With AI tools for businesses, data protection comes down to three things: where the data is stored, where the model computes and who can access it. You get these three points committed in the contract, and then choose one of four operating models: US provider (global), US provider with an EU region, EU provider, or self-hosted. What to do once the tool is running is covered in the GDPR checklist for AI in your business.
”GDPR-compliant” is not a product label
No tool is GDPR-compliant on its own, because you remain responsible for the processing. If you use an external cloud AI, the guidance on AI issued on 6 May 2024 by the DSK (Datenschutzkonferenz, the conference of German data protection authorities) says this will often be processing on your behalf (para. 32). Art. 28 GDPR then applies: you may only use providers that offer sufficient guarantees, and the provider may only bring in further sub-processors with your prior authorisation (Art. 28(2)). The guidance is not legally binding. It does show how the supervisory authorities read the matter. (As of 20 Oct 2026)
Three questions before choosing any tool
First: where is the data? This means storage of inputs, outputs, logs and uploaded files.
Second: where is it processed? The model can compute in a different region from where the data is stored. “Servers in the EU” does not answer this question by itself.
Third: who can access it? That includes the provider’s support, operations and abuse monitoring, plus its sub-processors, including those in a third country.
The answers belong in the contract or in documentation the contract refers to.
Three terms often get mixed up here. Self-hosted orchestration, such as n8n on your own server (see n8n vs Zapier vs Make), controls your workflows. But if a workflow calls a language model via API, the inputs still go to the model provider. A self-hosted model is something else: the language model itself runs on hardware you control or rent. And the EU region of a US provider means the model computes in a European data centre, while the provider remains a US company.
The four operating models compared
| Criterion | US provider, standard (global) | US provider with EU region | EU provider / German cloud (Mistral, STACKIT, IONOS) | Self-hosted (rented EU GPU or own hardware) |
|---|---|---|---|---|
| AVV (data processing agreement, Art. 28 GDPR) | yes | yes | yes | Rented server: yes, with the server provider; own hardware: not needed |
| Third-country transfer | yes (DPF/SCC) | reduced depending on commitment, check support/operations access | usually no, check sub-processors | no |
| Model quality | very capable models | very capable models | good (open models) to very good (Mistral) | open models, smaller sizes |
| Running costs | tokens | tokens, sometimes a surcharge | tokens | fixed costs (server) plus operations |
| Effort for an SMB | low | low to medium (contract, settings) | low | high |
| Evidence you should ask for | AVV, sub-processor list, DPF entry | contractual commitment on storage and inference | trust center, AVV, location | deletion concept, access concept, updates |
The DSK classifies open cloud systems as a risk for transfers to third countries, which fall under Chapter V GDPR (para. 17). With EU providers it still pays to look closely. According to its own help article, Mistral hosts data in the EU by default, but also states that data can be temporarily transferred outside the EU depending on the feature. With German clouds such as STACKIT or IONOS, check the processing location and AVV terms case by case as well.
US providers with an EU region: what is actually committed
Everything here depends on the specific offering. For the three below, EU processing is tied to a particular product, to conditions or to a setting. (As of 20 Oct 2026)
OpenAI API, regional processing “Europe”
Inference runs in the region, but only for endpoints and models that support regional processing. The prerequisites include Zero Data Retention or Modified Abuse Monitoring, plus approval by OpenAI; according to the documentation, you clarify eligibility with the sales team. Data residency explicitly does not apply to so-called system data. Without such arrangements, OpenAI retains data for abuse monitoring for up to 30 days by default.
Microsoft Azure (Foundry), Data Zone EU
Deployments of the “Data Zone” type process only within the chosen zone (US, EU or Asia-Pacific). The “Global” type may compute in any Azure region. The EU zone follows the EU Data Boundary, can include EFTA countries, and Microsoft can add regions without prior notice. You choose the type yourself when you create the deployment.
Anthropic: direct API, Amazon Bedrock, Google Vertex AI
The direct API has no EU option. Inference there can only be set to “us” or “global”, and the workspace location only to “us”. According to Anthropic’s regional compliance page, Europe is covered via Amazon Bedrock and Google Vertex AI. On Bedrock, AWS advises not to rely on the geographic prefix of an inference profile alone, but to check the destination regions requests can be routed to.
That leaves the question of whether an EU region prevents access by US authorities, for example under the US CLOUD Act. This is legally contested. Have it assessed case by case if your data is sensitive.
Self-hosted AI for SMBs: hardware, effort, model licences
The DSK writes: “Technically closed systems are therefore preferable from a data protection perspective” (para. 20, our translation). According to para. 16, closed means: processing takes place in a limited, technically closed-off environment, only a narrowly limited group has access, control over input and output data lies with the users, and the system is not designed for the provider to use the data for training. How much hardware self-hosting takes is shown by rented GPU servers, here using Hetzner as an example.
| Server | GPU memory | Price per month | Plus one-off setup fee | Location |
|---|---|---|---|---|
| GEX45 | 24 GB VRAM | €214 | €209 | Helsinki |
| GEX63 | 96 GB | €999 | €999 | Helsinki |
| GEX131 | 96 GB | from €1,499 | €1,499 | Germany |
Provider’s figures, net, as of October 2026. These are a server provider’s prices, not meine-bots prices, and they include neither the purchase price of your own GPUs nor electricity or administration time. No total cost can be derived from them.
