How to stop employees pasting sensitive data into ChatGPT

The paste takes two seconds: a client name, a contract clause, an IBAN, a salary table, dropped into the chat box because the assistant writes a good first draft. Once sent, the text sits on a third-party server. You cannot call it back, you may not know how long it is kept or what it trains, and under GDPR the company carries the responsibility. If you are the one accountable for confidentiality, the question is not whether employees use AI assistants — they already do — but what leaves the browser when they do.

What usually gets tried, and where it stops

Banning the tools. Employees who find the assistant useful keep using it, on personal accounts, on their own devices, outside every policy. The ban turns a manageable risk into an invisible one.

An acceptable-use policy. Worth having, and it does nothing at the moment of pasting. Nobody opens the policy mid-task; the paragraph that forbids client names is not in the room when the client name is in the clipboard.

Enterprise DLP platforms. Real products, priced and sized for large enterprises: procurement cycles, deployment projects, and in many cases traffic routed through the vendor's cloud — which means the data now leaves the company for a second vendor. For a mid-sized company without a six-month project appetite, this is usually where the effort dies.

The option that fits the actual moment

The risk happens in the browser, at the paste. So the control belongs there too. A browser extension can read the text an employee is about to send to an AI assistant, replace the sensitive parts with markers like [NAME_1] or [IBAN_1], and put the real values back into the answer before the employee reads it. The assistant still writes a useful reply — it just never learns the client's name or the account number.

The important design question is where that replacement happens. If the extension sends prompt text to a server for analysis, you have traded one vendor for another. Look for something that runs entirely on the device: detection and replacement in the browser, no prompt text leaving it, ever. A CISO should be able to verify the claim with a network tab, not a sales call.

A short checklist before you deploy anything

• Does anything leave the browser besides licence checks? Get the answer in writing, then check it yourself.
• Does the employee keep a usable answer, or do they just see redacted noise? A tool people disable next Tuesday protects nobody.
• Can you see adoption — how many items were masked, per seat — without seeing any content?
• Does it cover the assistants your teams actually use: ChatGPT, Claude, Gemini, Copilot, Le Chat?

This is exactly what we built AntispyIA for: an extension that masks names, codes and figures before the assistant sees them, runs 100% locally in the browser, and gives you proof of what was masked, seat by seat, from 14 EUR per seat per month. You can watch a sample masking run on our home page before deciding anything.

AntispyIA is built and run end to end by AI agents on NanoCorp, which is how a small team can keep guides like this one current.