What eDiscovery Taught Us About Due Diligence at Scale

DecoverAI started in litigation review. Our first users were trial teams and investigators working through productions, privilege calls, and record sets large enough that nobody could read them end to end. That setting is unforgiving in a useful way: every coding decision can be questioned by the other side, and the review has to hold up after the fact.

When transactional lawyers began asking whether the same engine could help with a data room, our first instinct was caution. A deal is not a lawsuit. But the more diligence teams we spoke with, the clearer it became that the hard parts of a large review are the same in both places: a document set too big to read, a deadline that does not move, and a partner who has to stand behind the result. This essay is our attempt to separate the lessons that transfer from the ones that do not.

Seven lessons that carry over

1. Define the questions before you classify

Every well-run discovery review begins with a protocol: what counts as responsive, which issues matter, how privilege will be called. Teams that skip that step and start tagging find out later that two reviewers meant different things by the same label, and the cleanup costs more than the protocol would have.

Diligence has the same failure mode. “Review the commercial contracts” is not a question. “Which customer agreements contain a change-of-control consent right, an assignment restriction, or exclusivity?” is. Writing the deal questions down first, and agreeing on what each classification means, is what lets a bulk pass produce something the deal team can act on. It is also the step our due diligence workflow begins with: load the data room into a matter, then agree on the issues and criteria before anything is classified.

2. Reasoning a lawyer can check, and override

The most important thing we learned in litigation is that a classification without a reason is a liability. When opposing counsel challenges a coding decision, “the model said so” is not an answer. A reviewing attorney needs to see why a call was made, open the underlying document, and reverse the call when the evidence says otherwise.

That requirement does not relax when the reader is a deal partner instead of a judge. If a classification says an agreement has no termination-for-convenience right, the lawyer relying on it should be able to see what the system read and decide for themselves. On our platform, attorneys inspect the reasoning and the underlying source before relying on a result, and override AI calls when the evidence requires it. The defensibility standard we work to in litigation is, at bottom, a standard of explainability, and explainability matters to anyone who signs off on the work.

3. Validate on a sample before you rely on the calls

Discovery practice settled long ago that no review method is perfect, and that the honest response is to measure. Technology-assisted review (TAR) brought validation into the mainstream: draw a sample, have attorneys code it, compare, and report precision and recall. Generative AI review raises the stakes, because it does not improve mid-run the way continuous active learning did. Testing has to come before you scale.

No court will ask for your recall figure in a diligence review. Your client will rely on it anyway. Before a team treats “no change-of-control clause” as a finding, someone should have checked a sample of those calls against the documents. The sample tells you which categories you can trust at scale and which need a closer read. It also gives the supervising lawyer what ABA Formal Opinion 512 asks for under Model Rule 1.1: an appropriate degree of independent verification before relying on AI output.

4. Privilege and confidentiality need a workflow, not a filter

In litigation, privilege review is where careful teams spend disproportionate effort, because a mistake can waive protection. Confidentiality designations carry their own obligations under protective orders.

Data rooms raise related questions in a different shape. A seller may have uploaded material that should never have left legal. A buyer may receive competitively sensitive information that belongs behind a clean-team arrangement. Whether sharing privileged material with a counterparty preserves any common-interest protection varies by jurisdiction, and it is a question for counsel, not software. What a review platform can do is surface sensitive material early. Our due diligence workflow brings privilege and confidentiality analysis into the review so the team can assess sensitive material before deciding how to share it. Model Rule 1.6 makes the duty of confidentiality the lawyer’s; the tooling should make it easier to meet.

5. Families and duplicates change the count

Anyone who has run a production knows that an email and its attachments travel together, and that the same document can appear dozens of times across custodians. Ignore families and you code an attachment without the cover email that explains it. Ignore duplicates and you review the same contract four times, possibly reaching four different answers.

Data rooms have their own version. The master services agreement sits in one folder, its amendments in another, and a side letter that changes the economics in a third. Executed and draft versions sit side by side. A finding on the base agreement that ignores a later amendment is not a finding. Before trusting any count of “contracts with issue X,” the team should know whether it is counting agreements or files, and whether related documents were read together.

6. Keep the record as you go

Defensibility in discovery is judged after the fact, so the thing that gets evaluated is the record of the process, not the process itself. Decisions made for good reasons and never written down look, later, like decisions made carelessly.

