AI M&A due diligence on a 10,000-page data room: all 8 planted deal-breakers found.

By Alan Finney — Founder, 3Dogs Nexus

We built a 100-document acquisition data room — roughly 10,000 pages — and put it to every option a buyer really has: a consumer AI, a due-diligence firm, the new "AI advisory board" tools, and 3Dogs Nexus. Then we ran two tests: would it bluff on a thin room, and could it find risks deliberately buried in a rich one?

Can AI actually do due diligence on a 10,000-page data room?

Yes. We built a data room of 100 documents totalling about 10,000 pages and planted eight specific deal-breakers across it — customer concentration, an environmental liability, an earn-out built to fail, obsolete inventory, inflated EBITDA add-backs, a key-person gap, litigation, and a revenue contradiction between the CIM and the tax filings. The panel read the room in full and surfaced all eight, then returned a decisive renegotiate.

Before analysis, one question kills most tools: can it open the room at all?

A 10,000-page data room is ~1.15 million tokens across 100 separate files. That single fact decides most of the contest before a word of analysis is written.

The tool most people reach for. It hits hard product limits before analysis begins:

• Gemini accepts 10 files per prompt — you have 100.
• ~1.15M tokens exceeds even Gemini Advanced's 1M-token window (~1,500 pages).
• ChatGPT's ~128K context ≈ ~200 pages — about 2% of the room.

The traditional answer — analysts read the room by hand: quality-of-earnings, working capital, legal, tax.

• Thorough, senior, defensible — the gold standard.
• $50,000 – $150,000+ for a deal this size.
• 6 – 12 weeks to a final report.

Its Deep Discovery engine map-reduces all 100 documents into one grounded, cross-referenced brief, then runs an adversarial panel over it.

• Reads every one of the 100 files — all ~10,000 pages.
• 1,302 model calls across 12 distinct models.
• A verdict in 28 minutes, not six weeks.

First we gave it a room with nothing in it. It refused to bluff.

Our first run used a deliberately empty 10,000-page room — volume with no real financials behind it. A single confident AI will happily manufacture an authoritative acquisition memo from thin air. 3Dogs read every page, detected the room was substantively empty, capped its confidence at Low (73%), and printed the dissent — three of eleven analysts arguing against proceeding on an incomplete room:

That's the point. The value wasn't a confident yes — it was an honest “not yet, and here's exactly what to demand first.” (601 calls · 13 models · 12m 03s.)

Then we buried 8 deal-killers in the room. It found all 8.

Our second run used a rich data room with eight specific, deal-changing risks hidden inside — each in one place, surrounded by thousands of pages of ordinary paperwork. Because we built it, the test is falsifiable: here's the answer key, and how 3Dogs surfaced each one.

What about the new "AI advisory board" tools?

A wave of multi-agent "decision" products has launched. They're genuinely useful — for the jobs they're built for. But this use case, an unstructured 10,000-page document room, exposes what each is actually architected to do.

Competitor capabilities and pricing reflect each vendor's published materials (July 2026) and are summarized for this specific document-heavy use case — not a knock on tools built for other jobs. The distinction here is architectural: persona boards and warehouse analysts aren't designed to read an unstructured 10,000-page due-diligence room.

Reading 10,000 pages is the easy part. Knowing what's buried — and what's missing — is the job.

The risk that sinks a deal is never on page one. It's the footnote in file 48, the aging report in file 23, the gap between the CIM and the tax return. A fast single-model read skims the top and sounds confident. 3Dogs read every page, cross-referenced the documents against each other, refused to bluff when the room was thin, and found all eight needles when they were there — with the dissent still on the record.

Questions this case answers directly

Plain answers on acquisition red flags and diligence failure, from a controlled test with planted problems and a known answer key. Every figure below comes from the delivered report for this case. These are clearly-labeled panel estimates from a multi-model adversarial analysis — not investment, legal or professional advice.

What are the most common red flags in acquisitions?

The eight planted in this controlled test are the eight that recur in real deals: customer concentration (a single account at 41% of revenue), undisclosed environmental liability, an earn-out structured so it cannot pay out, obsolete inventory carried at cost, EBITDA inflated by add-backs, key-person risk with no non-compete, pending litigation, and a revenue contradiction between the CIM and the tax returns. 3Dogs Nexus found all eight reading 10,000 pages blind — 1,302 model calls, 12 models, 28 minutes.

When should you walk away from an acquisition deal?

When the thing you cannot verify is the thing the price depends on. In this test the panel did not recommend walking — it recommended renegotiating with escrow conditions, which is the more common correct answer. The distinction matters: most red flags are price problems, not deal-breakers. The genuine walk-aways are the ones that survive escrow — unquantifiable environmental exposure, or a revenue figure that two documents describe differently and management cannot reconcile.

What are the biggest mistakes in M&A due diligence?

Reading at uniform depth, and reading for confirmation. A data room is designed to be exhausting; the material problems sit where fatigue peaks. The second failure is subtler — a team that has decided to do the deal reads to justify it. That is precisely what an adversarial panel is built to prevent: two seats, a devil's advocate and a risk officer, are permanent and cannot be reassigned, so somebody is always structurally paid to argue no.

How to spot fake revenue in a company you want to buy?

Cross-reference the documents that were written for different audiences. The planted contradiction in this test was exactly that — a revenue figure in the confidential information memorandum that did not match the tax filings. Marketing documents flatter; tax documents do not. Where they disagree, the tax filing is the one prepared under penalty. The panel surfaced this without being told to look for it.

See the actual 3Dogs reports

Both deliberated briefs — the empty-room integrity run and the 8/8 buried-risk run. Every claim, the confidence, the vote, the preserved dissent.

Act 2 — the 8/8 buried-risk run

Case 2026-0084 · RENEGOTIATE · 1,302 calls · 12 models · 28m 04s.

Act 1 — the empty-room integrity run

Case 2026-0083 · refused to bluff, Low confidence · 601 calls · 13 models · 12m 03s.

Questions this case answers

Can AI find planted red flags in acquisition due diligence documents?

In our test it found 8 of 8. We wrote an answer key before the run and scored the output against it afterwards, so the result is verifiable rather than impressionistic.

What happens when the data room is actually empty of substance?

We tested that too, on a corpus that turned out to be pure boilerplate. Rather than inventing findings, the system detected the room was substantively empty, capped its confidence at Low, and printed three dissenting models. Refusing to fabricate is the behaviour that matters most in diligence.

Is this buy-side due diligence AI suitable for distressed note purchases?

The same document-heavy method applies: read the full file, surface what contradicts the seller’s narrative, and price the risk explicitly. The recommendation always states its assumptions and what would change it.

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

People also ask

Related decision case studies

• Can AI Re-Litigate Military Decisions?

  Four of the most-taught command decisions in military history, given to a multi-model AI panel with none of the hindsight, then compare

• Should You Pay a Ransomware Demand?

  Two Las Vegas casino operators faced the same attacker in 2023. One paid, one refused. We put the decision to a 13-model panel and it s

• We Let ChatGPT Run a Decision Case and Review Us

  OpenAI's ChatGPT drove an entire 3Dogs Nexus engagement as the client, answered both clarification rounds, then wrote a review we publi

• What Does the Deepest Multi-Model AI Analysis Do?

  Our deepest analysis mode run on a real capital decision, with every step observable: two full ensembles, multiple research passes, and