Can you trust AI for due diligence? Yes, for reading and first-pass analysis, but only if it shows its work: every figure traced to the document it came from, and anything it cannot verify discarded rather than guessed. An AI summary you cannot check is the one thing you should not act on in a deal.
AI's limits are in the headlines again, and it is a fair question to ask before you point a model at a seven-figure decision. The honest answer is not a flat yes or no. It is "yes, under one condition," and that condition is verification.
Why trust means something different in an acquisition
In most jobs, a wrong AI answer is an annoyance you catch later. In diligence it is the basis you priced a deal on. If a model invents an EBITDA figure, rounds a churn number in the seller's favor, or softens a customer-concentration risk into something that sounds fine, the error does not announce itself. By the time you find it, you may have signed the LOI and spent a month and credibility you cannot get back.
So the real question is not "is the AI smart." It is "can I tell its right answers from its wrong ones." That is a verification problem, not an intelligence problem.
Where AI is trustworthy, and where it fails quietly
| Task | Trust level | Why |
|---|---|---|
| Reading a long CIM fast | High | It genuinely compresses hours of reading |
| Restating dense financials plainly | High | Strong at summarizing what is on the page |
| Extracting a figure with a citation | High, if cited | You can check each number against its source |
| Catching a document that contradicts itself | High, if asked | It surfaces both passages when told to look |
| Filling a gap with a plausible number | Dangerous | It will invent an average contract value or a clean litigation history sourced from nothing |
| An uncited summary taken at face value | Dangerous | You cannot tell a right number from a wrong one |
The pattern: AI is trustworthy when its output is checkable, and dangerous when it is not. The failure mode is not a rough paragraph, it is a confident wrong number.
The standard that makes AI output safe to act on: cite or cut
One discipline separates a tool you can trust from a chatbot that sounds sure of itself: every claim traces back to a line in a document, or it does not get used. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim never reaches you; it becomes a question instead.
That is the whole point. You are not trusting the model, you are trusting the citation, which you can open and read yourself. The verification layer, not the summary, is what makes the output safe.
What to demand from any AI diligence tool
- Every figure cited to a source document and page, openable and checkable.
- A visible discard log: the claims it could not verify, surfaced as questions rather than silently dropped.
- Cross-document tie-out: the same number checked across the CIM, the financials, and the tax return.
- Your documents kept isolated and encrypted, not fed into a general consumer chat.
- The judgment left to you. The tool reads and verifies; you decide.
If a tool cannot show you why to trust a given line, treat its output the way you would treat an analyst who refuses to show their work.
The bigger picture
The current debate about where AI needs limits is healthy, and it applies in miniature on every deal: the right amount of trust in AI is exactly as much as you can verify. Used with that discipline, AI does the slow first-pass reading in minutes and hands you the real questions faster. Used as an oracle you believe on faith, it is a liability dressed as a shortcut.
You can see the cite-or-cut discipline on a synthetic deal, no login, in the sample brief, where verified claims show their source, a contradiction shows both sides, and unverifiable claims are discarded in front of you. For the practical method, see how to use Claude for due diligence and what diligence automation actually catches.
Frequently asked questions
Can you trust AI for due diligence? For reading and first-pass analysis, yes, provided every claim is cited to a source you can check and unverifiable claims are discarded. Do not act on an uncited AI summary.
Does AI hallucinate financial figures? It can, especially when asked to fill a gap. The defense is not a better model but a discipline: require a citation for every figure and treat anything unsourced as a question, not a fact.
Is AI accurate enough to replace a quality of earnings report? No. AI replaces the slow first-pass read that decides whether a deal deserves a QoE. The formal QoE still happens on the deal you are closing.
How do I know an AI diligence tool is trustworthy? It shows its work: each claim traces to a document and page, it lists what it could not verify, and it ties the same figure across documents. If you cannot check the output, do not trust it.
See verified, source-cited diligence on a real-looking deal at Deal OS.
Top comments (0)