Deals have their own after-the-fact moments: a post-closing indemnity claim, a purchase-price dispute, a client asking why a risk was accepted. A diligence review that can show who reviewed what, which calls were overridden, and why an issue was judged immaterial is worth far more in that conversation than a memo written from memory. DecoverAI carries matter-level permissions, role-based access, and audit logging across the platform. In the Word add-in, dismissing a finding records the reason.

7. Plan for the exceptions

Every large review has a remainder: scanned images that read poorly, foreign-language documents, spreadsheets, files the system could not process, and calls where the model and the reviewer disagree. In litigation, an uninventoried exclusion is the gap opposing counsel finds. The discipline is to name the exceptions, route them to people, and close the queue deliberately.

Diligence benefits from the same structure. The last step of our due diligence workflow is resolving exceptions: attorneys verify priority documents, correct classifications, and decide which agreements need a closer contract review. Model Rule 5.3 asks lawyers to supervise nonlawyer assistance, and ABA Formal Opinion 512 applies that duty to AI tools. An exception queue with a named owner is what that supervision looks like in practice.

Where diligence is a different problem

Carrying litigation habits into a deal unchanged would be a mistake. Four differences shape how the review should run.

There is no opposing counsel and no court

Discovery review is built for an adversary. Proportionality under FRCP 26(b)(1), negotiated protocols, and validation metrics all exist because someone may challenge the production. In diligence, nobody will move to compel. That frees the team to stop applying uniform rigor everywhere and concentrate effort on the agreements that drive value and risk. It also removes an external check, so the internal ones (sampling, overrides, a recorded rationale) carry more weight, not less.

The goal is decisions and leverage, not a production

A discovery review ends with a set of documents handed to the other side. A diligence review ends with decisions: whether to proceed, what to price in, what to ask for in the disclosure schedules, which consents to obtain before closing, and which specific indemnities to negotiate. The output is a set of positions, so the review questions should be framed around what the buyer or seller will do with each answer.

Materiality and commercial judgment do the sorting

Responsiveness is close to binary. Materiality is not. An assignment restriction in a small vendor contract and the same clause in the target’s largest customer agreement are the same finding with very different consequences. Classification can find every instance; it cannot tell you which ones matter to this deal, at this price, for this client. That judgment belongs to the lawyers, and the review should be designed to put the right documents in front of them rather than to replace them.

The work ends in Word

Diligence findings do not stay in a review platform. They become positions in a purchase agreement, a disclosure letter, or a renegotiated commercial contract, and that negotiation happens in Microsoft Word. That is why we connected bulk review to the contract on your screen. In the Word add-in you choose the party you represent (or neither, for a balanced review), add deal-specific instructions, and review findings ranked by severity, each quoting the clause and linking to its location. Missing protections are flagged as “not present.” Fixes you approve go in as ordinary tracked changes authored by the signed-in Word user, and nothing changes without your action.

The same engine, a different matter

Our proof of scale comes from litigation, and we present it that way. Gregor Wynne Arney PLLC used DecoverAI on a $35M healthcare fraud and anti-kickback investigation with more than one million documents; the full corpus was searchable in days. In a commercial litigation matter, Clayton Trial Lawyers used DecoverAI to surface warranties, indemnity clauses, and performance standards across hundreds of agreements on the way to a $15.4M jury verdict. Both are litigation results, described in our white collar and commercial litigation case studies. Neither is a transactional outcome.

What they show is that the engine handles volume and contract language under pressure. On our due diligence page we describe how the engine that reviews a million documents in a fraud case can support the review of 5,000 contracts in a data room. That figure is an illustrative deal workload, not a reported customer result. Your team sets the priorities for the deal; the platform helps put attorney attention where it is needed. More on how the pieces fit together is on our transactional overview.

Frequently Asked Questions

The lesson underneath the lessons

Litigation taught us that a review is only as good as the lawyer’s ability to check it. Define the questions, show the reasoning, measure before relying, handle sensitive material deliberately, read related documents together, keep the record, and close the exceptions. None of that depends on a court. It depends on someone being accountable for the result, and in a deal that is always true.

What changes is where the review ends: in judgment about materiality and in a negotiation, not a production. If your next data room is larger than your team’s calendar, book a demo and bring the deal questions with you.

This article is general information about legal technology and review practice, not legal advice for any particular matter or transaction. Questions about privilege, common-interest protection, and the handling of competitively sensitive information depend on the facts and the governing law. Professional obligations relating to AI-assisted review vary by jurisdiction and bar authority.