Chat is a useful way to begin an investigation. It is a poor place to finish one.
A diligence conclusion usually depends on more than a single answer. The team needs to know which sources were in scope, which passage supports the claim, whether the answer conflicts with another document, who reviewed it, and where the approved conclusion was used.
That creates a gap between answer generation and professional work.
What a transaction team actually needs
A defensible diligence record should preserve five things:
- Scope: the files, systems, and time period considered.
- Evidence: the exact source context behind each material claim.
- Process: the questions, analyses, and workflow steps that produced the result.
- Review: the person who accepted, changed, or rejected the claim.
- Use: the finding, matrix, memo, or deliverable that ultimately included it.
A citation alone is not enough if it disappears when the answer is copied into a spreadsheet. A workflow alone is not enough if nobody can inspect what its agents did. A polished memo is not enough if the reviewer must rebuild its provenance by hand.
Treat chat as one surface in a larger system
The stronger pattern is a shared workspace where chat, structured analysis, workflow runs, findings, and deliverables operate on the same evidence base.
A useful answer can become a matrix question. A matrix result can become a finding. A reviewed finding can become a claim in a deliverable. The source connection should survive each handoff.
That is the shift from “chat with documents” to an evidence workroom. The interface still benefits from conversational AI, but the system around it is designed for collaboration, review, and reuse.
A simple evaluation question
When assessing an AI tool for diligence, ask:
If this answer becomes a material statement in our committee memo, can a reviewer move from the statement back through the decision and analysis to the original source?
If the answer requires searching through old chats and opening files manually, the team has gained speed at the cost of a new verification burden. If the chain is intact, AI becomes part of the operating record rather than a separate drafting tool.
Follow one claim through the room
Take customer concentration, a familiar diligence question. The first answer might come from an ARR workbook. The contract review adds renewal dates and termination rights. A management call explains a disputed account classification. A later workbook changes the revenue period.
The useful output is not one polished paragraph. It is a claim that can accumulate evidence, expose disagreement, and change state:
- the initial answer is drafted with a workbook-range citation;
- contract evidence adds commercial context;
- the call creates a conflict for the deal team to resolve;
- a reviewer accepts the calculation but changes the materiality;
- the approved claim enters the risk register and IC memo;
- the revised workbook marks the claim for another review.
A chat transcript can contain pieces of this history. It cannot reliably operate it. The team needs assignments, states, relationships, and downstream impact.
Evaluate the handoffs, not only the answer
Accuracy tests are necessary, but a production pilot should also observe what happens after a useful answer appears.
Ask five practical questions:
- Can the analyst promote the answer without copying and pasting it?
- Does the source context remain available to the reviewer?
- Can the reviewer edit the claim without erasing the original output?
- Does the approved version flow into the intended work product?
- If a supporting source changes, can the system identify the affected claim and output?
These tests reveal hidden work. A tool may save ten minutes of searching and create twenty minutes of source reconstruction, formatting, and review coordination.
Use chat where ambiguity is useful
Conversational work is strongest at the beginning of an investigation. Analysts can probe unfamiliar material, reframe questions, and follow surprising evidence. Structure should increase as the work becomes consequential.
A useful progression is:
Explore in chat. Ask broad questions and inspect evidence.
Compare in a matrix. Apply stable questions across files, companies, or contracts.
Record a finding. Give a material claim an owner, evidence, and review state.
Compose a deliverable. Use only the reviewed version in the memo, redline, or report.
The point is not to eliminate chat. It is to stop asking chat history to perform jobs it was never designed to do.
How we approach this topic
This field note is based on the workflow and product-design questions we encounter while building Underlying. It is educational, not legal, investment, or security advice. Product examples describe design patterns unless explicitly stated as generally available.
Reviewed by Underlying Product Team.