What happens when staff paste customer data into an AI chat
· By Bee Flow · 3 min read
People paste customer mails into ChatGPT because it saves time. A ban does not stop that. A check on the way out does, and it can run on your own server.

Somebody in your company has a difficult customer mail to answer. They copy it into ChatGPT, ask for a polite reply and paste the result into Outlook. It took two minutes and the reply is fine. The mail also contained a name, an address, a phone number and an IBAN. Those are now on a server of a company you probably have no data processing agreement with.
Nobody does this to cause harm. It is simply quicker, and that is why a rule in the staff handbook does not stop it.
Why a ban does not work
Many companies have a policy that says "do not enter personal data into AI tools." Everyone signs it. The trouble is that personal data does not look special while you are working. A customer mail is just a customer mail. And nobody checks every paragraph for a licence plate or a date of birth before pressing paste.
A rule that depends on people noticing something they do not notice needs help from software.
A check on the way out
Bee Flow has a filter called Privacy Shield. It reads each message before it goes to the AI model and looks for 21 kinds of data, among them names, addresses, phone numbers, IBANs, BSNs, licence plates, medical terms and API keys.
It works in two steps. Data with a fixed format is found with patterns and a checksum. An IBAN has to pass the same mod-97 test your bank uses, and a BSN has to pass the elfproef. For data without a fixed format, such as a name or an address, a small language model reads the context. That model also prevents silly mistakes: "factuurnummer 0123456789" is not a phone number, and it understands that.
Both steps run on your own server. The check itself sends nothing to anyone.
Block or replace
When something is found there are two options. By default the message is held back and the person sees which kind of data was in it. They remove it and send again. This is in every licence, including the free one.
The other option replaces each value with a placeholder before the message leaves. The model then receives something like "[person_1] at [address_1] has not paid invoice 2291" and writes its answer with those placeholders. Bee Flow puts the real values back before the person sees the reply. The model never had them. This option is in the higher licences; the plans page shows which.
If the detection service is down, chat messages are held back. We made it that way on purpose. Quietly passing on unchecked text is the worse failure.
What it does not do
Detection is not flawless. Most personal data is found, some is not. With a phone number in a very short message without context, for example, the filter is less sure. You can set it stricter, but then it also reacts more often to things that are harmless. So you still need agreements with your team. The filter is there for the moments people forget them.
The filter also has to be installed. It is a separate service that runs on an ordinary processor, so no graphics card is needed. A self-hosted installation without that service checks nothing.
Putting the check in place
In Bee Flow the check is part of the chat itself, so nobody has to remember an extra step before pressing send. Blocking is in every licence, including the free one. The privacy and security page shows how the filter fits with encryption and the rest of the setup, private AI chat covers the chat itself, and Bee Flow vs ChatGPT Business compares the two. To see it hold back a made-up customer mail, plan a 20-minute call.
This post belongs to a series on AI chat and client data. The starting point is is ChatGPT safe to use at work?, and is Microsoft 365 Copilot GDPR compliant? covers Microsoft's assistant.