How to stop staff pasting sensitive data into AI chats

Why the usual way fails

A written AI policy helps, but people forget it the moment they are busy, and nobody reads a policy while drafting a prompt. Blocking every AI site pushes staff to their phones or personal laptops, where you see even less. Network tools inspect traffic in transit, which often means decrypting connections, and anyone working off the VPN or from home may not pass through them at all. By the time a sensitive prompt is spotted in a log, it has already been sent. And a block list goes stale fast, because new AI tools appear every few weeks and the one your team uses tomorrow may not be on it.

How Milvago AI helps

Milvago AI works on the workstation, where the prompt is typed, instead of in the network. A browser extension and a lightweight agent apply your rules on each device, and sensitive data such as email addresses or bank details is masked in the prompt before it is sent to a supported AI service. For each service you choose to observe, block, redirect people to the AI you approved, or mask locally. The rules are signed, they expire, and they keep applying when a laptop is offline, while the usage record stays on your own server. It also records the names of files sent, never the files themselves, and the extension spots visits to more than 160 AI platforms by their address.

Step by step

1. Install the self-hosted server and console on infrastructure you control.
2. Roll out the browser extension and the agent to the workstations you want to cover.
3. Check the usage map to see which AI services people already reach.
4. Set a posture for each service, such as mask for one tool and redirect for another.
5. Publish the signed policy and review the activity log as events come in.

Tips

• Start in observe mode for a few weeks. Seeing which tools people actually use tells you where a mask or redirect rule will matter, and where a block would only push them elsewhere.
• Redirect rather than block when you already pay for an approved AI tool. People keep their habit of asking an assistant, and the data goes to the service you trust.
• Decide early whether you need prompt text at all. Recording facts about usage, not the content, is often enough for an audit and is easier to defend to staff.
• Tell your teams what the tool does and does not see before you roll it out. Clear rules and an approved option usually cut risky pasting more than surprise monitoring does.

Questions