Deep dive
How AI detects conflicting information in due diligence
Two documents disagreeing about the same fact is one of the easiest things to miss in a manual review — and one of the most systematic things to catch with automated cross-referencing.
01What does a "conflict" actually look like?
A conflict is when two or more documents in the same data room state different facts about the same thing. It's different from a gap (missing evidence) or an error (a single wrong figure) — a conflict specifically means the room itself contains a contradiction, and it's not obvious which source is correct without further investigation.
02What do real examples look like?
| Source A | Source B | The conflict |
|---|---|---|
| Management presentation | Audited financial statements | Different revenue figures for the same fiscal year |
| Customer contract | Sales pipeline spreadsheet | Different contract renewal dates for the same customer |
| Debt schedule | Loan agreement | Different outstanding principal balance |
| Org chart | Payroll records | Different total headcount |
| Board minutes | Cap table | Different share counts following an option grant |
Any one of these, found late or not at all, can materially change a deal's valuation or terms.
03How does automated detection actually work?
The mechanism builds on the same retrieval-and-citation approach used for question answering: every fact extracted from a document is tied to its specific source. When the system encounters multiple documents making claims about the same entity — the same customer, the same fiscal period, the same loan — it compares those claims. If they match, nothing needs to be flagged. If they don't, the system surfaces both sources and the discrepancy, rather than silently choosing one answer to present as fact.
This is only as good as the grounding underneath it — a system that isn't citing sources for every claim has nothing reliable to compare in the first place. That's why conflict detection and citation-grounded answering are really the same underlying capability, applied in two directions: one answers a specific question, the other compares answers across documents. See how AI reviews an M&A data room for how the citation step feeds into this.
04Why does manual review tend to miss these?
Not because reviewers aren't careful — because the two conflicting documents are often reviewed by different people, at different times, or in different categories entirely. A financial analyst reviewing the debt schedule and a legal reviewer reading the loan agreement may never directly compare the principal balance each document states, because neither document is "theirs" to cross-check against the other. Systematic, room-wide comparison catches exactly this kind of cross-category conflict that falls between two reviewers' areas of focus.
05What happens after a conflict is flagged?
Flagging isn't resolving. Once a conflict is surfaced, it goes to the deal team to investigate — sometimes it's a simple explanation (one document is an earlier draft), sometimes it's a genuine red flag worth raising with the target. Vaultrix marks conflicting items explicitly within request-list tracking so they don't get closed out as "evidence found" by mistake, and a person makes the final call on what the conflict means. See AI due diligence for how gap and conflict detection fits into the full review workflow.
06Frequently asked questions
Does the AI decide which source is correct?
No. It surfaces both sources and the discrepancy between them; deciding which is correct, or investigating further, is left to the deal team.
Can conflicts be found across different file types?
Yes — a conflict can be between a spreadsheet and a PDF, or a slide deck and a contract, as long as both are ingested and indexed.
Are all conflicts equally serious?
No. Some are explained by document versioning or timing differences; others point to a genuine discrepancy worth escalating. That judgment call belongs to the reviewing team.