Open models fit on servers like these. According to its model card, gpt-oss-20b runs within 16 GB, and gpt-oss-120b fits on a single 80 GB GPU. Mistral-Small-3.2-24B needs about 55 GB of GPU memory in bf16. These three models and Qwen3-32B are licensed under Apache-2.0. Llama 4 is different: its use policy does not grant rights to the multimodal models to individuals domiciled in the EU or companies with their principal place of business there; end users of products are exempt.
Self-hosting doesn’t take any duties off your hands. The record of processing activities under Art. 30 GDPR remains, as do the technical and organisational measures (TOMs) and the deletion concept. If you rent the server, you need an AVV with the server provider. And someone has to install updates and decide who may access the system.
When it isn’t worth it: when nobody in-house looks after the server on an ongoing basis, or when your task needs more model quality than open models in smaller sizes deliver.
And if the Data Privacy Framework collapses?
The EU-US Data Privacy Framework (DPF) is in force. The EU General Court dismissed the action against it (T-553/23, Latombe) on 3 Sep 2025, thereby upholding it. An appeal to the Court of Justice is pending; we found no ruling as of 20 Oct 2026. (As of 20 Oct 2026)
If the DPF fell, the main fallback would be standard contractual clauses (SCC), whose use has to be assessed case by case. So make exit-ability a selection criterion: can you export data, prompts and the knowledge base? Can the model be swapped out in your workflows, or does every workflow depend on one provider? Have you earmarked an EU alternative in case you have to switch?
EU AI Act: what matters when choosing
Art. 4 on AI literacy has applied since 2 Feb 2025. Even in its amended version, providers and deployers of AI systems must take measures to support the development of AI literacy among their staff; they are not obliged to guarantee a specific level of AI literacy. The transparency obligations under Art. 50 have applied since 2 Aug 2026; for systems already on the market before then, a transition period under Art. 50(2) runs until 2 Dec 2026. Art. 50(1) is addressed to providers of AI systems; whether and how it affects you as an SMB using AI depends on your role. These dates reflect the position after the Omnibus Regulation (EU) 2026/1744 of 8 Jul 2026, which has been adopted and entered into force on 27 Jul 2026. (As of 20 Oct 2026)
For choosing a tool, this means: ask the provider how it implements the transparency obligations that apply to it.
Decision guide: five questions to ask
- Which data should go into the AI at all? If a task works without personal data, leave it out. For health or client data, the DSK’s line points towards closed systems, meaning a closed-off environment with a narrowly limited group of users, in which you control the data and the provider does not use it for training (para. 16).
- Does the provider contractually commit to storage and inference in the EU, and what conditions come with that?
- Who has support or operations access, from where, and which sub-processors are on the list?
- Can you get out again with your data if the DPF falls or the provider changes its terms?
- Who runs the system? If nobody has time for updates, the access concept and deletion, self-hosting is off the table.
Sometimes the answer is also not to solve a process with AI at all, for example when a form or a fixed rule serves the same purpose. Have the choice in your specific case checked by your data protection officer or a lawyer.
meine-bots helps you choose a suitable setup through AI consulting, for example for individual AI agents. Which processors meine-bots itself uses is listed in the privacy policy. For an initial assessment of your project, book the free initial call (30 minutes, first assessment); the next step after that is the AI Check, which is paid and credited towards a follow-up project.
This article is general guidance (as of 20 October 2026) for businesses in Germany and is not legal or tax advice.
Sources
- Art. 28 GDPR, Processor (gdpr-info.eu) (retrieved 20 Oct 2026)
- DSK: Orientierungshilfe KI und Datenschutz, version 1.0, 6 May 2024 (German) (retrieved 20 Oct 2026)
- EU General Court, press release on T-553/23 Latombe, 3 Sep 2025 (retrieved 20 Oct 2026)
- WilmerHale: European Court of Justice to Review Challenge to EU-US Data Privacy Framework (retrieved 20 Oct 2026)
- OpenAI: Your data (API documentation) (retrieved 20 Oct 2026)
- Microsoft Learn: deployment types in Azure AI Foundry (retrieved 20 Oct 2026)
- Anthropic: Data residency (retrieved 20 Oct 2026)
- Anthropic: Regional compliance (retrieved 20 Oct 2026)
- AWS: Supported Regions and models for inference profiles (Amazon Bedrock) (retrieved 20 Oct 2026)
- Mistral AI Help Center: where data is stored (retrieved 20 Oct 2026)
- Hetzner: GPU servers (retrieved 20 Oct 2026)
- Model cards on Hugging Face: gpt-oss-120b, gpt-oss-20b, Mistral-Small-3.2-24B, Qwen3-32B (retrieved 20 Oct 2026)
- Meta: Llama 4 Acceptable Use Policy (retrieved 20 Oct 2026)
- European Commission, AI Act Service Desk: implementation timeline (retrieved 20 Oct 2026)
- buzer.de: Art. 4 KI-VO (German) (retrieved 20 Oct 2026)
- IHK Lüneburg-Wolfsburg: Digital omnibus regulation on AI has entered into force (German) (retrieved 20 Oct 2026)