A diligence request goes out as a checklist, a questionnaire and a data room, and what comes back is uneven. Some answers are complete. Some are technically responsive and still vague. Some files are stale, mislabeled or missing the schedule that gives them context. Before an attorney can judge what matters, somebody has to line up each answer with the documents that are supposed to support it.
An AI-integrated application can take on much of that lining-up, provided every note it produces points back to its source and nothing it drafts leaves the firm without a lawyer's approval.
Where the discrepancies sit
Take three illustrative cases from an acquisition review. A respondent uploads an insurance certificate but leaves the policy-limit question blank. The lease schedule and the rent roll disagree on how many leases are active. A seller answers "no pending disputes" while a correspondence folder in the same data room refers to an unresolved claim. Each item looks acceptable inside its own folder, and the conflict only shows when the answer and the document are read together.
The application can read the request item, the written answer and the attached files side by side and write a review note when they disagree. The note earns its place by naming what triggered it: the questionnaire answer, the PDF page or extracted field, and the checklist item it affects. Without that trail, an associate has to verify the note from scratch, and the tool has added work.
Who sends the follow-up question?
The model can draft a clarification in plain language. Sending it is a lawyer's decision. Drafts sit in a queue attached to the request item, where an associate can approve, edit, reject or reassign them, and only approved questions go back to the respondent through the portal. The queue records who changed the wording, when the question went out and whether the matter team accepted the answer.
Be wary of a "confidence" column on that screen. A score the model reports about its own output is not a calibrated probability, and it can be mistaken for one. The checkable facts are more useful: whether the document parsed, which fields were extracted, and which passage the note quotes. A file that failed to parse belongs in an exception list that someone reviews, where it cannot drop silently out of the count.
A replaced document reopens notes already reviewed
Respondents revise. A seller replaces a rent roll partway through review, or edits a questionnaire answer after an associate has cleared the note that depended on it. If the application overwrites the old file and regenerates its notes, the record of what the attorney actually reviewed is gone, and a "cleared" status now sits on a document nobody has read.
Follow one illustrative note through. An associate clears a note confirming that the rent roll and lease schedule agree. The respondent then uploads a new rent roll. The application keeps the original file, keeps the cleared note linked to that original, marks the note for re-review, and shows the associate both versions. The new file gets its own extraction and its own notes. The earlier clearance stays in the audit trail as a review of the earlier file, with the reviewer's name and date.
This is a data-model decision to make before the first matter goes live. Review notes need to reference a specific document version, and document storage needs to keep prior versions. If notes only reference a file name, the history of what was reviewed against which version cannot be rebuilt later.
Client confidentiality and the AI layer
ABA Formal Opinion 512, issued in July 2024, addresses lawyers' use of generative AI. For tools that learn from what users enter, it concludes that the client's informed consent is required before information relating to the representation goes in, and that boilerplate consent language in an engagement letter is not adequate. The opinion interprets the ABA Model Rules; the rules and ethics opinions of the firm's own jurisdiction govern what it must do.
That turns the model provider's terms into a design input. Before matter documents reach a model, the firm needs to know whether the provider retains inputs or trains on them. A commercial diligence product may already provide the controls described next, and the same question applies to the model provider behind it. The application needs matter-level permissions that decide which documents the AI layer can read, and a marker on every draft separating client-facing follow-up requests from internal attorney notes, so that an internal note cannot enter the approval queue as a question for the respondent.