# Data residency for AI workloads: a practical checklist for the EU and India > Map AI data flows and prepare a practical EU and India residency review covering models, retrieval, logs, backups and access. Source: https://hyperlake.cloud/blog/data-residency-for-ai-workloads-a-practical-checklist-for-the-eu-and-india Published 2026-10-21 · by Hyperlake Team · Hyperlake Data residency for an AI workload covers every place its data is stored or processed, including prompts, retrieved documents, model outputs, logs and backups. A regional database is only one part of that picture. Build a complete data-flow map, then assess the laws, contracts and operational access that apply to each flow. This is an engineering checklist for preparing that review. It does not classify a particular AI system under the EU AI Act or determine whether an organisation has met its GDPR or India DPDP obligations. ![AI data residency in practice: Map data flows, Review jurisdictions, Check support access, Rehearse recovery.](https://hyperlake.cloud/blog/img/production/6d59a133bced293602277d064316a7b55d92ebcc-1600x960.png?w=1600&fit=max&auto=format) *AI data residency in practice* ## What data can leave an AI application? Start with a single user request and follow it through the system. A question may pass through an application gateway, an embedding service, a retrieval store, a model endpoint and an observability service before the answer returns. A support engineer may later download the same trace into a ticket. Write down the payload at each step. A vector database may hold document fragments and metadata as well as embeddings. A log may contain the full prompt even when the main application claims to retain no conversations. Error reports can preserve input that a successful request would not retain. Include background jobs. Evaluation pipelines, fine-tuning runs, moderation services, backups and disaster recovery often follow different routes from interactive traffic. Their locations and access rules need the same scrutiny. ## Which questions belong in the EU review? For personal data, ask counsel or the privacy team to assess the relevant GDPR obligations, including processing purpose, lawful basis, minimisation, retention and transfers. Residency is one input to that assessment. It does not replace the assessment. The [EDPB guidance on Article 3 and international transfers](https://www.edpb.europa.eu/documents/guideline/guidelines-052021-on-the-interplay-between-the-application-of-article-3-and-the_en) is a useful starting point for reviewing cross-border arrangements. Map who receives or can access data, rather than reasoning only from the address of the storage facility. For AI Act review, describe what the system actually does and the organisation's role. A document-search assistant and a system used in a consequential decision may raise different questions. The [European Commission's AI Act overview](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) explains the risk-based framework; classification and applicable obligations need a current, system-specific review. Prepare a compact EU evidence packet: intended use, data categories, provider relationships, locations, access paths, retention settings and the people responsible for approval. Record unresolved questions instead of turning assumptions into compliance statements. ## Which questions belong in the India review? Use the final legal texts and applicable commencement provisions when assessing India DPDP requirements. Do not build a production checklist from an old consultation draft or assume that every provision takes effect on the same date. The [MeitY publication page for the Digital Personal Data Protection Rules, 2025](https://www.meity.gov.in/documents/act-and-policies/digital-personal-data-protection-rules-2025-gDOxUjMtQWa) is a starting point for obtaining the official material. Ask the legal reviewer to identify the organisation's role, covered processing, notices, individual requests, security and breach processes, and any applicable transfer or sector-specific constraints. Record the answer for the actual workload and data category rather than applying a blanket rule to all AI traffic. The engineering packet should show how the application can locate a person's data, apply the approved retention policy and propagate changes through derived stores. If documents are removed from the source system, explain what happens to chunks, indexes, caches and backups. Also identify where a human can intervene. A deletion request that depends on an engineer manually finding every copy is operationally different from a documented, repeatable process with an owner and an audit trail. ## How do the regional checklists compare? This table organises the review; it does not imply that the two frameworks have identical requirements. ## What should you ask a cloud or model provider? Ask about the selected service and deployment option, not the provider in the abstract. AWS, Microsoft Azure and Google Cloud each offer multiple services with different configurations. A statement about one regional storage product does not establish where a separate AI endpoint processes requests. Ask whether failover changes location, which telemetry leaves the region, how support accesses customer content and which subprocessors are involved. Request the contract language and configuration that correspond to the answer. Choosing an open model such as a Llama or Mistral model changes part of the deployment choice, but the model family alone does not determine residency. The host, serving software, retrieval components, monitoring and operational practices still create the data flows. A useful procurement question is, “Show us the complete request path when the primary endpoint fails.” Fallback behaviour is easy to miss in a normal demonstration and can undermine an otherwise careful regional design. ## Where does self-hosting help? Running components in a customer-controlled environment can make placement, network policy and operator access more directly configurable. HyperLake supports deployment in customer environments, including private and on-premises infrastructure, with modular data and AI components. That is a deployment capability, not an automatic compliance result. The organisation still needs to configure access, approve dependencies, control telemetry and operate the system. A self-hosted model that forwards traces to an external service still has an external data flow. Start with an egress inventory. For every outbound destination, identify the purpose, payload, owner and approval. Separate package and model downloads from runtime data transfers so the team can make different decisions about each. ## What should be checked before launch? Walk through one normal request, one failure and one administrative support case. Inspect the resulting traces and storage records. Confirm that the documented path matches the implementation rather than relying solely on a diagram. Then rehearse a restore and a data-removal request. These exercises expose forgotten copies and unclear ownership. Record limitations honestly, including retention constraints that mean a backup copy is removed on a schedule rather than immediately. Finish with a signed-off responsibility map: who approves a new model endpoint, who changes retention, who reviews access and who responds to an incident. Residency remains an operating responsibility after the first deployment. ## Frequently asked questions ### Does EU hosting make an AI system GDPR compliant? No. Hosting location is one consideration. Purpose, lawful basis, individual rights, security, relationships between parties and applicable transfer arrangements also need assessment. ### Does using an open model keep data local? Only if the complete deployment keeps the relevant data local. Model hosting, embeddings, retrieval, monitoring, backups and support access all need to be considered. ### Should prompts be included in the data inventory? Yes. Prompts can contain personal, confidential or commercially sensitive information, including material copied from retrieved documents. Include outputs and traces as well. ### Who should approve the final checklist? The owners of the system, security controls and privacy or legal review should approve their respective responsibilities. Engineering evidence supports that decision but does not replace legal interpretation. ## Related reading - [Sovereign AI explained](https://hyperlake.cloud/answers/what-is-sovereign-ai) - [Deploying an LLM on premises](https://hyperlake.cloud/answers/how-to-deploy-an-llm-on-premises)