EU-Hosted, GDPR-Compliant AI Chatbot: Which Providers Keep Customer Data in the EU (2026)
If you want an AI chatbot but need customer data to stay in the EU, you have four realistic options: (1) a frontier model through a hyperscaler’s EU boundary (Azure OpenAI’s EU Data Zone, Amazon Bedrock’s EU inference profiles or Google Vertex AI’s EU endpoint); (2) OpenAI’s own Europe data residency, which needs sales approval; (3) a European provider such as Mistral, or open models served by IONOS, STACKIT, OVHcloud or Scaleway; or (4) an open-weight model you host yourself on EU GPUs.
Two things catch people out. Anthropic’s own Claude API has no EU option yet (only US or global), so EU teams that want Claude use it through Bedrock or Vertex AI in the EU. And the model is only one of the places customer data goes: the vector database, conversation logs, tracing, error monitoring, analytics and the helpdesk your bot hands off to all need to be in the EU as well.
This guide is about where to host and which providers keep what in the EU. For the wider GDPR design (lawful basis, data subject rights, human oversight), see our GDPR-compliant AI agents guide. It is practical engineering guidance, not legal advice.
- Frontier models with EU processing
- Azure, AWS, Google
- Claude via Anthropic’s API
- US or global only
- OpenAI Europe residency
- Sales approval, new projects
- AI Act chatbot disclosure
- Applies since 2 Aug 2026
The options compared
What counts is not where a provider has data centres but what it documents as staying in the EU: processing of prompts and outputs, stored data (files, conversation state, embeddings) and the logs it keeps for abuse monitoring. This is what each route says as of November 2026.
| Option | EU guarantee | What stays in the EU | Caveats |
|---|---|---|---|
| Azure OpenAI (Microsoft Foundry) | Data Zone deployments (Standard, Provisioned, Batch) in an EU resource process prompts and responses only inside the EU data zone; Standard and Regional Provisioned stay in the resource’s Azure geography | Prompts, completions and embeddings; data at rest (files, Responses API history, stored completions) and the abuse-monitoring store in your resource’s geography; human abuse reviewers for EEA deployments are in the EEA | The EU data zone follows the EU Data Boundary, which can include Norway and Switzerland. Global and Developer types can process anywhere. Modified abuse monitoring is being withdrawn for models released from 1 November 2026 |
| Amazon Bedrock | EU geographic cross-Region inference profiles (eu. model IDs) keep requests within EU Regions | Prompts and outputs processed in EU Regions; anything kept for abuse detection is stored in the EU Region that served the request; model providers have no access to prompts or logs | Called from London or Zurich, the EU profile can also run there. Global profiles route worldwide. Claude Fable models require up to 30 days’ retention with possible AWS human review |
| Google Vertex AI | The eu multi-region endpoint keeps ML processing within EU member states; some models also run in single EU regions | Prompts and outputs processed in the EU; data at rest stays in the location you choose; the in-memory cache (up to 24 hours) sits in the serving region | The global endpoint has no residency guarantee. eu excludes the UK and Switzerland. Grounding with Google Search keeps query logs for up to 3 days. Google may log prompts for abuse monitoring |
| OpenAI API (direct) | Europe data residency (EEA and Switzerland) on eu.api.openai.com, for eligible organisations via sales, on new projects only | Regional processing for Responses, Chat Completions, Embeddings, Realtime, Batch and hosted tools; customer content at rest in Europe | Files and vector stores are stored in Europe but get no regional processing. System data is excluded. 10% uplift on eligible models released from 5 March 2026 |
| Anthropic Claude API (direct) | None today: inference_geo is us or global, and workspace (storage) geo is US only | Nothing guaranteed in the EU | Use Claude through Bedrock’s EU profiles or Vertex AI’s EU endpoint instead |
| Mistral AI | European company; data hosted in the EEA by default, with a US endpoint only if you choose it | Inference and stored data, by default | Some features may temporarily send data outside the EU to sub-processors listed in its Trust Center; enterprises can switch some off. Check the training opt-out |
| EU clouds serving open models (IONOS, STACKIT, OVHcloud, Scaleway) | Inference in the provider’s own data centres: IONOS in Germany, STACKIT in Germany and Austria, Scaleway currently in Paris, OVHcloud “in Europe” | Everything they process; each says prompts are not stored (IONOS: never logged, no third-party sub-processors) | Open-weight models only, so test answer quality. Product-page statements are not always contractual; get location and retention in the DPA |
| Self-hosted open-weight model | Wherever you put the GPUs: Hetzner (Germany, Finland), OVHcloud, Scaleway, STACKIT, or your own EU cloud account | Everything, under your control | You own uptime, scaling, patching, safety filtering and model upgrades |
How to lock each option to the EU
Each route has a default that leaves the EU. These are the settings to change and the policies that stop someone undoing them later.
