"Are we even allowed to use ChatGPT without breaking the GDPR?" — short answer: yes, but it depends heavily on how. The difference between clean and risky often comes down to a single setting almost nobody checks. Here's the honest, practical classification — no panic, no legalese.
The GDPR applies the moment personal data is involved
As long as you use AI for general work — drafting a marketing text, structuring an idea — data protection isn't a concern. It gets sensitive the moment personal data wanders into the chat window: customer names, email addresses, job applications, sick notes, contract details. Then the same rules apply as for any other data processing — legal basis, purpose limitation, data minimization. Most violations in small businesses don't happen out of bad intent, but because nobody defined that line.
The one lever that decides almost everything: free vs. business account
The most important thing first, because it makes the biggest difference:
- Free & personal accounts (ChatGPT Free/Plus & co.) may process your inputs further — including to train the models — and you get no data processing agreement. For personal data, that makes them effectively off-limits.
- Business, Enterprise and API access do not train on your data by default, and the provider signs a Data Processing Agreement (DPA) with you.
So simply switching from a personal to a business account moves you from the red zone to the green one — provided you sort out the rest.
Three things to settle with the provider
1. The DPA. The moment a provider processes personal data on your behalf, a data processing agreement is mandatory. With the business tiers of the major providers it's available — you just have to actively sign it.
2. The US data transfer. Most AI services process in the US. That's currently covered by the EU-US Data Privacy Framework, which the EU General Court upheld in September 2025. But: the decision is under appeal at the Court of Justice — and its two predecessor agreements were both struck down. So don't assume this path stays open forever.
3. The storage location. EU data residency now exists — but it's not on automatically. Default processing often still runs via the US; you have to actively configure the EU storage location. This is the setting that's missed most often.
For sensitive data: put in less — or bring the AI to the data
Two robust principles that work independently of any agreement:
Data minimization. Anonymize or pseudonymize before you type anything in. Often the AI doesn't need the real name at all — "Customer A" is plenty.
Run AI where the data lives. For genuinely sensitive data, the cleanest path is not to transfer it at all: a model that runs in the EU or on your own premises. Small, specialized models (SLMs) make that economical too — and the data-protection question solves itself, because nothing leaves your building.
What you don't need to do
The most common overreaction: a blanket AI ban "for data-protection reasons." That's just as wrong as carelessly pouring everything in. For the bulk of everyday work — anything without personal data — you need no special setup. Match the measure to the data: non-critical → public tools with a short policy; personal → business account with a DPA; highly sensitive → EU or your own models. More effort than necessary only costs you speed.
Starting pragmatically
The easiest first step is a clear, short AI policy: what may go in, what may not, which account is allowed. You can generate one in minutes. And because data protection and the AI regulation interlock, it's worth a parallel look at the EU AI Act for small businesses.
For sensitive data, we show on Custom AI models how AI can run under your control — or talk to us directly if you want to know what the clean path is in your case. (This article is for orientation and is not legal advice.)