AI Compresses Discovery. Accountability Stays With the Lawyers.

Start with the business problem. In complex commercial litigation, why has the sheer volume of documents become such a pressing commercial issue?

Discovery is where most of the cost and most of the risk in a commercial case now sit. Twenty years ago a large matter might have meant a few hundred boxes of paper. Today a single dispute over a supply agreement can involve email, chat logs, shared drives, and system data running into millions of records per custodian. Someone has to look at that material, decide what is responsive to the requests, and withhold what is privileged.

Historically we did that with contract reviewers and keyword searches, billed by the hour. Clients resent paying for it, and courts have grown less patient with it. Proportionality is now a live standard: a judge can ask whether the burden of a review is justified by the amount in controversy. If your only tool is human eyes on every page, you lose that argument, or you spend the client's budget before you reach the merits. The commercial pressure is plain. We had to find a defensible way to review more material, faster, for less.

Technology-assisted review is not new. What has actually changed that makes AI viable now in a way it was not five years ago?

You are right that the category is old. Courts have accepted technology-assisted review since 2012, when Judge Peck approved predictive coding in Da Silva Moore, and later rulings such as Rio Tinto v. Vale reinforced that a sound process is defensible. That earlier generation of tools classified documents statistically. You trained a model on a sample the lawyers coded, it ranked the rest by likely responsiveness, and you validated the result with sampling. It worked, but it mostly told you where to look.

What changed is the arrival of large language models that read a document and reason about its content in something close to plain language. Instead of a relevance score, we can ask what a memo actually says, whether it touches a specific issue, or why it might be privileged. That shifts the technology from sorting to comprehension. The caution is that these newer models can be confidently wrong, so we treat their output as a draft to be checked, not an answer to be trusted.

Walk us through how the workflow actually operates, from collection to production.

It runs in stages, and lawyers set the parameters at each one. We collect and process the data, then use classic technology-assisted review to rank the full population by likely responsiveness, so reviewers see the most probable material first rather than working through noise.

On top of that ranking we layer generative models for the harder judgments. For a promising document the system drafts a short summary, proposes issue tags, and flags language that may carry privilege, such as apparent legal advice from counsel. A reviewing attorney sees the document beside that draft and confirms, corrects, or overrides it. Privilege and responsiveness calls are never left to the machine.

Everything is logged: the seed decisions, the model version, the validation samples, and who reviewed what. That record is not bureaucratic overhead. If our production is challenged, we can show the court and opposing counsel exactly how the review was conducted, which is the heart of defensibility.

This is highly confidential material, often privileged. How do you handle security, and what ethical or procedural rules constrain you?

Confidentiality is the first design constraint, not an afterthought. Client data sits in an environment we control, governed by a data processing agreement with any vendor, and client material never trains a public model. Before a matter starts we address AI expressly in the protective order and rely on Rule 502(d), so an inadvertent privilege disclosure can be clawed back without waiver. Recent amendments to the Federal Rules have pushed everyone to make those provisions explicit rather than assumed.

The ethical obligations are just as binding. Our duty of competence now includes technological competence, our duty of confidentiality governs what we feed any tool, and our duty of supervision means a lawyer answers for the work product however it was produced. Rule 26(g) requires an attorney to certify that a discovery response is reasonable after a reasonable inquiry, and you cannot certify what you have not supervised. Governance and ethics point the same way, and both keep a named lawyer answerable for what we file.

How do you know the system is accurate, and who is accountable when it is wrong?

We measure it, and we assume it will be wrong some of the time. For the classification stage we use the established metrics: recall, the share of truly responsive documents the process found, and precision, how much of what it flagged was actually relevant. We validate with statistical sampling of the documents the model set aside, so we can defend the completeness of a production with numbers rather than faith.

The generative layer needs a different kind of check, because its failures look plausible. A model can produce a fluent summary that misstates a document or, in legal research, invent a citation outright. Courts have sanctioned lawyers for exactly that, and the fines are no longer trivial. So nothing the model asserts reaches a filing or a production without a person confirming it against the source. Accountability is unambiguous: the supervising attorney owns the output. We treat the tool the way we would treat a junior associate's draft, useful, faster, and never filed without a qualified lawyer reading it first.

