How to Use Claude for Due Diligence (Without Trusting a Word It Says)

Claude can read a hundred-page CIM in a couple of minutes and hand you a clean summary. In an acquisition, that summary is the one thing you cannot afford to trust on its own. The failure mode is not a rough paragraph; it is a confident, wrong number you cannot tell apart from a right one, sitting in the basis you priced a deal on.

So the goal is not to get Claude to summarize. It is to get Claude to read, and to make every factual claim trace back to the document it came from, or cut the claim before it reaches you. Used that way, it genuinely compresses the slowest part of diligence. Used as a summarizer you trust, it is a liability.

The one rule: cite or cut

Every figure traces back to a line in a document, or it does not get used. That single discipline is what separates a tool you can act on from a chatbot that sounds sure of itself. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim does not reach you at all; it becomes a question.

A practical method

Step	What to do	Why
1. Give it the real documents	Attach the CIM, financials, tax return, and key contracts, not a paste of one page	It can only cite what it can see
2. Ask for citations, not prose	Require every claim to quote the source and page, verbatim	Makes the output checkable
3. Ask what it cannot support	Tell it to list claims it could not tie to a document	Surfaces the gaps as questions, not silence
4. Cross-check the figures	Ask it to tie the same number across the CIM, financials, and tax return	Contradictions are where deals move
5. Verify the citations yourself	Spot-check the quotes against the documents	The model can still misquote; trust is earned per claim

The order matters. If you ask for a summary first, you anchor on a confident narrative and spend the rest of the time defending it. If you ask for cited claims and a list of what could not be verified, you start from evidence.

Where Claude is strong, and where it fails

It is strong at reading volume fast, restating dense financial language plainly, drafting a question list, and catching a document that contradicts itself when you ask it to look. It fails, quietly, when you let it fill gaps: it will produce a plausible average contract value, a tidy "no outstanding litigation," or a churn number that sounds right and is sourced from nothing. Those are not bugs in the model; they are what a summarizer does when you do not force it to cite.

If you want the discipline pre-built rather than prompting it each time, there is a free, MIT-licensed pack of twenty single-job agents for reading deals this way, described in 20 Claude agents for M&A diligence, each built to show its work and cite or cut.

When a turnkey tool fits better

Prompting Claude by hand works, and for a quick read it is enough. The friction shows up at volume: re-pasting documents, re-writing the prompt, keeping each deal isolated, and checking citations every time. If you are screening deal after deal, a hosted tool that does the cite-or-cut read on your uploaded documents and returns a source-linked brief can fit how you actually work better than a fresh chat each time. That trade-off is laid out in the best due diligence software and in what diligence automation actually catches. You can see the cited output on a synthetic deal, no login, in the sample brief.

Frequently asked questions

Can I use Claude for due diligence? Yes, for reading and first-pass analysis, as long as you force it to cite every claim to a source document and to list what it could not verify. Do not treat an uncited summary as fact.

Is it safe to paste a CIM into a general AI chat? Treat confidentiality and verification separately. On verification, an uncited summary is not safe to act on; on confidentiality, use a tool that keeps each deal isolated rather than a general consumer chat. A purpose-built workspace handles both.

How do I stop Claude from making up numbers? Require citations, ask it explicitly to list claims it cannot tie to a document, and cross-check the same figure across the CIM, financials, and tax return. Anything it cannot source should become a question, not a stated fact.

Claude for due diligence versus a turnkey tool? Hand-prompting is flexible and free but high-friction at volume. A turnkey diligence tool runs the cite-or-cut discipline on your documents automatically and returns a cited brief, which fits better when you are screening many deals.

See the cite-or-cut discipline running on a real-looking deal at Deal OS.