Red-Team the Record: Find Your Client's Worst Documents First

Red-teaming the record starts with the questions opposing counsel will ask, put to your own client's documents first.

Red-Team the Record is the practice this article proposes for a client's own documents, and the research section below sets out what the rules expect of it.

Every litigator assumes the client's documents contain something unpleasant. Few know which document it is. Fewer still know which element of the case it damages. A speaker on a 2023 episode of the Personal Jurisdiction podcast put the underlying problem plainly: lawyers often do not know the answer when a client first comes to them, and knowing where to look is part of the job.

Where to look is, I think, less mysterious than it seems. Look where the other side will look.

So the first question to ask of any new matter is a blunt one: what would I least like opposing counsel to find in these files? The difference is not semantic, as it were, because it decides whether the worst email in the collection reaches your Rule 30(b)(6) witness in a quiet preparation session or, for the first time, across a conference table with the court reporter typing. What follows is a working definition of the word worst, a way of turning claim elements into questions, and a method for ranking what those questions return, beginning with the reason the step can no longer be postponed.

Why should you find your client's worst documents before the other side does?

AI-assisted review now runs cents per document, against dollars per document and roughly $19K/GB for manual review, so reading your own client's collection adversely is finally affordable.

Before deciding whether an adverse pass on your own record is worth running, I would put three questions to the matter:

1. Does the collection hold more email, chat and files than anyone at the firm has actually read?
2. Will a corporate witness be prepared for a Rule 30(b)(6) deposition before your own review is finished?
3. Would one bad document change the settlement number, the pleading or the budget if you knew about it today?

A yes to any of them is reason enough. Yet the habit of the profession runs the other way. The common assumption is that only the other side's review is dangerous, and that one's own collection need merely be sorted for production. That assumption is, in a sense, a courtesy we extend to our clients; it is also the reason the worst document so often surfaces in the hands of the person least inclined to be gentle with it.

The reluctance is partly cultural. A speaker on a 2023 episode of the Personal Jurisdiction podcast observed that many law students want to litigate but do not want to spend their first two years reviewing documents. The same speaker, who did a great deal of e-discovery early on, called knowing how to do it essential to running a case, and described finding damaging emails as one of the enjoyable parts of litigation. Both things are true at once. The work is resented. The work is also where cases turn.

What has changed is less the law than the arithmetic and the architecture. In our own deployment model, processing can run in-account on AWS under the client's own keys, a trust posture no consumer AI tool and almost no small-firm tool can match. Originals stay immutable, with content hashing, append-only audit trails and documented chain of custody, so an adverse pass leaves a process record that holds up if opposing counsel challenges it. (Whether one wants a durable record of having hunted one's own worst documents is, of course, a fair question; I would rather have the record than the surprise.) Anyone weighing the cloud bill should look at what AI coding 100,000 documents costs in your own AWS account before assuming the step is out of reach.

Read together, 4 evidence items point one way: the adverse read is now cheap, defensible and private enough to do first. Ask, then, who on your side has actually read the custodian with the most email. The habit has not caught up.

A practical first step is to list the elements your client must prove or defeat, then ask which of them a single email could undo.

Can a cheaper document review find the damaging documents on its own?

Cutting review cost does not find damaging documents by itself; an adverse question does, and only if each answer links back to the exact exhibit for one-click verification.

A responsiveness review is a sorting exercise. It asks whether a request for production covers a document, and it codes the answer as produce, withhold or log. That question is legitimate, and indeed required. It is simply not the question that hurts you.

Consider the same email read twice. Under a responsiveness review, a message from the client's operations manager about a delayed shipment is coded responsive to Request 14, and the reviewer moves on. Under an adverse reading, the identical message is asked something else: does it show the client knew of the defect before the date it told the buyer everything was on schedule? Coded the first way, the email is a line in a production log. Read the second way, it is the exhibit opposing counsel will hand your witness, and the one your witness has never seen.