How did the firm adopt this in practice? Lawyers are not famous for embracing new tools.

There was real skepticism, and some of it was healthy. Senior litigators worried, correctly, that a fluent summary invites overconfidence. Our first mistake was letting reviewers lean on the machine's issue tags without reading closely, and in one internal test the summaries were smooth enough that people stopped scrutinizing them. Fluency is persuasive in a way a spreadsheet is not.

We fixed that by redesigning the workflow rather than the tool. Reviewers now see the source document first and the draft second, and we audit a fixed percentage of confirmations to make sure people are genuinely checking. We trained everyone on where these models fail, not just how to prompt them, and we made clear that speed is never a defense for an error that goes out the door. The associates adapted quickly once the role was reframed. Their job is no longer first-pass reading. It is judgment, exception handling, and building the case, which is the work they trained for and the work clients pay a premium to get.

Beyond document review, you mentioned using AI to pressure-test case theory and prepare for trial. What does that look like?

Once the record is organized, the same comprehension helps build the case. We use the system to assemble a chronology across thousands of documents, extracting dates, actors, and events into a timeline a person can then interrogate. That surfaces gaps and contradictions faster than a team reading in sequence.

For depositions it is genuinely useful. Before we question a witness, we can pull every document that person touched and a summary of what each one shows, so we walk in knowing the record cold and are far harder to surprise. We also use it against ourselves. We ask the model to argue the other side's strongest version of the facts and to find the documents that cut against our theory. That is uncomfortable, which is the point. Better to meet a weakness in a conference room than in front of a jury.

None of this replaces the lawyer's theory of the case. It stress-tests it. The narrative, the strategy, and the judgment about what a fact-finder will believe remain ours.

What measurable results have you seen, and how should a reader weigh them?

I would offer them with caveats, because honest numbers in this field are always contextual. On first-pass document review we have cut the time by roughly seventy percent on the matters where we have deployed this, which on a large case means days saved at the stage that usually delays everything downstream. We have triaged millions of documents on a single matter with attorney oversight, work that would have been painful to staff by hand.

The caveats matter. Those figures depend on the data being reasonably clean and the issues well defined; a messy collection narrows the gain. The seventy percent is first-pass throughput, not a claim that we removed lawyers from the process, because we did not. And the savings are real only because the validation held up. If a production later proved unreliable, the speed would be worthless. Measured against the right baseline, the economics have shifted enough that we can take matters, and defend them, that the old cost structure made marginal.

What would you say to a skeptical peer, or to a client nervous about AI touching their case?

To the client, I would say the accountable lawyer has not changed and the confidentiality protections are stronger than most people assume. The technology lowers cost and speeds review, and a person still stands behind every decision. To a skeptical peer I would be blunt about the failure modes, because the risks are concrete. These models can fabricate authority, they can misread a document with complete confidence, and they can leak information if you are careless about where the data lives.

The answer to all three is discipline, not enthusiasm. Verify every citation and every factual assertion against the source. Keep the data in a controlled environment. Document the process so it is defensible. Treat output as a draft. The firms that get into trouble are the ones that adopt the speed without the controls, and colleagues have already been sanctioned for exactly that. Used carefully, this is a serious improvement. Used carelessly, it is a malpractice claim waiting to happen. The difference is entirely in the supervision.

Can you give a concrete moment where this mattered, and what it tells you about where the profession is heading?

On one commercial matter the opposing party produced a large volume late, close to a filing deadline. The conventional path would have been an emergency team of reviewers and a lost weekend. We ran it through the workflow, ranked and summarized it overnight, and by morning had a chronology and a shortlist of documents that changed how we framed a key argument. One summary flagged an internal email that contradicted the other side's stated timeline; an associate confirmed it against the original, and it became a meaningful point at deposition. Speed did not win that. It bought us the time to exercise judgment.

For the profession, that is the pattern. Work that can be structured, reading, sorting, first-pass summarizing, is compressing toward minutes. What clients will pay for is what remains: strategy, credibility with a court, and the judgment to know which fact matters. Firms that treat AI as a way to shed lawyers will misread it. Those that use it to redirect senior time toward the hardest questions will serve clients better.