform engagement, so the work either goes out at cost or gets absorbed by associates at night. That middle is where a firm-owned workflow pays for itself.

What AI Document Review Covers in Litigation

AI document review covers four jobs, and courts treat them differently.

Worth separating them before anything else, because "document review" sounds like one decision and is not. The defensibility question attaches differently at each stage, and firms get into trouble by buying for one stage and assuming it covers the others.

The pattern shows up in procurement conversations. A litigation group buys something that culls a collection well, then discovers it has bought nothing that helps with the privilege log, which is the stage that had been consuming partner time all along. The four stages below have different evidence behind them, different judicial treatment, and different answers to the question of who decides.

One idea runs through all four. Courts have never approved a tool. They have approved processes that happened to use one, and the approval always came with conditions attached.

What Courts Have Approved: From Da Silva Moore to Schulte v. LinkedIn

Fourteen years of authority, and four decisions mark the line.

Da Silva Moore (2012): The First Approval of Technology-Assisted Review

Magistrate Judge Andrew Peck, Southern District of New York, issued the first judicial opinion approving predictive coding, 287 F.R.D. 182. He held it was an available and judicially approved tool in appropriate large-volume cases, provided counsel ran a reasonable process with quality control and proportionality safeguards.

The conditions were part of the approval from day one. That detail gets lost in the retelling, and it is the whole basis of the defensibility section below.

Rio Tinto (2015): TAR Becomes Black Letter Law

Peck again, three years later, describing the position as settled. A producing party does not need permission to use technology-assisted review, and it should not be held to a higher standard than any other review method just because the technology produces metrics that keyword searching never did.

That second half matters more than firms realise. The metrics a machine-assisted process generates can be used against you, and Rio Tinto says they should not be.

Hyles (2016): The Court Will Not Force You Either

Here is where the received wisdom gets it wrong, so it is worth being precise.

Hyles v. New York City, No. 10 Civ. 3119 (AT)(AJP), 2016 WL 4077114 (S.D.N.Y. Aug. 1, 2016). A plaintiff asked the court to compel the City to use TAR rather than keyword searching. Peck refused. His opening line: the short answer is a decisive "NO."

He was not hostile to the technology. He wrote that for most cases TAR was the best and most efficient search tool, and that he would have liked the City to use it. But Sedona Principle 6 controls: the responding party is best situated to decide how to search its own documents. Then the sentence worth carrying into 2026: there may come a time when TAR is so widely used that it might be unreasonable for a party to decline to use it, and we are not there yet.

That was ten years ago.

Schulte v. LinkedIn (2026): Generative AI for Responsiveness Calls

Schulte v. LinkedIn Corp., 2026 WL 1905851, No. 22-cv-00237-HSG (N.D. Cal.), decided 1 July 2026 by Magistrate Judge Laurel Beeler. The first federal decision accepting generative AI for final responsiveness determinations in discovery.

Two details carry more weight than the holding. The use of generative AI for those calls was neither challenged by the plaintiffs nor questioned by the court. And Judge Beeler described the generative tool as a form of technology-assisted review, which places it inside the Da Silva Moore line rather than opening a new category.

The analysis explaining that decision was co-authored by Judge Peck, now retired, who wrote the 2012 opinion that started all of this.

What Courts Now Require: Conditions on Generative AI in Discovery

A group of 2026 decisions permits generative AI in discovery subject to conditions, and those conditions read like a procurement checklist.

Commentary on these decisions describes the practical effect as a line between closed enterprise systems and publicly available consumer tools.

That line is worth noticing, because a different set of courts drew the same one for a different reason. In the privilege context a federal court held that a defendant's use of a consumer AI tool defeated confidentiality, partly on the terms of the platform's own privacy policy. Two unrelated legal questions, one distinction: what the platform's terms say about your data.

A caution on this table. These conditions and the decisions behind them reached us through e-discovery commentary rather than from the opinions. We have flagged them for confirmation before anyone relies on the specific citations. The pattern is well corroborated. The individual case references are not yet.

Is AI Document Review More Accurate Than Manual Review?

On the published evidence, yes, and the finding is old enough to have shaped the case law that followed it.

