The Data Layer in AI Due Diligence (Part 1 of 3)

Most debates about AI in due diligence are about which model is smarter. That question matters, but it skips an earlier one. A model can only reason over the evidence in front of it, and in diligence the expensive failure is rarely an obviously wrong answer. More often it is a plausible answer built on a missing competitor, a market estimate from three years ago, or a management assumption nobody tested.

Those gaps don't announce themselves. The memo reads well, the citations look fine, and the conclusion holds up until someone at the investment committee asks about the regional player the research never found.

So the quality of AI-driven research depends on what the system finds, what it leaves out, what it challenges and what it can't establish. That makes evidence discovery and validation part of the analysis itself. At Binocs we treat it as its own layer, and it runs before the AI writes a word.

Collecting sources is the easy part

More sources don't automatically make better research. A hundred pages that repeat the same press release add nothing one of them didn't. What a deal team needs is an evidence base that covers the relevant market, competitors, hypotheses and counterarguments, and is honest about where it falls short.

We find it useful to separate three activities that usually get lumped together. Source collection is finding material relevant to the question. Evidence discovery is noticing what's missing, following new leads and testing whether the evidence supports the conclusion that's starting to form. Evidence validation is establishing what can be substantiated, what stays uncertain, and where credible sources disagree.

General-purpose AI tools have become good at the first. The second and third are where diligence is won or lost, and they have to be engineered deliberately.

Four things diligence-grade research has to get right

Coverage. Has the research found the evidence that matters? The most visible companies and reports are a starting point. Smaller competitors, niche trade surveys and customer comments on earnings calls rarely rank well in search, and they are often what changes the answer.

Discovery. What does the evidence itself say to look at next? A competitor mentioned once in a trade journal may point to a segment nobody scoped. Good research expands based on what it finds, and it also knows when to stop.

Validation. Can the material claims be supported? Sources differ in relevance, credibility and recency. When two credible sources disagree, that disagreement belongs in front of the reader, because quietly picking one hides exactly the uncertainty a deal team needs to see.

Consistency. Would a second run, or a second analyst, reach a comparable result? Judgment will always vary between people. The method shouldn't.

How we build it: Source Monte Carlo

Source Monte Carlo is Binocs' implementation of the evidence layer. It builds the evidence base the way an experienced analyst would with unlimited time, in four steps, and the AI starts its analysis only once that base is ready.

Discover. The system searches across public sources, proprietary databases, deal-room documents and alternative signals like podcasts and interviews. For competitive work it looks for companies by what they do, the markets they serve and the customers they target, so it isn't limited to names someone already knows.

Expand. It looks at what it has found, spots the gaps and searches again. The competitor from the trade journal becomes a new line of inquiry. The loop keeps running until new searches stop adding anything useful, which gives the research a defined stopping point.

Rank. Every source is scored on relevance, credibility and recency, so a recent government dataset outweighs a sponsored whitepaper from 2021. Each report type also has its own rules for what counts as proof. In competitive trends, a pattern has to show up in at least two companies before we call it a trend. A single company's move might be the start of something or might be a one-off, so we surface it as a watch item and leave the call to the reader.

Triangulate. Material claims are checked against independent sources. Where they agree, confidence goes up. Where they disagree, the disagreement is flagged, since conflicting data often points to different definitions, a market in transition, or an area that needs more diligence. Uncertainty usually can't be eliminated, so the aim is to make it visible.

Some rules exist purely to keep the output honest. Every link in an industry trends draft is checked, and if more than half fail, the draft is thrown out and rebuilt. Financial facts have to come from filings, and figures we can't confirm are marked undisclosed and never estimated. M&A searches use a 24-month window and widen to 36 only when nothing qualifies, so a quiet sector doesn't get padded with irrelevant deals.

What an independent benchmark showed

Claims about research quality are easy to make, so we asked Caliper Lab to evaluate Binocs independently against frontier LLMs.

On source breadth, Binocs drew on 833 distinct domains, against 64 to 96 for the frontier models. It surfaced 1,062 distinct entities, and 84% of its cited sources were reachable when checked.

We'd be skeptical of anyone who treated breadth as proof of quality on its own, and we don't. What these numbers do suggest is that the evidence base reaches well past the first page of search results, and that the citations behind it can be opened and checked by the person who has to defend them.

Where it changes the decision

Take a target that says it's gaining share in a fragmented market. Its management presentation supports the story, and so do the most visible industry reports. A quick research pass finds those, plus the obvious competitors, and concludes the narrative holds.

A wider evidence base turns up a few regional competitors growing faster than the target, along with customer commentary pointing to rising price pressure. None of that disproves management's case. It does change what the deal team asks next:

• Is growth coming from real share gains, or from a market expanding underneath everyone?

• Are pricing dynamics shifting, and is the target holding share by discounting?

• Are the newer competitors stronger in particular regions or customer segments?

• What evidence would we need before underwriting the growth assumption?

That is where the evidence layer earns its place. The call stays with the deal team, and they get to make it on a more complete, better-tested base.

What a diligence-grade evidence system should demonstrate

AI research tools vary a lot, and the better general-purpose ones can run iterative searches and check citations. So the useful comparison is against a standard. If you're evaluating any AI research system for diligence, Binocs included, these are the questions we'd ask of it.

What deal teams get

• Deeper insights. A wider search finds what a quick one misses: the smaller competitor gaining share in one region, the segment growing faster than the market, the data point that challenges the management plan.

• Consistent output. Because the rules for searching and ranking are fixed, running the same question again gives a comparable result. You can rerun it, compare it with last month's version, or hand it to a colleague.

• Stronger evidence. Every claim links back to a specific, ranked source. When someone on the IC asks where a number came from, you can show the document and the other sources that back it up.

It also takes the slowest part of early diligence, finding and checking sources, off the associate's plate. That leaves more time for judgment calls.