Azure OpenAI (Microsoft Foundry)
Create the resource in an EU region and deploy with Data Zone Standard (or Standard, to stay in one Azure geography). Microsoft publishes an Azure Policy rule that blocks a deployment type; use it to deny GlobalStandard so nobody adds a global deployment for extra quota. Microsoft’s own documentation says EU-zone processing can include EFTA countries such as Norway and Switzerland, so tell your DPO if “EU” must mean EU member states only.
If you rely on modified abuse monitoring (no storage of prompts for human review), note Microsoft’s 17 November 2026 update: it continues for approved customers on models released before 1 November 2026, but will not be available for newer models from a date still to be announced, with at least 30 days’ notice. Also note that Anthropic’s docs only mention a US Data Zone for Claude hosted on Azure.
Amazon Bedrock
Call Bedrock from an EU Region inside the EU and use the model’s eu. inference profile. For Claude Sonnet 5.5, for example, the EU profile called from Frankfurt can route to Frankfurt, Stockholm, Milan, Spain, Ireland and Paris; called from London it adds London, and from Zurich it adds Zurich. AWS says not to rely on the prefix alone: call GetInferenceProfile from each source Region and check the destination list. Claude Sonnet 5.5, for one, has no single-Region option in the EU, so the EU profile is the realistic choice.
Block other Regions and global. profiles with Service Control Policies. Set the account’s data-retention mode to none where your model allows it; Claude Fable 5.x models require aws_review, which means up to 30 days’ retention and possible human review by AWS (not shared with Anthropic).
Google Vertex AI
Call the eu multi-region endpoint, or a single EU region where your model supports one, never global. Gemini and Claude models are listed for the EU multi-region. For zero retention, Google’s checklist is: leave request-response logging off, disable caching, avoid Grounding with Google Search (use Web Grounding for Enterprise), and don’t enable Live API session resumption.
For Claude, use it in Vertex AI on the EU endpoint. Anthropic’s separate “Claude Models on Google Cloud (Anthropic managed)” offering uses US workspace storage and only US or global inference.
OpenAI API (direct)
Ask OpenAI sales to make your organisation eligible, create a new project with Europe residency (existing projects can’t be switched) and send traffic to eu.api.openai.com. Keep store=false on Chat Completions, since store=true isn’t allowed outside the US. Files and vector stores are stored in Europe but get no regional processing, so keep retrieval in your own EU vector database. OpenAI’s launch post said EU-resident API requests are handled with zero data retention; the current guide doesn’t restate that, and standard API abuse-monitoring logs are kept for up to 30 days, so get the retention terms in writing. The UK region is storage only.
European providers and self-hosting
With Mistral, stay on the default EU endpoint and check the training setting: its help centre says free-mode data may be used for training, with an opt-out on free and pay-as-you-go. With IONOS, STACKIT, OVHcloud or Scaleway, ask for the location and no-logging statements in their docs to be written into the DPA. STACKIT, for example, says it stores no customer data from requests and doesn’t train on it; OVHcloud says data “is not stored or shared”; Scaleway says it doesn’t read or reuse prompts.
To self-host, put an open-weight model on EU GPUs. Hetzner, for example, runs its GPU servers in ISO/IEC 27001-certified data centres in Germany and Finland, with 96 GB cards on its GEX131. For EU-operated hyperscaler infrastructure, the AWS European Sovereign Cloud (Brandenburg, Germany, generally available since January 2026) runs open-weight models on Bedrock with in-Region inference.
Which option fits
| If you… | Start with |
|---|---|
| Run on Microsoft and want GPT models | Azure OpenAI, Data Zone Standard in an EU region |
| Want Claude | Bedrock eu. profile, or Vertex AI eu endpoint |
| Run on Google Cloud or want Gemini | Vertex AI eu endpoint |
| Need OpenAI’s own API features directly | OpenAI Europe data residency (talk to sales early) |
| Want an EU-headquartered provider | Mistral, IONOS, STACKIT, OVHcloud or Scaleway |
| Handle health, public-sector or similarly sensitive data | Self-hosted open model in the EU, or a hybrid: self-hosted for sensitive steps, EU cloud model for the rest |
For most UK and EU customer service bots we start with a frontier model through Bedrock or Azure in the EU, with retrieval, logs and conversation history in the customer’s own EU account. It gives the best answer quality with the least operational work. Residency can cost extra (OpenAI adds 10% on newer models); our API pricing comparison has the base rates.
