AI and GDPR: Using ChatGPT Without Exposing Personal Data

The risk isn't AI. It's where your data goes. A practical guide to using ChatGPT and other AI without breaching the GDPR — and how we build systems where your data stays yours.

Hugo Matthaey

Hugo Matthaey

Jul 18, 2026

Personal-data records flowing to an AI model through a protective shield that keeps them in a safe zone

Someone on your team pastes a client list into ChatGPT to clean it up. Someone else drops in a CV to summarize it, or a contract, or the transcript of a sensitive meeting. It's fast, it's useful, and it happens every day. It's also, quite often, a GDPR breach hiding in plain sight.

The problem isn't AI. It's where the data goes the moment you hit send.

Why this is a GDPR problem

When you paste personal or confidential data into a consumer AI tool, that data leaves your control: it's sent to a processor that is usually outside the EU, may be retained, and — depending on the plan — may be used to improve the model. Under the GDPR, that raises the obvious questions: what's your legal basis, is there a data-processing agreement, where is the data stored, and can you actually delete it on request? For most casual ChatGPT use, the honest answer is "none of the above."

This isn't specific to ChatGPT. It's true of any AI tool where you don't control the model or the infrastructure it runs on.

"So we ban AI?" No.

Banning AI just pushes it into the shadows — people use their personal accounts instead, which is worse. The goal isn't to avoid AI; it's to use it without exposing data you're responsible for. In practice, there are four levers, from lightest to strongest:

  1. Enterprise tiers with the right contract. Business/API tiers from the major providers offer no-training-by-default, a DPA, and (increasingly) EU data residency. For lower-sensitivity data, this is often enough — but read the terms, and confirm where the data actually sits.
  2. Anonymize before you send. Strip or pseudonymize personal identifiers before anything reaches the model. Good hygiene, but easy to get wrong at scale.
  3. Self-hosted, open-source models. Run the model on your own infrastructure or on EU-hosted GPUs so the data never leaves your perimeter. This is the strongest answer when the data is genuinely sensitive — regulated, confidential, or contractually locked down.
  4. Explainability and a human in the loop. Whatever the setup, you should be able to see what the AI did and keep a person in control of anything that matters. No black box.

How we build it: your data stays yours

This is the core of how we work. When the data is sensitive, we build AI systems that run on infrastructure you control — self-hosted open-source models, EU hosting, no training on your data — wrapped in pipelines that are explainable and quality-checked, with human review where it counts. Same capability you'd get from a consumer tool, without handing your data to one.

We've written about the concrete version of this for self-hosted AI meeting notes, and about running our own GPU inference in Europe.

The honest version

For low-sensitivity work, an enterprise AI tier with a proper DPA is usually fine — don't over-engineer it. For personal data, regulated data, or anything under an NDA, assume the consumer tools are a liability until proven otherwise, and move that work to a self-hosted or tightly-controlled setup.

If you want AI in your business but your data can't leave your walls, that's exactly what we build. Get in touch and tell us what you're working with.

This is not legal advice; check your specific obligations with your DPO or counsel.