Back to the blog

Choosing a GDPR-proof alternative to ChatGPT: five questions for the vendor

· By Bee Flow · 4 min read

Every vendor of a business AI chat calls itself secure and European. Five questions show the difference: where conversations are stored, the processing agreement, which model runs where, what happens to personal data and who can log in.

A clipboard with a checklist and a pencil lying across it

On Monday morning the managing director asks the office manager whether the company can just get its own ChatGPT. Everyone uses it anyway, so it might as well be something that fits the GDPR. A week later there are four quotes on the table. All of them call themselves secure, European or GDPR-compliant, and their websites look much alike.

You only see the difference once you ask the vendor a few concrete questions. Here they are, in the order we would ask them. This is a buying checklist, not legal advice. If the question is still whether ChatGPT itself will do, start with is ChatGPT safe to use at work?.

Where are the conversations stored?

A conversation with an AI assistant is stored text. That text soon holds a customer's name, a salary or something from an HR file. So ask in which country the conversations are kept, by which company, and for how long. Ask who sets the retention period, and what happens when an employee deletes a conversation.

In Bee Flow, conversations live in the database of the installation. If you host Bee Flow yourself, they are on your own server. Your organisation sets the retention period. By default a conversation stays until the employee deletes it.

The data processing agreement

When a vendor processes personal data for you, Article 28 of the GDPR asks for a data processing agreement (what goes in one). Two things deserve attention.

The agreement with the chat vendor is separate from the one with the company behind the language model. If the chat service passes your questions on to another company's model, that company is a sub-processor and belongs in the agreement, together with the country where it processes the data.

Also ask whether your conversations are used to train models, and have the answer put in writing.

If you host Bee Flow yourself, there is no extra party between you and your data for the chat itself. You still need an agreement with the provider of the model you choose. Run the model on your own hardware and that one goes too.

Which model, and where does it run?

A European vendor can still send your questions to a model in the United States. So ask which model answers your questions and in which data centre it runs.

In Bee Flow you choose this yourself, for the whole organisation and, if you like, per assistant. There are roughly three options. Large cloud models from Anthropic, OpenAI, Google, Mistral and others. Open models hosted in Europe, at Scaleway in France and the Netherlands or at IONOS in Germany. Or models on your own hardware. You can mix them as well, with a local model for simple work and a cloud model for the hard questions.

What happens to personal data in a prompt?

Even with a good policy, people paste customer mails into the chat. Ask what happens to a name, an IBAN or a BSN in a message: is it held back, replaced by a placeholder, or sent along as it is?

Bee Flow has a filter for this that runs on your own server, called Privacy Shield. How it works and where it falls short is in what happens when staff paste customer data into an AI chat. For this checklist, the point is where the check happens. A filter that runs at an outside service has to send the text out before it can check it.

Who can log in?

An AI chat that can reach your documents is as safe as the way people sign in to it. Ask whether staff can sign in with the account they already have, whether there is two-factor authentication, and whether you can decide per group who sees which assistant and which documents. Ask too how you shut out someone who leaves the company.

Bee Flow supports sign-in with a password, or with a Google, Microsoft or Nextcloud account. Two-factor authentication works with an authenticator app or a security key. Knowledge bases are shared with the whole organisation or with specific groups, and an assistant only answers from sources the person asking may see. Not every sign-in method is in every licence; see the plans page.

How we would go about it

Put these five questions next to every quote. A vendor who cannot answer one of them clearly usually has nothing good in place for it either. After that we would run a one-month trial with one team, with real questions but without the most sensitive files, and only then roll it out further.

Bee Flow's own answers to the five questions are in this piece, and the privacy and security page goes into more detail. What a private ChatGPT for business looks like in Bee Flow is on its own page. The trial month can run on your own server with the free licence, which covers chat, knowledge bases and blocking personal data. A comparison with a well-known service is on Bee Flow vs ChatGPT Business. To set up that first team, plan a 20-minute call.

#privacy #compliance