How to Keep AI Prompts and Documents Off the Public Cloud

Step 1 — Inventory your AI exposure

Before you can close a leak you have to find it. List every tool your staff paste work into: the obvious chat apps, the browser extensions, the “summarize this” buttons baked into other software. Ask people honestly what they use; most teams find more than they expected. This list is your exposure map, and it’s usually longer than the owner thinks.

Step 2 — Classify what can’t leave

Not all of it is sensitive, so mark the data that legally or contractually can’t go off-site: patient records, payment data, signed contracts, NDA-bound material, source code. Those classes are the ones that must never touch a public model. Everything else you can decide on later — the point is to know your hard line first.

Step 3 — Stand up a local model and point it at your data

Install an in-house assistant on a server you own, then connect it to your document store over the LAN. Now staff get the same paste-a-question, get-an-answer experience, except the question and the documents never leave the network. Open models handle most business tasks well, and there’s no per-seat meter running while people work.

Step 4 — Block the leakers and verify isolation

Once the sanctioned tool is live and faster, policy-block the public ones — you’re removing the reason to leak, not just the access. Then prove it: pull the internet connection and confirm the local assistant still answers. If it works with the cable out, your prompts and documents are genuinely staying in the building.