How to run document due diligence with AI without losing control

TLDR: Document due diligence is AI review’s extreme case: hundreds or thousands of documents, a short deadline and a report with consequences. The layered method makes it manageable: the written scope (what gets sought and at what threshold), the data room inventory (where the gap is a finding) questioned through NotebookLM, which answers only from your documents and opens the passage behind every answer, the analysis with Paxton extracting data and raising flags so lawyers deep-review the critical and sample the rest, and the report answering the mandate’s question with method and limits declared. Gamma presents it. House rule: AI extracts, the lawyer weighs and whoever signs answers.

The section’s notice: this guide describes the documentary working method. A real due diligence’s legal scope, conclusions and report are lawyers’ work, and nothing here is legal advice.

1. Define the scope before opening the first folder

Due diligence without a written scope drowns: the data room is infinite if nobody knows what is sought. The mandate gets written before opening anything: which decision the report feeds (the acquisition, the financing, the investment), which areas it covers (corporate, contracts, employment, litigation, property, per the mandate), which materiality threshold applies (the minor-value contract does not deserve the same treatment as the one conditioning the deal) and which deadline and format the deliverable has.

That scope document is also the report’s defense: what stayed out stayed out by mandate, not by oversight, and that difference gets written on day one.

The checklist per deal type (what always gets reviewed in a company acquisition, what in a financing) is the firm’s reusable asset, maintained like the contract checklists.

2. Inventory and classify the data room with help

The first work on the data room is not reading: it is knowing what exists. The classified inventory (which documents exist, of which type, from which dates, in which languages) gets built with the tool’s help and produces the second most important finding of any due diligence: what is missing. The documentary gap (contracts mentioned but absent, years without accounts, interrupted minutes) is information heading to the report and to the request list for the other side.

The mechanics: automatic classification by type gets sampled for review, the inventory gets crossed against the scope’s checklist (what should exist and does not) and the missing-documentation requests go out early, because they arrive late when discovered in the final week.

The sector’s data rules apply cubed: the data room enters only tools with validated guarantees (Paxton operates with the category’s certifications), with the deal’s confidentiality agreements in front and the notebooks created under the firm’s corporate account, never the reviewer’s personal one.

3. Analyze in layers: AI extracts and the lawyer weighs

Mass analysis works in three layers with clear roles. Machine extraction: Paxton (from $499 a month) processes the volume extracting the systematic data (parties, dates, terms, renewals, the clauses the checklist demands: change of control, exclusivity, termination) and raising flags on the unusual, with every datum linked to its document and page.

The deep human review: critical documents (by materiality or by flag) get read whole by the lawyer, because weighing what conditions the deal does not get delegated. And sampling the rest: a fraction of non-critical documents re-reviewed by hand measures the extraction’s quality, and its error rate decides whether the sample suffices or must widen.

This is how a finding travels from the data room to the report:

The tool’s citation orients and the document rules, exactly as in legal research.

The same circuit applied to finding case law and doctrine lives in the legal research guide.

4. Write the report that answers the mandate’s question

The due diligence report gets judged on whether it answers the mandate’s question: findings ordered by materiality (what conditions the deal first, the relevant next, the minor in an annex), each with its source document referenced, documentary gaps declared as such, and the method explained (what was deep-reviewed, what by extraction with sampling, which thresholds applied), because the report that explains its method is the defensible one.

The drafting comes from the same circuit (verified findings turned into report prose) for the review and signature of whoever answers for it, and Gamma (free credits, and from $10 a month for the paid plan) builds the executive presentation when the client wants the meeting besides the document.

The complete method turns the unapproachable into the manageable: the machine swallows the volume, the team supplies the judgment and the report ships on time with its limits stated. The rest of the operation lives in AI for legal and law firms.

Common mistakes

Starting to read without a scope. The data room without a written question swallows the team’s hours in documents feeding no decision: the written mandate is the first deliverable.

Ignoring what is missing. The documentary gap is among due diligence’s most valuable findings: the checklist-crossed inventory uncovers it in the first week, not the last.

Reporting from extraction without verifying. The datum heading to the report gets checked in its document: the machine’s error rate exists and the signed report cannot inherit it.

The report without a declared method. What was deep-reviewed and what sampled gets written: the report silent about its method is indefensible when something surfaces later.

Frequently asked questions

How much does AI speed up a due diligence?

The inventory and extraction phase drops from weeks to days in mid-size deals: the deep review of the critical keeps its time, which is where the judgment lives. The total deadline usually shortens notably, and coverage rises.

What fraction of documents gets hand-reviewed?

The critical, whole, always: by materiality and by flag. Of the rest, the sample the error rate justifies: the exact fraction gets decided by the deal’s risk and what the initial sample returns.

Does it work for multilingual data rooms?

Serious tools process several languages and the inventory works: critical documents in languages the team does not command demand professional translation of their key clauses before weighing.

Should the client know AI was used?

Method transparency is gaining ground and the report declaring it defends better: the final call is the firm’s within its ethics frame, and the report’s responsibility does not change with the answer.