AI without the CLOUD Act. Where a company's data really goes when it uses artificial intelligence.
A provider tells you the data stays on servers in Europe, and the question seems settled. Under the CLOUD Act the server location is not what counts. What counts is who the provider is, and almost every AI model is American.
The CLOUD Act applies to every provider subject to US law, wherever it keeps the data. Section 2713 of Title 18 of the US Code, added in 2018, requires the provider to hand over data it controls even when that data sits outside the United States.
The GDPR says the opposite, from the European side. Article 48 recognises a request from a non-EU authority only if it rests on an international agreement, and in 2019 the European Data Protection Board wrote that a US request on its own is not a legal basis.
The most widely used AI models are American, and EU data residency does not take them out of the CLOUD Act. It lowers the risk; it does not remove it. What a company really controls is where it stores its own data and how much text it sends to the model.
There are eight questions worth asking a provider, and they are in the table further down: who provides the model, where the data sits at rest, whether the model keeps it, and who approves before anything goes out.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
This piece sits next to the guide on the AI Act and the GDPR and covers one side only: data that leaves Europe. It is written for anyone deciding whether customer data goes into an AI system, and it stops where a lawyer's work begins.
For day-to-day use, meaning what you can type into ChatGPT at work and why personal accounts are the biggest risk, there is the guide ChatGPT, GDPR and the CLOUD Act. This page is about choosing a provider and how a system is built.
What the CLOUD Act is, in one line of law
The CLOUD Act (Clarifying Lawful Overseas Use of Data Act) is a 2018 US federal law, enacted as part of Public Law 115-141. It added Section 2713 to Title 18 of the US Code, which requires providers of communication and remote computing services to preserve and disclose data in their possession, custody or control.
The part that matters to a European company is the end of the sentence. The obligation applies «regardless of whether such communication, record, or other information is located within or outside of the United States», so it makes no difference whether the data is in the US or anywhere else.
So «where are the servers?» is the second question. The first is «which country's law does the company running my data answer to?». A US provider with a data centre in Frankfurt is still a US provider.
Servers in Europe are not enough: the provider is what counts
The gap between where data sits and who controls it is where almost every promise of «European AI» gets muddled. This table lines up the providers a small business meets most often when it uses artificial intelligence.
| Provider | Based in | Servers in Europe and US law |
|---|---|---|
| OpenAI | United States | GPT models and ChatGPT. European servers available with EU data residency; still subject to US law. |
| Anthropic | United States | Claude models. European regions through cloud platforms; still subject to US law. |
United States | Gemini and Vertex AI, with European regions; still subject to US law. | |
| Microsoft | United States | Azure and Copilot, with European regions; still subject to US law. |
| Amazon | United States | AWS and Bedrock, with European regions; still subject to US law. |
| Cloudflare | United States | Network and website protection on a global scale; still subject to US law. |
| Hetzner | Germany | Servers and data centres in Nuremberg, Falkenstein and Helsinki. European company. |
| Mistral AI | France | Language models from a European company. European company. |
The column that counts is where the company is based. A provider based in Europe is not reached by the CLOUD Act the way a US company is. Hetzner also runs data centres in the United States and Singapore, though: for your own data, what matters is which data centre you pick.
What the GDPR says, and what the European Board said
On the European side the rule is Article 48 of Regulation (EU) 2016/679, in Chapter V on transfers to third countries. A judgment or administrative decision from a non-EU country ordering the transfer or disclosure of personal data can be recognised «soltanto se basate su un accordo internazionale», in other words, only if it rests on an international agreement such as a mutual legal assistance treaty.
In July 2019 the European Data Protection Board and the European Data Protection Supervisor answered the European Parliament on exactly this point. Their conclusion: a request from a US authority, on its own, is not a legal basis for transferring data to the United States. On 2 December 2024 the Board adopted draft Guidelines 02/2024 on Article 48.
In practice this is a conflict of laws. The US provider is required on one side to hand the data over and on the other not to without an agreement, and if it happens, the loser is the European company that entrusted it with the data.
The Data Privacy Framework still stands, and it does not solve the CLOUD Act
For ordinary data transfers to the United States there is the Data Privacy Framework, the Commission's adequacy decision (EU) 2023/1795 of 10 July 2023. On 3 September 2025 the General Court of the European Union dismissed the action seeking its annulment (case T-553/23); an appeal to the Court of Justice is still possible.
The Data Privacy Framework makes it lawful to send data to a certified US company. It does not change the fact that the company, once it holds the data, stays subject to the CLOUD Act. These are two separate questions, and it pays to keep them apart when you read a contract.
The step almost nobody talks about: the model
An AI system has two parts. The company's data lives in a store, a database, and from there a piece of text is sent to the model, which reads it and answers. The store can sit entirely in Europe, with a European provider. The model almost always belongs to a US company.
