July 27, 2026 · Diligence, AI
AI diligence: what can actually be verified, and what can't
By Emily Buckley, founder of SamCIO
Most conversations about AI diligence tools for funds skip the only question that matters: which parts of diligence can a model actually verify, and which parts is it just restating back to you with more confidence than the evidence supports. Getting that line wrong is expensive in a specific way. A tool that verifies nothing but sounds certain doesn't save you work, it moves the error earlier in the process where it's harder to catch.
The honest answer is that AI diligence tools for funds are good at a narrow, valuable band of work: extracting the load-bearing claims out of a deck and checking each one against the public record. They are not good at the judgment calls that decide the deal. Knowing which is which lets you delegate the first category aggressively and keep the second where it belongs.
What a model can genuinely verify
A deck makes claims. Some of them are checkable against sources that exist outside the deck: market size figures, named customers, published funding history, regulatory status, competitor claims, founder background, whether a partnership that got a logo on slide 9 was ever announced anywhere. These are research tasks with a right answer, and they are exactly the tasks that get skipped when an analyst has four decks to get through before Thursday.
This is where automation earns its place. In SamCIO, fact-checking runs inline with the memo: Sam pulls the load-bearing claims out of the deck, runs each one through a web search plus a verification pass, and returns a verdict per claim with citations. The verdicts are deliberately four-valued, not two: supported, mixed, contradicted, or unverifiable. That fourth bucket is the important one. A claim nobody can source is a different animal from a claim that's wrong, and collapsing them into "false" produces a memo that's confidently incorrect in the other direction.
The output isn't a score. It's a list of claims with citations attached, which is a thing a partner can argue with. That matters more than it sounds like it should.
What a model cannot verify, no matter how good the prompt
Everything that depends on the future, on private information, or on your fund specifically.
Whether the founder can execute. A model can tell you what this person did before. It cannot tell you whether they will hold a team together through an eighteen-month slog. Track record correlates with outcome, it doesn't determine it, and the correlation is weak enough at the seed stage that treating a model's read on founder quality as evidence is a mistake.
Whether the market timing is right. Every plausible-sounding market timing argument is available in text, on both sides, for every market. A model trained on that text can produce a fluent case for either direction. Fluency here is not information.
Anything private. Real revenue quality, churn behind the headline number, how the last board meeting actually went, whether the lead investor is wobbling. None of this is on the public web. It comes from reference calls, from the data room, and from asking the founder a direct question and watching how they handle it.
Whether it fits your fund. This is the one funds most often assume the tool has covered. A model does not know your reserve strategy, your LP base's tolerance for a concentrated position, or the fact that you already own something adjacent. It knows those only to the extent you have written them down and given them to it, which is a different claim entirely.
Design the process around the split
Once the line is clear, the workflow follows.
Run verification first, before the human read. The point is to arrive at the deck already knowing which three claims don't hold up, so the analyst's attention goes to the interesting parts rather than to confirming a market size number that was fine all along.
Treat unverifiable as a question, not a finding. A claim that can't be sourced belongs on the founder call agenda. Some of the best diligence questions come from that bucket, because they're specific and the founder can tell you're paying attention.
Keep the citations attached to the memo. A verdict without a source is an opinion with better formatting. When a memo says a market size claim is contradicted, the next person to read it should be able to click through and judge the source themselves. If your process strips citations to make the memo shorter, you have built a system that generates unfalsifiable assertions, which is worse than not checking.
Fail open, not closed. Fact-checking depends on external search, and external search sometimes fails. A diligence system that blocks the memo when verification is unavailable trains people to route around it. The correct behavior is to produce the memo anyway and mark the section as not run, so the gap is visible instead of silent.
Consistency is the real return
The underrated benefit of automating this band of work isn't speed, it's that deal 40 gets the same treatment as deal 4. Manual claim-checking degrades under load in a predictable pattern: it's thorough in January, cursory by June, and completely skipped on the deal everyone already likes. That last case is the expensive one, because enthusiasm is exactly when verification has the most to offer.
A system that runs the same checks on every deal removes the correlation between how much you like a company and how hard you look at it. That's not a small thing. Most diligence failures I have seen were not from missing information; the information was available and nobody went looking for it, because the deal felt good.
What to ask a vendor
Three questions separate real capability from a wrapper:
Does it cite sources per claim, or produce a summary? A summary you cannot audit is not diligence.
Does it distinguish contradicted from unverifiable? If not, it's optimizing for a clean-looking output over an accurate one.
What happens on a claim it can't check? "It says so" is the right answer. Silence is not.
Anything that answers those three well is doing real work. Anything that answers them vaguely is generating text.
The part that stays yours
Verification is a floor, not a decision. The best case for an AI diligence layer is that it clears the ground so the judgment happens on solid footing: the claims are checked, the sources are attached, the open questions are listed, and the partner meeting is about the actual disagreement rather than about who read the deck most recently.
If you want to see how the claim-level checks and citations sit inside a memo, the diligence page walks through the flow. And if you're still setting up the structure the checks feed into, how to write an IC memo covers the document itself.