Blog · Published 2026-10-18
Connecting a data room to an AI model: what to check first.
Before you connect a data room to an AI model, find out where the documents are sent and hosted, whether they are used for training, how long they are kept and how you delete them, who can see them, and what your NDA with the seller allows. A connector is only a way of moving the documents; the risk is in the answers to those questions.
MCP in plain words
MCP, the Model Context Protocol, is described on its own site as "an open-source standard for connecting AI applications to external systems" (modelcontextprotocol.io). It compares itself to a USB-C port: a standard plug between an AI application and the places its data or tools live.
For a data room this means an AI application could be given access to files in some system without a custom integration. The plug is standard; what flows through it, and where it ends up, is decided by the application and the service on each side. So the safety questions do not change.
Questions to ask before connecting anything
| Question | Why it matters | Where to find the answer |
|---|---|---|
| Where are the documents sent and hosted? | Determines which jurisdictions and subprocessors touch confidential deal material. | The vendor's security page, data processing agreement and list of subprocessors. |
| Are documents used to train models? | Training use would put seller information into a system you do not control. | Terms and data processing pages. Look for an explicit statement, not silence. |
| How long is data kept, and can you delete it? | A deal that dies should not leave its documents behind. | Retention and deletion sections; test the deletion yourself. |
| Who can see the documents? | Includes the vendor's staff, your team, and any other tenant. | Access control and isolation descriptions; ask whether access is logged. |
| What can the connection do? | Read-only access is a smaller risk than access that can change or send files. | The permissions shown when you authorize the connection. |
| Does your NDA allow it? | Many NDAs restrict sharing the seller's information with third parties. | Read your own NDA, then ask the seller or their broker if it is unclear. |
How Deal OS answers the same questions
Deal OS works from documents you upload into a workspace of your own. Its pages state that the workspace is hosted in the EU (France), documents are encrypted at rest, each workspace is isolated from others, documents are never used for training, retention is set by you and documents can be deleted on request, and a mutual NDA can be signed in the workspace before you upload anything. Read the information security page and the data processing agreement rather than relying on this summary.
To see what the output looks like before you upload anything, open the sandbox; for how a single deal is priced, see pricing.
Worked example from the sample brief
What a reviewer gets back once documents are read, from the sample brief (a synthetic deal):
- A claim contradicted by the document's own table. The CIM says on page 8 that no customer is more than 15% of revenue. Page 9 lists the top account (Northgate Property Group) at $1,366,000 against $6.21M of revenue, which the brief computes as 22.0%. The brief shows both passages word for word, each with its page, and marks the page 8 statement as false on the CIM's own numbers.
- A headline figure that two documents state differently. Trailing twelve month revenue is $6.21M on page 6 of the CIM and $5.95M as gross receipts on the tax return. The brief shows the $260K gap side by side and says to reconcile it before applying any multiple.
- A definition that changes what a number means. Page 7 says about 78% of revenue is recurring. Page 22 defines recurring as all revenue from accounts with an active inspection agreement, including repair and call-out work. The brief flags the two passages as inconsistent and asks for a revenue cut that separates fixed inspection billings from variable work.
- Add-backs sorted by whether they are likely to survive. The seller presents adjusted EBITDA of $1.42M. The brief itemizes the add-backs and rates each one; $90K of them are rated questionable or unlikely, which gives $1.33M that holds up on the documents. The arithmetic is computed, and the ratings are a first judgment for a human to confirm.
- A statement the documents do not support, thrown out. While drafting, the model proposed that the company has no outstanding litigation. No passage says so, so the verifier discarded it and the brief lists it in a verification log, with a request for a 10-year litigation history instead.
The example is Sentinel Fire & Safety, Inc., a synthetic company built to show the product. Nothing in it describes a real business, and the same output shape applies to a real data room. See the whole thing at /sample-brief.
Keep reading
See it on a deal.
Open a real brief on a synthetic deal without an account, or check how a single deal is priced.