Maura Grossman and Gordon Cormack published their study in the Richmond Journal of Law and Technology in 2011, concluding that technology-assisted review yields more accurate results than exhaustive manual review with much lower effort. It has since been cited in multiple reported judicial decisions and sits underneath the whole line of authority above.

Newer independent work points the same way for this class of task. Vals AI, with Legaltech Hub, tested legal AI products against a control group of practising lawyers and found the tools ahead on document questions, summarisation and transcript analysis, and behind on amending documents. Review work sits entirely on the winning side of that split.

Now the counterweight, because accuracy at reading is not accuracy at everything. Peer-reviewed testing of legal research tools measured hallucination rates between 17% and 33%. The difference is what the system is being asked to do. Reading material it was given is a different problem from supplying authority it was not, and everything in this article sits on the first side of that line.

There is an uncomfortable implication for a firm still running linear review. The empirical case against reading everything by hand has been public since 2011 and cited by courts since 2012.

The Mid-Market Gap: Too Big to Read, Too Small for a Platform Engagement

The problem at a hundred lawyers is not whether to use technology. It is that the available options are priced for a different size of matter.

Three routes exist today, and a mid-market litigation group usually cycles through all of them.

Send it out to a provider. This works, and it transfers margin on work the firm could have kept.

Absorb it internally. This shows up as associate hours that get written down at the pre-bill, which means the firm pays for the review twice: once in salary, once in realisation.

Run a full platform engagement. Appropriate for the matter sizes a large firm sees regularly, harder to justify for a recurring medium-sized production.

The case this article is about is the fourth one: a practice group that sees the same kind of production constantly. Insurance defense, employment, commercial disputes with document-heavy discovery. Never enormous, never small, and always arriving.

That is also the lowest-risk place to automate, for a reason that has nothing to do with technology. Review work is frequently where firms already write time off. Automating billable work reduces the invoice, since on hourly matters you bill the time you spent. Automating work that was already being written down costs nothing in collected revenue.

For a single enormous matter, a provider or a platform engagement is still the right answer, and we would tell you so on the call.

How to Make an AI Document Review Process Defensible

Defensibility is a process you can describe, not a product you bought.

Six elements, with what your firm should be able to produce for each.

The disclosure row is the one firms discover too late. In Schulte the producing party disclosed its intended use under the operative ESI order, which is part of why the question arrived at the court in an orderly way rather than as a dispute about concealment.

Here is the test a COO can apply without understanding any of the technology. If opposing counsel asks how the review was conducted, can someone at your firm answer in specifics, today, without calling the vendor? If the answer is no, the process belongs to the vendor rather than to you. That is a position, not a protection.

What Stays With a Lawyer: Privilege Review

A system can flag likely privilege and draft a first-pass log entry. The determination stays with a qualified lawyer, every time.

Privilege is the clean example because it is two tasks in sequence and only the first is a pattern-matching problem. Classification asks whether a document looks like an attorney-client communication or work product. Judgment asks whether it is, which requires knowing the relationship, the scope of the representation, and the jurisdiction.

The asymmetry in the stakes is what settles it. A missed responsive document usually gets found and produced later. An inadvertently produced privileged document has already been read, and a clawback provision is a remedy rather than a cure.

So the framing for this entire article, in one line: the machine reduces what a lawyer has to read. It does not reduce what a lawyer has to decide.

How Codebridge Builds Document Review Triage

We build triage into the systems a firm already runs, so a production arrives sorted, labelled and prioritised rather than raw.

The criteria are the firm's and they are written down. The approval checkpoint is designed in during the first conversation rather than added after a security review. And the record of what ran, on what set, with whose approval, is part of the build instead of something reconstructed later, which is what the defensibility table above requires a firm to have.

One workflow goes live in three weeks. Your firm owns the repository, the prompts and the configuration from day one.

The closest reference we can offer, labelled for what it is: Knowledge Cloud, built for a Big Four tax and legal practice, runs an expert review queue with an immutable audit log so a senior practitioner approves each output before the firm acts on it. A research platform rather than a discovery system. What it demonstrates is the review and audit pattern this kind of workflow needs.

Our founding team spent more than a decade at KPMG.

If you want to look at which productions are costing your firm the most, book a 20-minute call.