Why Your Due Diligence Report Is Probably Wrong
We ran our own report through a three-step fact-checking process. Four claims were fabricated. Two were misleading. Here's what we learned — and why it matters for every deal.
AI-generated research is fast, cheap, and getting better every quarter. It's also full of plausible-sounding fabrications that are nearly impossible to catch without a systematic verification process.
We learned this the hard way. Last week, we built a due diligence report on a California FRP manufacturer — the kind of analysis a search fund entrepreneur would use to decide whether to pursue a $3.5M acquisition. The report was thorough: market sizing, competitive landscape, regulatory environment, financial analysis, risk assessment. It read like something you'd pay a consulting firm $15,000 for.
Then we fact-checked it. Here's what we found.
The four fabrications
| Claim in the report | What was wrong | Real answer |
|---|---|---|
| "U.S. FRP market: $15.1 billion" | That's the global market. The U.S. market is ~$2.1B. | Off by 7x. The number was real — just for the wrong geography. |
| "OSHA is tightening styrene limits to ≤0.5 ppm" | Completely fabricated. OSHA's current PEL is 100 ppm. No rulemaking exists for 0.5 ppm. | The model invented a regulation that sounded plausible. |
| "CalEPA Rule 1365: ≤50 ppm VOC" | Wrong units. The limit is ≤50 g/L, not ppm. Different magnitude, different meaning. | The regulation exists. The number was close. The unit was wrong. |
| "Marlite" listed as a major competitor | Marlite is a brand of Molded Fiber Glass, not a company. | The model invented a corporate entity from a product name. |
These aren't edge cases. Every one of these claims would have made it into a client deliverable without a verification step. The styrene claim alone could have led a buyer to budget $1–2M for abatement equipment that isn't required.
The most dangerous error was the one that felt right
The market size error is instructive. The $15.1 billion figure was real — it came from a legitimate Grand View Research report. The problem was that it described the global FRP market, not the U.S. market. The model cited a real source for a claim that source didn't support. This is called citation laundering, and it's the hardest type of hallucination to catch because the number checks out — you just have to read the source carefully enough to notice it's answering a different question.
When we sent a subagent to verify the claim, it confirmed the number existed without catching the geography mismatch. The verification itself needed verification.
What we built instead
After finding these errors, we built a mandatory three-step fact-checking phase into every report we produce:
- Source Audit — Every factual claim gets tagged with its origin. Sourced and verified. Listing-derived. Calculated. Industry benchmark. Or unverifiable — in which case it's removed or qualified with honest language.
- Cross-Reference Verification — High-stakes claims (market size, revenue, multiples, regulations, competitor names) are verified against an independent second source. If sources disagree, the report shows the range — not whichever number fits the narrative.
- Hallucination Pattern Scan — A nine-point checklist targeting the specific ways LLMs fabricate: invented URLs, fake company names, over-precise numbers, internal inconsistencies, temporal mismatches, plausible-sounding statistics, geographic impossibilities, regulatory fabrication, and citation laundering.
The rule: No report leaves our hands until every factual claim has a source tag, high-stakes claims are cross-referenced, and the hallucination scan is clean. We publish the verification report alongside the deliverable.
Why this matters for your deal
If you're an independent sponsor or search fund entrepreneur, you're probably using AI tools in your diligence process. That's smart — the speed advantage is real. But raw AI output is a first draft, not a finished product. The errors aren't random noise; they're systematic, predictable, and concentrated in exactly the places that matter most: regulations, market data, and competitive intelligence.
A PE firm can't tell their LPs they based diligence on a Claude chat. And a search funder who budgets for abatement equipment that isn't required has just overpaid for a business by $1–2 million in their own head.
The solution isn't to avoid AI. It's to build verification into the process — systematically, not as an afterthought. Every claim gets a source. Every source gets a second source. Every number gets checked for internal consistency. And if something can't be verified, the report says so.
That's what we do. It's slower than raw generation. It's more expensive than a ChatGPT prompt. It's also the difference between a report you can stake a deal on and one you can't.
Preflight Research delivers fact-checked due diligence reports for independent sponsors and search fund entrepreneurs. Every report passes a mandatory three-step verification process before delivery. Get a report →