GDPR applies to your prompts, not to the AI
Under GDPR Article 3, the regulation's territorial scope is not triggered simply because a data subject lives in the EU or because software is used there. It generally applies where processing is carried out in the context of an establishment of an organisation in the EU, where an organisation outside the EU offers goods or services to people in the EU, or where it monitors their behaviour there. The precise scope depends on the facts. Under Article 4(1), personal data includes any information relating to an identified or identifiable person. A name, an email address, a phone number, a voice recording, an IP address, even a pseudonymised contract clause that can be linked back to a person.
So when you paste "summarise this email from jan.novak@company.cz" into a chatbot, you are processing Mr Novák's personal data. Whether GDPR applies is determined by the relevant establishment, offering-of-services or monitoring criteria, not by the user's or data subject's location alone. The model's location is also only one part of the analysis.
The responsibility is not automatically left entirely with the person typing. The employer or other organisation usually determines the purposes and means of business processing and is therefore the controller, while a cloud provider may act as a processor. Responsibility is shared according to the actual roles and arrangements; an individual using a service for purely personal purposes may be assessed differently.
The five-point GDPR checklist before you use cloud AI
1. Classify the input, not the tool
There is no "GDPR-compliant tool" — only compliant processing. A privacy-friendly chatbot can still receive data it should never see. Before you press Enter, ask: does this text contain anything that identifies a natural person? If yes, check the lawful basis, purpose limitation, retention period and deletion route, among other governance requirements. These are practical compliance checks, not an absolute test that a reader must personally be able to answer in full before any processing is possible. This is general information, not legal advice.
2. Check where the data actually lands
The EU offers a real choice today, but the details depend on the product, contract and configuration:
- Microsoft Azure OpenAI — Azure's EU Data Boundary can support processing and storage within the EU for eligible services and configurations. Regional processing and zero-data-retention options are product- and workload-dependent, so verify the applicable terms rather than assuming they are available everywhere.
- Google Cloud / Vertex AI — EU regions and data-residency controls are available for supported services, but routing, storage, logging and retention depend on the model, endpoint and configuration. Google states that customer data is not used to train models without permission under the relevant enterprise terms; check the specific service documentation.
- OpenAI directly — EU data-residency options may be available for eligible ChatGPT Enterprise, Edu or API customers, subject to plan, product and configuration. The standard API and consumer products should not be assumed to use EU-only infrastructure; verify the current regional and retention terms. Routing through Azure is a separate arrangement.
- Mistral (France) and Aleph Alpha (Germany) — both offer European-based products and hosting options, but the applicable region, subprocessors, model access, retention and subscription terms vary by product and plan. An EU headquarters or brand does not by itself guarantee that every workload remains in the EU.
- Anthropic — its direct services are available in Europe, but regional processing, retention and residency depend on the particular plan and offering. AWS Bedrock and Google Vertex AI may provide EU-region or geographically routed options for supported Claude models, subject to their own endpoint, subprocessor and contract configuration. Do not assume that an EU endpoint makes every part of the service EU-only.
If a US provider is not certified under the Data Privacy Framework, transfers may need another lawful mechanism such as standard contractual clauses (SCCs) — which can mean more documentation and risk assessment for you, not less. Certification and its scope should be checked for the specific provider and service.
3. Sign a data processing agreement (Article 28)
A consumer chatbot's terms of service are not necessarily a DPA. If you use cloud AI for work and real personal data flows through it, an Article 28 agreement may be required where the vendor is acting as your processor, together with an assessment of its subprocessor list and other contractual terms. "But I only use the free tier" is not a defence the Czech DPA or any other EU supervisory authority will find charming.
4. Switch off training and retention
Different products behave differently:
- API access — OpenAI, Anthropic, Google and Mistral generally state that eligible API data is not used for model training by default, but exclusions, abuse monitoring, logs, retention settings and product terms vary.
- Consumer chat — ChatGPT, Claude and Gemini may use conversations to improve models unless you change the setting or use a plan with different terms. Disable chat history or training use before pasting anything sensitive, and check the current product policy.
- Zero retention — Azure OpenAI and Vertex AI may offer zero-data-retention or equivalent controls for eligible API workloads and configurations. Availability, meaning and exclusions differ, so treat it as a contractual option to verify rather than a universal guarantee.
5. Test the erase path (Articles 15–17)
GDPR gives data subjects the right to access and erasure. So ask yourself: can you actually delete one conversation, one uploaded file, one transcript? With a cloud API you may be able to delete it via API or support ticket, but the exact route and effect on logs, backups and safety records vary. Try to document the process once. A difficult or unclear deletion process is a governance risk, even though the legal assessment depends on the circumstances.
When local AI is the right answer
Local AI is not a magic GDPR shield — but it can solve the three most annoying parts of the checklist: data location, external subprocessors and deletion. Where processing is genuinely confined to your organisation's own systems, no processor DPA is needed for that processing. However, hosting, support, telemetry, backups, software updates and connected services can still involve processors or transfers. When a model runs on your own machine and the whole workflow stays inside your systems, no data leaves your office. Deleting data means deleting the files yourself, including backups and chat logs.
Here I write from operational experience. At ai-jarvis.eu we run production AI services — article pipelines, transcription and TTS — on our own servers, and we benchmark local models in AI Arena on an RTX 5060 Ti 16 GB using Ollama. The practical sweet spot for that card is 7B–14B parameter models at 4-bit quantisation. A 7B model occupies roughly 5–7 GB of VRAM, a 14B model around 9–10 GB, and a 24B model about 13–14 GB. That is enough for summarisation, rewriting, entity extraction, structured data extraction and decent RAG.
