The Diligence Memo Nobody Wrote

2026-08-15

Half of the world's dealmakers now run an AI system before they take the first call. A survey of 1,000 senior dealmakers across 27 countries, published by Datasite and FT Longitude this July, found that regular AI use during due diligence delivers the highest reported return of any stage in the deal process — and that 62 percent of respondents now consider a purely human review indefensible on a complex transaction. Seventy-one percent think firms that skip this step will struggle to compete within five years.

Nobody is going to argue investors should stop typing company names into ChatGPT, Gemini, or Perplexity before a meeting. It is faster than the twenty minutes of searching it replaced, and it surfaces things a human reviewer might miss. The question worth asking is what, exactly, gets read — and what happens to that reading once the meeting is over.

A study published this month traced the answer directly. Researchers ran a fixed set of diligence-style questions — funding raised, headcount, customer traction, financial risk — against twenty privately held software companies across six AI systems: ChatGPT, Gemini, Perplexity, Claude, Copilot, and Google's AI Mode. That produced 1,440 individual answers, hand-checked claim by claim against the true figures.

The pattern was consistent enough to be a structure, not noise. Every answer filled the same rough set of slots — identity, funding, team, traction, risk, and a closing verdict — regardless of which system produced it. And each slot leaned on a narrow, predictable set of sources: 58 percent of every citation in the panel traced back to just three source types — the company's own site, a funding database, and a review platform.

That concentration is where the trouble starts. Of the answers that cited a specific funding figure, 31 percent named a round the company had already been superseded by, with a median staleness of 22 months among companies that had since raised again. Headcount estimates ran a median of 41 percent low. Ask two different AI systems the identical funding question and you got conflicting numbers 44 percent of the time — one system pulling the current round, the other still working from an older snapshot. And the risk section — the part of the answer most likely to shape an investment committee's first impression — drew only 12 percent of its material from anything the company itself published. A third came from review sites. A fifth came from forum threads. In several cases, the stated evidence for "financial risk" was nothing more than the age of the last funding announcement, read as a signal of trouble rather than a fact about time.

Perhaps the most consequential finding: two-thirds of answers closed with a verdict-shaped sentence — "worth pursuing if," "a concern would be" — and that sentence almost never carried a citation at all. It reads as a conclusion. It is really a tone summary of everything above it, generated fresh each time, unanchored to any single source a reader could check.

None of this means the systems are useless, or that the people running them are careless. It means the profile an investor, acquirer, or board nominating committee reads before the first conversation is frequently a reassembly of records the company stopped tending months or years ago, presented with a confidence that outpaces its accuracy. And the analyst reading it rarely thinks to ask where any individual line came from, because the answer doesn't invite the question.

Here is the part that should concern anyone advising a company through a raise, a sale, or a board appointment: the company being described almost never sees this. It doesn't know which figure was quoted, which review was surfaced, or which risk line colored the conversation that followed. By the time a term sheet arrives with a lower valuation than expected, or a board seat quietly goes to someone else, the AI-generated profile that shaped the room is usually long gone — most systems retain no durable, retrievable record of an individual answer once the session that produced it ends. Even the company that corrects its own facts is fixing the input, not recovering the output that already did its work in someone else's head.

That is a different problem than getting the facts right, and it is the one nobody in the deal room is set up to solve. Correcting a stale figure on your own site is necessary. It does nothing for the meeting that already happened on the strength of the old one, and it leaves no way to establish, after the fact, what a given AI system told a given decision-maker, when, and from where.

Due diligence has always run on the assumption that a claim can be traced back to its source and dated. That assumption held when the sources were filings, references, and data rooms. It gets harder to sustain when the first and most influential document in the process is a nine-second answer that no one saved, attributed to nobody, and gone by the next query.