AI due diligence in a data room: check every answer before IC
AI can read a data room in minutes. How to check its answers on change of control, debt, customers and add-backs before they reach the investment committee.
Key takeaways
- AI answers from a data room read well whether or not they are right. Check the ones the IC memo relies on.
- An answer without a page and passage is unverified, however confident it sounds.
- The riskiest answer is "none": it is only as good as the documents that were searched.
- Keep the checked answers with their sources. They are the diligence record.
A data room can hold thousands of documents: material contracts, financial statements, a quality of earnings report, debt documents, leases, employment agreements and disclosure schedules. AI assistants can now answer questions across all of it in minutes, which changes how fast a deal team can form a view.
It does not change who is accountable for the view. Whatever goes into the investment committee memo still has to be true, and an AI answer is a lead until someone has seen the passage behind it.
Which answers to check first
Not every answer needs the same scrutiny. Start with the ones that change price, structure or the decision to proceed.
| Question | Where the answer should come from | What usually goes wrong |
|---|---|---|
| Change of control and assignment | Material contracts, with their amendments | Consent rights missed in an amendment or a side letter |
| Customer concentration | Revenue schedules and customer contracts | Group entities counted as separate customers |
| EBITDA and add-backs | Quality of earnings report, financial statements | Management figures quoted instead of diligence-adjusted ones |
| Debt, liens and guarantees | Credit agreement, security documents, notes to the accounts | An undrawn facility or a guarantee left out of the total |
| Litigation and claims | Disclosure schedules, legal letters | Matters listed in a schedule but not in the summary |
| Key people | Employment and incentive agreements | Change of control payments or accelerated vesting missed |
When an answer has no citation
Some AI tools cite only some of their answers, or only some of the files in a room. Others cite the document but not the page. Either way, the rule is simple: an answer without a page and a passage is unverified. Treat it as a question to put back to the documents, not as a finding.
Watch for files the tool may not have read at all. Scanned PDFs without a text layer, spreadsheets, and documents uploaded after the room was indexed are common gaps. If a file was not read, nothing in it can be cited, and nothing in it can contradict the answer either.
The riskiest answer is "none"
"None of the customer contracts has a change of control clause" is a claim about every contract in the room. It is only as good as the list of documents that were actually searched. Before an absence goes into the memo, confirm three things:
- Coverage. Which documents were searched, and does that match the index of the room?
- Wording. Consent rights can sit under assignment, termination or a defined term such as "Change in Control", not only under a heading with the obvious name.
- Amendments. A clause added in a later amendment or side letter is easy to miss if only the base agreement was read.
A check routine before IC
- Fix the question list. Ask the same questions of every document in a set, so gaps show up as gaps.
- Require a page and passage for every answer. No source, no finding.
- Read the passage, not only the answer. Confirm it is the operative clause, in the governing version of the document.
- Recompute every number. Concentration percentages, leverage and add-back totals, each with its inputs and formula.
- Test every "none" for coverage. List the documents behind it.
- Mark each answer. Supported, contradicted, or not found in the documents.
- Keep the record. The checked answers and their sources belong in the deal file, next to the memo.
The same discipline applies to the memo itself once it is drafted. See how to tick and tie a credit memo.
Running a diligence sweep with CiteLLM
In ChatGPT or Claude, attach up to 20 data-room documents to one CiteLLM review and ask the same questions of each. You get one table, and every cell opens at its clause with the passage highlighted:
- Supported answers show the passage they rely on.
- Calculated answers show the formula, with each input highlighted.
- Contradicted answers show the passage that disagrees.
- Not in the document is reported as such, so an absence is visible rather than assumed.
For a whole room, CiteLLM Ask answers across thousands of documents with the page behind each answer, and runs in our cloud or in your own infrastructure.
Frequently asked questions
Can AI replace a legal due diligence review?
No. It speeds up finding and summarising the relevant clauses. The conclusions still need a person who has read the passages they rest on.
How many documents can one CiteLLM review hold?
Up to 20 PDFs, each in its own tab. For a full data room, CiteLLM Ask searches across thousands of documents.
What should I do with an answer that has no source?
Treat it as unverified. Ask the question again against the specific documents, and keep it out of the memo until a passage supports it.