A 32B model at Q4 needs roughly 20 GB of VRAM, and a 70B model around 40 GB — simply not realistic on a 16 GB card without heavy offloading that makes the experience painful. So local AI is not "free frontier AI". It is a trade: you give up top-tier coding and reasoning quality for the ability to process sensitive data without external disclosure. For many European businesses, that trade is exactly right.
Local vs cloud under GDPR — the comparison table
| Dimension | US cloud (with DPA) | EU-hosted cloud | Local (your hardware) |
|---|---|---|---|
| Data location | May be US-based or globally routed, depending on provider and configuration | May use EU data centres or EU endpoints, subject to routing and subprocessors | Your device / office, if the workflow is genuinely isolated |
| DPA (Art. 28) | Usually required where the vendor is your processor | Usually required where the vendor is your processor | Not needed for processing confined to your own systems; connected services may differ |
| Training on your inputs | No (eligible API terms), varies (consumer chat) | No (under applicable terms), varies by product and configuration | No external model training by default, unless software or telemetry sends data out |
| Subprocessors | May include many, listed according to the provider's terms | May include several, listed according to the provider's terms | None in a genuinely isolated setup; hosting, support and connected software can add them |
| Deletion | API request / ticket, subject to logs and backups | API request / ticket, subject to logs and backups | Delete files + backups |
| Auditability | Certifications and audit reports vary by provider, service and configuration | Certifications and EU sovereignty programmes vary by provider, service and configuration | More direct control, but your organisation must operate and audit it |
| Upfront cost | €0 for some entry tiers; commercial terms vary | €0 / subscription, depending on product | €400–550 for a 16 GB GPU, depending on model and market |
| Running cost | Pay per token or subscription | Pay per token or subscription, potentially plus VAT | A few €/month electricity at normal office usage, depending on workload and energy price |
| Viable model size | Frontier (100B+), depending on service | Frontier, depending on service | 7B–24B practically |
Cloud API list prices change frequently and vary by region, model, tier and billing terms. Check the current official pricing page for the specific model, pricing tier and region before budgeting; do not rely on figures quoted in older articles. A local GPU, by contrast, costs electricity only once purchased — roughly the equivalent of a few euros per month at normal office usage, depending on workload and energy prices.
The fine print: local AI is not GDPR-free
Running a model locally removes most of the data-transfer problem, but not all obligations:
- Security (Article 32): an unencrypted laptop full of local model files and client data is still a breach waiting to happen. Local AI means you are now the security department.
- AI Act transparency: from 2 August 2025, providers of general-purpose AI models have been subject to the applicable obligations, including publishing training-data summaries under the AI Act (Article 53). Article 50 transparency obligations for certain AI systems have applied since 2 August 2026, with the exact duty depending on the system and use. If you distribute a chatbot or app that uses AI, check whether those obligations apply to you too — including requirements to inform users when they are interacting with a machine.
- Model licensing: open weights are not public domain. Llama, Qwen, Mistral and Gemma all have use restrictions. Licence obligations apply regardless of where the company is based, and they must be checked for each model and version.
- High-risk systems: if you use an AI system to make decisions about people — hiring, credit, medical triage — the AI Act's high-risk obligations do not all begin on one date. For high-risk systems listed in Annex III, including certain employment, education and essential-services use cases, the main obligations have applied since 2 August 2026. For AI systems that are themselves products, or safety components of products, under the product-safety legislation listed in Annex I, the obligations generally apply from 2 August 2027, subject to the system's classification and the applicable transition rules. This applies whether the model runs locally or in the cloud.
The workflow we would actually recommend
- Default to an EU-hosted API (Azure OpenAI in an EU region, Google Vertex AI, or Mistral) for anything containing personal data, after checking the specific endpoint and terms. Get the DPA signed once where required, and keep the subprocessor list in your contract folder.
- Run local models for internal drafts, templates and low-sensitivity work — with Ollama or LM Studio on a 16 GB GPU class machine. You get private, unlimited drafts for the cost of electricity, provided the surrounding software and backups stay local.
- If you use consumer chat, disable training and history — and still pseudonymise. Replace names and emails before pasting; a simple script can do this automatically.
- Review quarterly. Residency options, prices and AI Act deadlines move fast. What was US-only last year may have an EU region now — and what was EU-hosted may have changed its subprocessors.
Does GDPR force me to use a European AI provider?
No. GDPR does not require data to stay in the EU as long as transfers have a lawful mechanism: an adequacy decision (like the EU-US Data Privacy Framework) or SCCs, where appropriate. But an EU-hosted provider can make life significantly simpler — fewer transfer decisions and potentially fewer cross-border subprocessors, depending on the configuration.
I have 16 GB VRAM — which models can I realistically run?
For text work, 7B–14B models at 4-bit quantisation are the sweet spot. A 24B model fits on 16 GB with careful settings and is noticeably smarter; 32B and 70B models will not fit comfortably and force heavy offloading to RAM, which hurts speed.
Who is the data controller when an employee uses ChatGPT at work?
For business purposes, the employer usually decides what data is sent and why, so the employer is the controller, while the provider may be a processor. The employee is the person who actually types — which is exactly why companies need an internal AI policy and a role-based assessment, not just a vendor contract.
The bottom line for European users is refreshingly simple: cloud AI is about capability, local AI is about control, and GDPR is about knowing which one you actually need for each prompt. Classify the data before you type, verify where it lands, and never assume that a big brand's privacy page covers your obligation. It does not — the organisation responsible for the processing still needs to assess its role and obligations, rather than assuming the model or vendor does that work for it.
