Data loss prevention (DLP) for AI tools: four approaches compared
DLP stands for "data loss prevention". With AI tools the question is simple: how does a company stop customer data, bank details or passwords from ending up in an AI chat? Such leaks are rarely intentional. They happen when someone quickly pastes a text into ChatGPT, Copilot or Gemini.
Why classic protection often misses this
- The text leaves the browser encrypted. Whatever is pasted into an AI chat goes directly to the provider via HTTPS. A firewall only sees that an AI site was opened, not what was sent.
- Personal accounts: according to an analysis by Cyberhaven (2024), most ChatGPT use at work ran through personal rather than company accounts. The company's contracts and settings do not apply there.
- The input itself is the risk: once a text has been sent, it cannot be taken back.
Four approaches compared
1. Block AI sites
Easy to do, but use often moves to personal phones, where the company sees nothing at all. More in the guide Shadow AI in small businesses.
2. Business accounts with a contract
The AI providers' business plans govern how input is used; OpenAI, for example, does not train on input from its business plans by default. This is an important foundation. It does not stop anyone from entering data that should never go into an AI service at all, such as data under professional secrecy.
3. Network or endpoint DLP
Full DLP systems inspect traffic through a proxy or with an agent on the device. They cover many channels but need licences, setup and maintenance. Network solutions often route traffic through an additional service. For larger companies with their own security team this makes sense; for many small businesses it is too heavy.
4. A browser extension where the text is typed
An extension checks the text where it is created: in the input field of the AI site, before it is sent. AImori does exactly that, locally in the browser, without servers and without user accounts:
- detects IBANs, national ID numbers, card numbers, passwords, keys and others of its 242 data types (list), plus names and addresses with a local AI model,
- shows what it found before sending and offers Redact, Cancel or a deliberate Send anyway,
- can steer users from AI sites that are not approved to the allowed tool: a notice or block page with a button, set centrally via policy,
- can be deployed and configured centrally via Intune or Group Policy (for companies).
We state the limits openly: AImori checks input only on ChatGPT, Claude, Gemini, Copilot and Microsoft 365 Copilot in the browser, not in desktop or mobile apps. It reads images with local text recognition and checks tables and text files for customer lists and hidden instructions; it does not read PDF or Word files. The AI model does not catch every name. It is an aid against mistakes and does not replace a full DLP system. What AImori does not check.
What fits which company
| Company | Sensible combination |
|---|---|
| Small company without in-house IT | One-page rules, business account, browser extension |
| SME with an IT partner | Policy, training, extension via Intune or Group Policy |
| Larger company with a security team | Full DLP system, plus a hint right at the input field |
In every case, clear rules and short training belong to it: AI policy template · AI literacy under AI Act Art. 4.
AImori is free. First see what would be detected in your texts, without installing anything: free test · Install AImori
Read next: Can employees put customer data into ChatGPT? · IBANs, ID numbers & co.
Sources
Information as of 17 October 2026. This guide is not legal or security advice.
- Cyberhaven: Shadow AI: how employees are leading the charge in AI adoption (2024)
- OpenAI: How your data is used to improve model performance
- Swiss Federal Act on Data Protection (FADP), SR 235.1, Art. 8 (data security)