Here the tension sharpens, because a cheaper review can make the problem worse rather than better. When per-document cost falls, the temptation is to run the same responsiveness pass faster, on more data, and call it diligence. Speed is not the same as attention. A review that sorts more documents more cheaply still sorts them on the wrong axis, and a damaging document coded responsive is no less damaging for having been coded efficiently. Semantic search carries a version of the same risk, which is why I would read why AI semantic search still misses the documents that matter before trusting any single query to surface the worst of the record.

The second friction is trust. An AI answer to an adverse question is useless, and possibly dangerous, if the attorney cannot see where it came from. Our own design answer is architectural: every output links back to the exact exhibit it came from, so the attorney verifies before relying on it, which is the structural response to the sanctions risk that Mata v. Avianca made familiar. The machine proposes; the lawyer still reads.

The third friction is scale of practice. Enterprise platforms are built for matters with an administrator; the small firm holding one messy collection usually has none. In our analysis of where adverse self-review breaks down, that is the segment priced out first, which is why we built for enterprise-grade defensibility at small-matter economics. The proof we offer is deliberately plain: bring a messy collection and a hard question, and watch the tool read, code and cite your own evidence rather than reading a case study about someone else's.

So the cheaper review and the adverse review are not rivals. They answer different questions, and the next step is learning to write the adverse ones well.

How do you keep privileged documents out of an adverse review of your own record?

Run the adverse questions inside a single-tenant environment or your own AWS account, under your own keys, with no vendor retention and no model training, after privileged material is screened out.

The worry is reasonable, and I would not wave it away. An adverse review is, by design, a search for the client's most sensitive material, and some of that material will be privileged: the general counsel's email weighing the very defect the plaintiff alleges, the outside lawyer's memo on exposure. To hand those documents to a tool whose data handling you cannot describe is to create, in the act of protecting the client, a new way of harming it.

The resolution is an order of operations rather than an act of faith. I would put it in three steps:

1. Screen privileged communications out of the set before any adverse question is asked, so the questions run only against material that could plausibly be produced.
2. Run the questions in an environment where privileged evidence never leaves your control, so nothing you feed it trains anyone's model or waives privilege.
3. Keep the answers, the citations and the production record in the same tool, so a hot document found today can be tracked through privilege, Bates and chain of custody tomorrow.

The first step deserves its own design work, and designing a privilege gate so privileged ESI never reaches review walks through it. The third step is where our own product thinking began. Enterprise platforms give you the defensible spine of Bates, privilege and chain of custody; newer AI entrants give you cited Q&A and chronologies. We built Relevant e-Discovery as the combination, because an adverse answer that cannot be carried into production is only half an answer.

There is a temptation, at this point, to believe the tool does the thinking. It does not. In a 2022 interview on the Lawyers Who Lead podcast, the co-founder of HyperDraft recalled writing a script in 45 minutes, in the first year of practice, that turned a spreadsheet into a disclosure schedule; done by hand, the same task was estimated at 13 or 14 hours. The anecdote is usually told as a story about speed. Read more closely, it is a story about a lawyer who already knew what the schedule had to contain. The machine compressed the labor. It did not supply the judgment.

So too here. The software keeps privileged evidence contained and makes each answer traceable. What it cannot do is decide which questions a hostile reader of the record will ask, and that decision belongs to the lawyer who knows the elements of every claim in the case, which is precisely where the adverse questions have to start.

What should you do before your next corporate witness is prepared?

Write one adverse question for each element of each claim and defense, screen privilege out first, run the questions against your client's collection, and read every cited hit before witness preparation begins.

I expect this step to move from the eve of production to the week of intake. The reason is not virtue; it is price. Once an adverse read costs a fraction of a manual one, the remaining question is whether you prefer to learn your client's worst facts privately or publicly.

The tools will keep improving. The judgment will not move to them. Deciding which questions a hostile reader will ask remains the lawyer's work, and it is, perhaps, the most interesting work in discovery, which is why I would begin with the element you are least comfortable defending and let the record answer it first.