How to Verify AI-Generated Due Diligence Findings Before the IC Memo

A single model's blind spot is invisible from inside that model's own output. If a model drew a conclusion about customer concentration, competitive position, or legal exposure from a thin or ambiguous source, the sentence describing that conclusion looks exactly the same as one built on solid ground — same fluent tone, same declarative structure. There is no confidence interval printed next to it.

This matters more in M&A specifically because deal materials mix genuinely well-documented facts — audited financials, signed contracts — with judgment calls the model has to infer, like management credibility or competitive durability. A single model doesn't reliably flag which sentence in its own output is the first kind and which is the second.

Running the same underlying question through more than one model changes what you're looking at. Instead of one confident answer, you get a comparison: where do independent models converge on the same read of customer concentration, and where does one model flag a risk another doesn't mention at all? Convergence across models trained on different data is a stronger basis for a finding than any single model's confidence — though it still isn't proof, since overlapping training data can produce agreement without independent verification.

Disagreement is the more useful signal in practice. When one model characterizes a customer relationship as low-risk and another flags contract-renewal timing as a material dependency, that specific split is exactly the thing worth tracing back to the underlying document before the finding goes into a memo — not something to average away into a single blended answer.

Illustrative example: an AI-drafted note on a mid-market target states that customer concentration is "manageable, with the top account representing roughly 15% of revenue." Run through a panel, one model corroborates the 15% figure directly from the disclosed schedule. A second model flags that the schedule combines two entities under common ownership that function as a single buying relationship — pushing effective concentration closer to 30% once combined. Neither model is being dramatic; they're reading the same schedule two different ways. The disagreement is the finding here: it tells the analyst exactly which line needs a second look before "manageable" goes in front of the committee.