Reference architecture for an EU-only chatbot
Every arrow below carries customer data, so every box needs an EU location and a DPA.
| Component | EU-only choice | Notes |
|---|---|---|
| Chat widget | Your own widget, or a vendor’s with EU hosting | Static files can come from a CDN; chat calls and transcripts go only to your EU backend |
| Backend | Containers or functions in one EU region, in the same cloud account as the model | Session state, secrets and tool credentials live here |
| Model and embeddings | An EU-pinned endpoint from the table above | Use the same EU endpoint for embeddings and reranking; treat vectors of customer text as personal data |
| Vector database | pgvector in your EU Postgres, or a managed vector DB in an EU region | Tag every chunk with a customer ID so erasure requests can find it |
| Conversation store | EU database with a scheduled retention job | Redact emails, phone and card numbers before storing |
| LLM tracing | Self-hosted Langfuse, or Langfuse Cloud EU (Ireland, with database backups copied to Frankfurt) | Traces hold full prompts; give them the same retention as transcripts |
| Error monitoring | Sentry with EU data storage (Frankfurt) | Chosen at organisation creation and can’t be changed; account data stays in the US |
| Analytics | EU-hosted or self-hosted analytics | Send events and outcomes, not transcripts |
| Helpdesk handoff | Zendesk with the Data Center Location add-on (EEA), or Intercom with EU hosting | Check the plan: Intercom offers EU hosting to new customers on certain plans; Zendesk’s add-on lists which products it covers |
| Email, CRM, voice | Providers with EU processing | Confirmation emails, CRM sync and speech-to-text are processors too |
The retrieval side is the same as any RAG system; our production RAG chatbot guide covers chunking, retrieval and evaluation. For what a support bot should handle and when to hand off, see AI customer support agents.
The legal pieces that still apply
- DPA and sub-processors. Article 28 GDPR needs a contract with every processor. Get each provider’s sub-processor list and change notices, including the model provider behind a cloud platform.
- Transfers. EU hosting removes most transfers, not all: support access, “system data” (OpenAI excludes it from residency) and account data (Sentry keeps it in the US) can still leave. The tools are the 2021 standard contractual clauses or, for certified US companies, the EU–US Data Privacy Framework. The General Court upheld the framework on 3 September 2025; an appeal (C-703/25 P) went to the Court of Justice in October 2025 and we found no judgment as of November 2026, so keep SCCs ready as a fallback.
- UK. The Commission renewed its UK adequacy decisions on 19 December 2025, to 27 December 2031. That matters when a route touches London.
- DPIA. Article 35 requires one where processing is likely to be high risk. The UK ICO lists AI as an innovative technology that needs a DPIA when combined with one of the nine EDPB-endorsed criteria, such as large-scale or sensitive data. A customer chatbot that sees account data usually deserves one.
- Retention. Map the provider’s windows as well as yours: OpenAI’s standard abuse logs (up to 30 days), Bedrock’s retention mode, Vertex AI’s 24-hour cache and 3-day search grounding logs, Azure’s abuse-monitoring store.
- Training. Azure, Vertex AI and the OpenAI API document that your prompts are not used to train models by default, Bedrock doesn’t share them with model providers, and Mistral offers an opt-out. Confirm each in the contract.
- EU AI Act. Article 50 has applied since 2 August 2026: providers of AI systems that talk to people must make sure people know it is an AI, at the latest at the first interaction, unless it is obvious. The 2026 Digital Omnibus delayed the high-risk rules but, per law-firm briefings, not this duty. Put “AI assistant” in the widget header and the first message.
Where EU-only setups leak
- A global deployment added for quota. One
GlobalStandarddeployment orglobal.profile undoes the setup. Block them with policy. - Calling an EU profile from London or Zurich. Fine for many teams, but not EU-only.
- Provider-side files and vector stores. OpenAI stores them in Europe but may process them elsewhere; keep retrieval in your own EU database.
- Web search and third-party tools. Grounding with Google Search keeps logs, and MCP servers follow their own residency rules, not your model provider’s.
- Observability on US defaults. Tracing, error tracking and analytics SDKs capture full prompts unless you choose EU regions and redact.
- The handoff. A bot that is EU-only but escalates into a US helpdesk instance moves the whole conversation.
