Skip to content

Where do your prompts go?
EU data residency by provider

OpenAI, Microsoft Azure, Google Vertex AI and Mistral all offer EU processing for business customers, usually as a setting you choose. Anthropic’s own API offered only global or US processing when we checked, so EU processing for Claude goes through AWS Bedrock or Google Vertex AI. Training and retention defaults differ, so check each account.

Why the location of processing matters

An AI agent sends each conversation to a model provider. If the provider processes or stores that data outside the European Economic Area, GDPR Chapter V applies, and you need a transfer mechanism such as an adequacy decision or standard contractual clauses. EU processing does not remove your other GDPR duties, but it removes the transfer question for the model call, and it is easier to explain to customers.

Three things are often mixed up. Inference location is where the model runs. Storage at rest is where saved data sits, such as files or conversation state. Retention is how long the provider keeps prompts and outputs, for example for abuse monitoring. Check all three.

The comparison table

EU processing, retention and training defaults, checked 19 November 2026
OptionEU processingDefault retentionTraining on your dataHow to turn on EU
OpenAI APIYes. Europe region (EEA and Switzerland) with in-region processing for supported endpointsAbuse monitoring logs up to 30 days. Zero data retention only with approvalNo, unless you opt inBecome eligible through OpenAI sales, create a project with the Europe region, call eu.api.openai.com
ChatGPT Enterprise and EduStorage at rest in Europe for new workspaces. Ask about in-region processingNot stated on the pages we read; ask OpenAINo, by default for business plansSet up a new workspace with Europe data residency, through OpenAI sales
Anthropic Claude API (direct)No. Inference geo is global or US only; workspace storage is US onlyDeleted within 30 days. Some newest models require 30-day retentionNo, by default for commercial productsNot available on the direct API. Use Bedrock or Vertex AI
Claude on AWS BedrockYes. EU geographic inference profiles keep processing in EU regionsDepends on model and your retention mode. Newest Claude models require up to 30 days within AWSContent is not shared with the model providerCall an EU inference profile (eu. prefix) from an EU region
Claude on Google Vertex AIYes. EU multi-region listed for all Claude models; some older models also in BelgiumSee Google’s retention settingsNo, without your permissionUse the EU multi-region or a regional endpoint, not the global endpoint
Azure OpenAI (Microsoft Foundry)Yes. Standard deployment in your geography, or the EU Data Zone (any EU member state)Abuse monitoring data store in your geography; no day count stated on the page we readNo. Not available to OpenAICreate the resource in an EU region and use a Standard or DataZone EU deployment, not Global
Google Gemini on Vertex AIYes. EU multi-region (EU member states only) or single EU regions, by modelIn-memory cache up to 24 hours by default; Search grounding logs 3 daysNo, without your permissionUse the eu multi-region or an EU regional endpoint; disable caching for zero retention
Mistral APIYes. Hosted in the EU by default; api.eu.mistral.ai guarantees EU or EFTA inference30 rolling days for abuse monitoring; Agents API data kept until account closureOpt-out toggle in admin Privacy settingsCall api.eu.mistral.ai (10% price uplift); request zero data retention if needed

Notes on each provider

OpenAI API. Data residency is set per project and cannot be changed afterwards. The Europe region supports regional processing for Responses, Chat Completions, Embeddings, Realtime, Audio, Images, Batch and other endpoints. Files and vector stores are stored in Europe but not processed there. You cannot set store=true on Chat Completions outside the US. Models released on or after 5 March 2026 cost 10% more on data residency endpoints. Account, billing and support data are not covered. See our OpenAI integration page.

Anthropic. On the direct Claude API, the inference_geo setting has two values: global (the default) and us. Workspace storage is US only. There is no EU option on the direct API at the time of writing. For EU processing of Claude, use Amazon Bedrock with an EU inference profile, or Google Vertex AI with an EU endpoint. On Bedrock, the cloud provider is the processor, and Bedrock states that it does not share your content with model providers. See our Anthropic integration page.

Microsoft Azure OpenAI. Microsoft states that prompts and outputs are not available to OpenAI and are not used to train models. A Global deployment can process data in any geography. A DataZone deployment in an EU member state can process data in any EU member state. Data at rest stays in your chosen geography. For resources in the EEA, human reviewers for abuse monitoring are in the EEA. Microsoft is changing its Modified Abuse Monitoring programme for models released from 1 November 2026, so check before you rely on it.

Google Gemini on Vertex AI. The eu multi-region covers EU member states only; the UK and Switzerland are outside it. Global endpoints give no residency guarantee. Gemini caches data in memory for up to 24 hours by default; you can turn caching off. Grounding with Google Search keeps query logs for up to three days, and this cannot be turned off. The consumer Gemini app and the Gemini Developer API are different products, and we did not assess them here.

Mistral. The help centre says data is hosted in the EU by default, but some features can send data to subprocessors outside the EU. The global endpoint does not commit to an inference location, so use the EU endpoint where it matters. Regional endpoints do not support Agents, Batch or the Files API. Zero data retention is available on pay-as-you-go for stateless calls, on request.

Contracts: the DPA and what residency does not cover

Each provider offers a data processing agreement (DPA). OpenAI has a DPA for the API and ChatGPT business products. Anthropic’s DPA is incorporated into its Commercial Terms. Microsoft’s Products and Services DPA covers Azure. Mistral publishes a Data Processing Addendum. For Google Cloud, check the data processing terms that apply to your contract.

EU residency has limits. It usually does not cover account, billing and support data. It does not cover tools the model calls, such as web search, third-party connectors or MCP servers. It does not cover your own systems: the hosting of the agent, its logs and your CRM. Map all of these in your DPIA, and ask the questions in our AI vendor questionnaire.

How to choose

  1. Decide whether you need EU processing, or EU storage with a valid transfer mechanism. Ask your DPO if you are not sure.
  2. If the agent handles health or other special category data, prefer EU processing and short retention.
  3. Pick the model for the task first, then the route to it. Claude in the EU means Bedrock or Vertex AI; GPT models in the EU mean the OpenAI Europe region or Azure.
  4. Price the uplift. OpenAI and Mistral charge 10% more for regional endpoints on the models we checked.
  5. Turn off training and caching settings you do not need, and record the settings.

Our model comparison covers quality and cost. When we build an AI agent, we set the EU route and retention as part of the build and document them for your DPIA.

Common questions

Can I use Claude with EU data processing?

Not through Anthropic’s direct API when we checked on 19 November 2026: its inference setting offers only global or US. You can use Claude with EU processing through Amazon Bedrock EU inference profiles, or Google Vertex AI EU endpoints. The model is the same, but the contract, the processor and the retention settings come from the cloud provider.

Does OpenAI keep my API prompts?

By default, OpenAI keeps abuse monitoring logs for up to 30 days, and does not use API data for training unless you opt in. Zero data retention, which removes customer content from those logs, needs OpenAI’s approval and only covers eligible endpoints. Stateful features such as stored responses or files keep data until you delete it.

Is EU data residency enough for GDPR?

No. It helps with international transfers for the model call, but you still need a lawful basis, a privacy notice, a DPA with each processor, retention rules and a way to handle rights requests. Residency also does not usually cover account and support data, or tools the model calls during a conversation.

Does EU processing cost more?

Sometimes. OpenAI adds 10% for data residency endpoints on models released from 5 March 2026, and Mistral charges 1.1 times list price on its EU endpoint. Azure, Bedrock and Vertex AI publish regional pricing that varies by model and deployment type, so compare the price for the exact model and region you plan to use.

Sources

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.