ACG Minnesota · PE Summit · October 20, 2026

# Live: Watch AI Work a Data Room ## This morning: test the claims that kill deals. Now: watch it happen in a room. Note: Callback to Faster Diligence at 11:00. Its line was: AI doesn't make diligence faster by reading faster; it makes diligence better by testing what matters, early. This session is the demo that talk promised. For anyone who missed the morning: 46.6% of broken lower-middle-market LOIs in 2025 died on what diligence found (Axial). The point of AI in the room is to find those things in week one, not month four. Everything you will see runs live: our own agent, built on Claude, reading a Box data room. No Box AI, no data-room vendor's AI. --- ## The room - **Northstar Precision Components**, a fictional middle-market manufacturer - 48 documents in a standard eight-folder diligence index - Contracts, audits, leases, board minutes, insurance, HR - The buyer's question: what in here would change price, structure or the go/no-go? Note: Say it plainly: the company, the people and every document are made up. Customer names too. We planted the kinds of problems a deal team actually loses money on. I won't tell you where they are; watch the agent find them. 48 documents is small, about nine hours of reading by hand. A real room is thousands of documents, which is the point: the method is the same, only the hours change. --- ## What to watch for - A claim tested **across** documents, not read in one - Every finding cites a page. Click it. - Late documents re-read, and old findings resolved - Who can download what - Seller questions: one per finding, not three hundred Note: Each bullet is a point from this morning, now live. Across documents: the management presentation says no customer is over 18% of revenue; the revenue schedule says 31%. Neither document is wrong on its own. Citations: a citation is a pointer to proof, not proof. So we click through to the page in Box, every time. Late documents: sellers drop files mid-process. The agent re-reads them and compares versions. Access: 48% of dealmakers in the SS&C Intralinks and Reuters survey reported an access-control lapse. One is planted here. Questions: AI makes questions free; management's time is not. The question list is short on purpose. Switch to the dashboard and Box now. --- ## Which claim does this thesis depend on, and what in the room contradicts it? This is one example of what can be built around the way your firm works. Note: Close on this morning's last slide: stop asking how fast AI can read the room; ask which claim the thesis depends on and what contradicts it. The offer from this morning stands: backtest a claim-first workflow on a room you already know the answers to. Built in your cloud, owned by you. Take questions. Likely ones: what model (Claude, through the API); where the data goes (the agent runs on this laptop, reads Box through Box's API, and sends document text to Claude through Anthropic's API); how long to build (this one: days, not months).