GDPR and AI: How to run business AI in the EU

You want to use AI to run your business. You’re in the EU. Your legal team has questions. This guide answers them.

The core tension

AI models need data to work. GDPR restricts how you process personal data. These aren’t incompatible. But the default setup of most AI tools violates GDPR by design.

When you type “Follow up with Sarah Chen about the Acme proposal” into ChatGPT, you’ve just sent personal data (Sarah’s name, her business relationship with Acme) to OpenAI’s US servers. That’s a transatlantic data transfer. Under GDPR, you need a legal basis for that.

The three layers to get right

Layer 1: Where your AI runs

This is the most important decision. Your options:

Provider Region GDPR Impact CLOUD Act
Anthropic Direct US Transfer to US. Requires SCCs, DPIA. ⚠️ Yes
Mistral AI EU (based in Paris) EU processing (default). French company as processor. No US parent; transfers to sub-processors outside the EU possible
Local model (LLaMA, Qwen) Your hardware No transfer. You’re the only processor. ✅ None

Recommendation: For maximum AI quality, use Anthropic (Claude) via the direct API and accept the US transfer with SCCs. For EU sovereignty, pick Mistral AI: a French company with its own models and native tool calling, hosted in the EU by default. lynox supports both natively via the LLM-provider setup.

Layer 2: Where your data lives

If you use a cloud AI platform (Lindy, Dust, etc.), your business data lives on their servers. You’re adding another processor to the chain, and another compliance surface.

With self-hosted software like lynox, your data stays on your server. The data flow is:

  1. You → Your server (no third party)
  2. Your server → AI provider (conversation context only)
  3. Your server → lynox.ai (nothing — no telemetry, no analytics)

The AI provider is your only processor. You sign their DPA directly. No middleman.

Layer 3: What data reaches the AI

Even with EU-based processing, you should minimize what reaches the AI model. Best practices:

  • Avoid sending raw API keys or passwords. lynox uses an encrypted vault designed to keep secrets out of conversations.
  • Be deliberate about personal data. The AI doesn’t need passport numbers to write a follow-up email.
  • Use the right model for the task. Quick tasks can use smaller models that process less context.

Article-by-article checklist

GDPR Article What it means How to comply
Art. 5 — Minimization Only process data you need Send conversation context, not bulk exports
Art. 17 — Right to erasure Users can request deletion With self-hosted: delete the SQLite file
Art. 20 — Portability Users can export their data SQLite files are portable by default
Art. 25 — Privacy by design Build privacy into the system Self-hosted + encrypted vault + no telemetry
Art. 28 — Processor agreements Need DPA with processors Sign DPA with your AI provider directly
Art. 44-49 — International transfers Restrictions on non-EU transfers Use Mistral AI (EU processing) or a local model; rely on SCCs for Anthropic (US)

What to tell your DPO

Give them this summary:

  1. Software (lynox): Source-available, runs on your server. Not a service — no DPA needed with lynox.
  2. AI Provider (Anthropic or Mistral AI): Anthropic (US) under SCCs, or Mistral AI, hosted in the EU by default. Mistral may transfer some data to sub-processors outside the EU; Mistral lists them in its Trust Center. Sign the relevant DPA directly. Mistral AI is a French company with no US parent.
  3. Data flow: User → our server → EU-based AI provider. Other third parties only come in if we connect them, such as a web search or an integration, or if our server runs at a hosting provider.
  4. Storage: SQLite on our server. Encrypted backups. No external database.
  5. Telemetry: Zero. The software doesn’t phone home.

Then point them to lynox.ai/trust/eu-compliance for the detailed breakdown.

The bottom line

Running AI in the EU isn’t hard. It’s just about making the right architectural choices upfront:

  1. Self-host the software (removes one processor from the chain)
  2. Pick the right AI provider for your risk profile: Anthropic (US, under SCCs) for top model quality, Mistral AI (France, no US parent), or a local model for zero external transfer
  3. Minimize data sent to the model (encrypted vault for secrets)

lynox was designed for exactly this setup. One command to install, one config change for your preferred provider. Choose between Claude via the Anthropic direct API (best AI quality, US transfer under SCCs), Mistral AI (French, hosted in the EU by default, no US parent), or local models (no external data transfer). Consult your legal team for your specific compliance requirements.