Checklist
Before you sign
- DPA with every processor in the chain
- Sub-processor list reviewed, with change notices
- EU residency in the contract, not just a docs page
- Abuse-monitoring retention terms in writing
- No-training terms confirmed, opt-outs set
- SCCs or DPF for anything that leaves the EU
When you build
- Model pinned to an EU endpoint
- Global options blocked by Azure Policy or SCPs
- Embeddings on the same EU endpoint
- Vector DB, transcripts and backups in the EU
- PII redacted before logs and traces
- Tracing, errors, analytics and helpdesk in EU regions
At launch and after
- DPIA done and signed off
- AI disclosure in widget and first message
- Privacy notice lists processors and retention
- Retention jobs running and tested
- Erasure request tested end to end
- Quarterly check of provider region and retention changes
FAQ
We want to host an AI chatbot but keep customer data in the EU. What are our options?
Four: a frontier model through a cloud provider’s EU boundary (Azure OpenAI EU Data Zone, Amazon Bedrock EU inference profiles or Google Vertex AI’s EU endpoint); OpenAI’s own Europe data residency, which needs sales approval; a European provider such as Mistral, IONOS, STACKIT, OVHcloud or Scaleway; or an open-weight model you host on EU GPUs. Keep the vector database, logs, tracing and helpdesk in the EU too.
Can we use Claude with EU data residency?
Not through Anthropic’s own API today: its inference_geo setting offers only us or global, and workspace storage is US only. Use Claude through Amazon Bedrock’s EU cross-Region inference profiles or Google Vertex AI’s EU multi-region endpoint.
Can we use OpenAI models and keep data in the EU?
Yes, two ways: Azure OpenAI with an EU Data Zone or EU regional deployment, or OpenAI’s own Europe data residency on eu.api.openai.com for eligible organisations and new projects. With OpenAI direct, files and vector stores are stored in Europe but get no regional processing, and system data is excluded.
Is EU hosting enough to make a chatbot GDPR compliant?
No. It removes most international transfers, but you still need a lawful basis, a DPA with every processor, a DPIA where risk is high, retention limits, a way to handle access and erasure requests, and an updated privacy notice.
Does the EU AI Act apply to a customer service chatbot?
Its transparency rule does. Since 2 August 2026, Article 50 requires that people are told they are interacting with an AI system, at the latest at the first interaction, unless that is obvious. A typical customer service chatbot is not in the Act’s high-risk list.
Is there an agency that does both AI and GDPR compliance?
stackcone builds customer service chatbots and AI agents for UK, US and EU businesses with EU-region models, EU-hosted vector stores and logs, PII redaction, retention controls and the documentation a DPO review needs. Legal sign-off stays with your DPO or counsel.
Sources
Provider documentation as published up to November 2026. Regions, model lists and retention terms change often; check the live page and your contract before relying on them.
- Microsoft Learn: Deployment types in Microsoft Foundry Models
- Microsoft Learn: Data, privacy and security for Foundry Models sold by Azure
- Microsoft: What is the EU Data Boundary?
- AWS: Geographic cross-Region inference in Amazon Bedrock
- AWS: Supported Regions and models for inference profiles
- AWS: Claude Sonnet 5.5 on Bedrock (Regional availability)
- AWS: Amazon Bedrock data protection
- AWS: Amazon Bedrock abuse detection
- AWS: Amazon Bedrock data retention
- AWS News Blog: Opening the AWS European Sovereign Cloud
- AWS Security Blog: Open-weight models on Bedrock in the AWS European Sovereign Cloud
- Google Cloud: Generative AI on Vertex AI data residency
- Google Cloud: Vertex AI and zero data retention
- OpenAI: Data residency guide
- OpenAI: Data controls in the OpenAI platform
- OpenAI: Introducing data residency in Europe
- Anthropic: Claude API data residency
- Mistral AI: Where do you store my data?
- Mistral AI: Do you use my data to train your models?
- IONOS: AI Model Hub data handling
- STACKIT: AI Model Serving FAQ
- STACKIT: LLMs on STACKIT
- OVHcloud: AI Endpoints capabilities
- Scaleway: Generative APIs data privacy
- Scaleway: Generative APIs FAQ
- Hetzner: GPU servers
- Langfuse: Data regions
- Sentry: Data storage location
- Zendesk: Regional Data Hosting Policy
- Intercom: Regional data hosting
- EUR-Lex: General Data Protection Regulation (Articles 28 and 35)
- European Commission: Standard contractual clauses
- European Commission: EU–US data transfers
- EUR-Lex: Case C-703/25 P, Latombe v Commission (appeal)
- WilmerHale: Court of Justice to review challenge to the Data Privacy Framework
- Hunton: European Commission renews UK adequacy decisions
- ICO: When do we need to do a DPIA?
- EU AI Act, Article 50: Transparency obligations
- McCann FitzGerald: EU AI Act transparency compliance
- Jones Walker: August 2 still matters for AI Act transparency
