Data residency is the requirement, or the guarantee, that data is stored and often processed inside a specific country or region. It can come from law, from sector regulators, from a contract or from a company’s own policy. For AI systems it reaches further than people expect: prompts, uploaded documents, embeddings, logs, caches and model outputs are all data, and each can end up in a different place.
Residency, Sovereignty and Localisation
These terms are often mixed up:
- Data residency is about where the data physically sits and is processed.
- Data sovereignty is about whose laws apply to it. Data stored in the EU by a company subject to another country’s law can still be reachable under that law.
- Data localisation is a legal requirement that certain data must stay in a country, sometimes with a ban on copies abroad.
A system can meet residency and still raise sovereignty questions, which is why some organisations want both an in-region location and a provider or deployment model they control.
Why It Matters
In the EU, the GDPR does not require data to stay in Europe, but it restricts transfers of personal data outside the European Economic Area to destinations with adequate protection or with safeguards such as standard contractual clauses. For the United States, the European Commission adopted an adequacy decision for the EU-US Data Privacy Framework (Decision 2023/1795), which lets personal data flow to US companies that participate in the framework. Many organisations still prefer in-region processing to reduce dependence on such decisions. Healthcare, finance and the public sector often add stricter rules of their own, and enterprise customers frequently write residency into contracts.
Where Data Goes in an LLM System
| Data | Where it can leak out of region |
|---|---|
| Prompts and responses | The model provider’s inference region |
| Documents for retrieval | Storage, the vector database, the embedding API |
| Logs and traces | Observability and analytics tools |
| Cached prompts | Provider-side caches (prompt caching) |
| Fine-tuning data | The provider’s training environment |
| Backups and support access | Disaster-recovery regions, vendor support teams |
Providers now separate storage location from processing location. In Microsoft’s Foundry deployment types, for example, data at rest stays in the chosen Azure geography, while “Global” deployments may process prompts in any region, “Data Zone” deployments keep processing inside the US, EU or Asia Pacific zone, and regional deployments keep it within one geography. Other major providers offer comparable regional options. Read the exact terms: storage and processing guarantees are often different.
Ways to Achieve Data Residency for AI
- Regional deployment of a hosted model in the required zone, with logging and retention settings checked.
- Self-hosted open-weight models in your own cloud account or data centre, so prompts never leave your infrastructure.
- Redaction before the call. A local small language model or rules engine removes personal data before anything goes to an external model.
- In-region supporting services. Vector stores, observability, queues and backups placed in the same region as the model.
Common Failure Modes
- Checking only the model. The model runs in Frankfurt, but traces go to a US analytics tool.
- Confusing storage with processing. A provider stores data in the EU but processes it globally by default.
- New features, new regions. A new model version or feature launches first in global deployments only, and teams switch without noticing.
- Assuming self-hosting solves everything. It solves location, but security, access control and retention still need work.
In Practice
Residency is mapped flow by flow: where prompts, documents, embeddings, logs, traces, caches and backups are stored, and who can reach them, including the vendor’s support staff. The deployment model is then chosen per flow, for example a regional cloud endpoint for the model and the client’s own database for embeddings, and each choice is recorded with the contract terms that back it. Any change of model or vendor reopens the map, since a new provider may process data in another region by default.
This is the focus of our private LLM and data residency service, and it shapes the architecture of RAG and enterprise knowledge assistants. Data rules are also part of AI governance and the EU AI Act.