Five questions to ask any "AI data room" before you upload a thing

• Security
• AI

You can't un-share a confidential document. Once a file leaves your control, no setting reverses it. So the questions worth asking about an "AI data room" aren't about features — they're about where your documents go the moment AI touches them. Ask these five before you upload anything.

1. Which model is actually reading my documents?

"AI-powered" tells you nothing. A specific, named model running on infrastructure the vendor controls is a verifiable claim; "we use AI" is a slogan. If a provider can't name the model and the service behind it, assume the answer is whatever's cheapest this quarter — and that it can change without telling you.

Our answer: a named foundation-model service inside our own cloud account — Amazon Bedrock — handles generation, retrieval, and embeddings. No mystery model, no undisclosed third-party API.

2. Where do my documents physically go?

This is the question vendors most often dodge. Plenty of "AI" features work by shipping your raw files to an outside model provider's API. For a confidential deal, that's the whole document crossing a boundary you don't control and can't audit.

Our answer: documents never leave our cloud boundary. Text extraction, OCR, embedding, semantic search, and answer generation all run inside a single account and region. Your files are not forwarded to any external AI service.

3. Will my documents train someone else's model?

If your content trains a shared model, fragments of your deal can resurface in a stranger's answer. "We don't train on your data" should be a property of the architecture, not a checkbox buried in a settings page.

Our answer: the managed inference service we use does not use your inputs to train its models, and we run no fine-tuning on customer documents. Nothing you upload becomes training data.

4. How long is the AI's copy retained — and can you delete it?

AI features create derived data: text chunks, embeddings, indexes. Ask who holds that derived copy and whether deleting a document deletes it too.

Our answer: the only derived data is chunks and embeddings stored in our own database, under our control. Delete a file and its chunks are hard-deleted with it — there's no second copy sitting in an outside vector store you can never reach.

5. Is the AI permission-aware?

An AI that ignores your permission model is a leak with a search box. If every viewer's question runs against the entire corpus, the AI will happily quote a document that viewer was never allowed to open.

Our answer: retrieval is scoped to the caller's actual access and fails closed — no permission, no results. Deleted and archived material is excluded, and retrieved text is fenced so a malicious document can't hijack the AI's instructions and exfiltrate the rest.

Notice that none of these are feature questions. They're about custody, control, and the ability to verify. A data room earns the word "secure" by answering them plainly — and by being willing to show you the answer in its own code. Ask the five questions. Insist on real answers.