US providers offer data residency in Europe. On 5 February 2025 OpenAI announced its own: for the API, requests from new projects are processed in the European region without storing prompts or responses; for ChatGPT Enterprise and Edu, conversations and files are stored at rest in Europe. It is real protection against data retention, not against the provider's jurisdiction.
There are models from European companies, such as Mistral AI, and models you can install on your own server. Today the most capable models are still American, and the choice almost every company makes, ours included, is to use them in the way that lets the least data out.
The eight questions to ask an AI provider
These are the questions that separate a promise from an architecture. Each one has an answer you can check in a contract or on the provider's own pages.
| Question | Why it matters | A good answer |
|---|---|---|
| Where is my data stored at rest? | it is the part you control | a named EU data centre |
| Which company runs that store? | it decides which law applies | a provider based in the EU |
| Which AI model reads my text? | it is the step outside the store | the model and its provider, by name |
| Does the model keep prompts and responses? | the less is kept, the less can be handed over | no retention, in writing |
| Is EU data residency switched on? | it moves processing to the EU | yes, on the project in use |
| Is my data used to train the model? | it would leave your control | no, by contract |
| Is there an Article 28 GDPR agreement? | it is a legal obligation | yes, with the list of sub-processors |
| Who approves before anything goes out? | a mistake stays inside the company | a person, every time |
How Itria works
We spell it out, because it is the question a customer asks right after reading the sections above. The new architecture we are moving Itria's systems onto keeps databases, logins, automations and analytics on a Hetzner server in its Helsinki data centre: a provider based in Germany, with the data stored in Europe.
The AI model is the only point where text leaves our server. We use US models, like almost everyone, and the architecture is designed to run them with EU data residency and no retention of prompts. The model only receives the text a task needs, never the store. The model provider is still a US company and the CLOUD Act still applies to it, which is why the real defence is to send it as little as possible.
A person approves every message that goes out to a customer. The list of systems we actually use is on the AI transparency page, and the principles on the AI principles page.
Questions and answers
Is there an AI model outside the reach of the CLOUD Act?
Yes, if the model provider is not subject to US law, such as a European company or a model running on your own server. The most widely used models, from OpenAI, Anthropic, Google and Microsoft, belong to US companies and stay subject to the CLOUD Act even when the servers are in Europe.
If the servers are in Europe, is the data safe from the CLOUD Act?
Not if the provider is American. Section 2713 of Title 18 of the US Code requires the provider to hand over data it controls wherever that data sits. Server location matters for the GDPR and for retention; the provider's home country matters for the CLOUD Act.
Does the GDPR allow a provider to hand data over to the United States?
Only on a specific basis. Article 48 of the GDPR recognises an order from a non-EU authority only if it rests on an international agreement, and in 2019 the European Data Protection Board and the European Data Protection Supervisor wrote that a US request on its own is not a legal basis.
Is OpenAI's data residency in Europe enough?
It lowers the risk; it does not remove it. For the API, since 5 February 2025 requests from new projects are processed in Europe without storing prompts or responses, and for ChatGPT Enterprise and Edu the data is stored at rest in Europe. OpenAI is still a US company, so the jurisdiction does not change.
What can a small business actually do?
Three things. Store its own data with a provider based in Europe, send the model only the text a task needs, with EU data residency switched on and no retention, and have a person approve everything that goes out to a customer.
Notes on sources
- 18 U.S. Code § 2713, Legal Information Institute, Cornell University: text of the section added by Public Law 115-141 of 2018 (the CLOUD Act).
- Regulation (EU) 2016/679 (GDPR), EUR-Lex: Chapter V on transfers to third countries, Article 48 and Article 28 (Italian version).
- EDPB and EDPS, joint response to the LIBE Committee on the CLOUD Act, July 2019, with the annexed assessment.
- General Court of the European Union, case T-553/23, 3 September 2025, action against adequacy decision (EU) 2023/1795 dismissed, as reported by Lewis Silkin.
- OpenAI, Introducing data residency in Europe, 5 February 2025, updated 16 January 2026.
- Hetzner, Data centers and connection: the Nuremberg, Falkenstein and Helsinki (Tuusula) parks. Hetzner Online GmbH is headquartered in Gunzenhausen, Bavaria.
- Mistral AI, company page: a European company headquartered in Paris.
This article is an operational overview, not legal advice. On data transfers outside the Union and on contracts with providers, the answer comes from a professional who looks at your company.
Before you put customer data into an AI system, map where it lives and what leaves.
The eight-question table takes an hour to fill in, and it works for every provider you already use. If you are about to build a system that touches customer data, it pays to design it with the store in Europe, the least possible text going to the model and a person approving. Thirty minutes on a call, with the Dashboard open.