# Do I need a DPA for AI agents? > Usually, yes. If an AI agent processes personal data on a vendor's infrastructure or model, Article 28 GDPR normally requires a DPA with each such vendor, including the model provider. - Source: Sanbox knowledge base, by Sanlabs GmbH - URL: https://sanbox.cloud/kb/eu/dpa-for-ai-agents/ - Type: explainer - Section: EU, GDPR and data residency - Language: en - Published: 2026-11-06 - Updated: 2026-11-06 - German version: https://sanbox.cloud/de/kb/eu/avv-fuer-ki-agenten.md Usually, yes. If your AI agents process personal data on a vendor's infrastructure or send it to a vendor's model, that vendor is normally your processor, and Article 28 GDPR requires a data processing agreement (DPA) with it. With Sanbox you sign one DPA with Sanlabs GmbH for the Computers and keep a separate one with your model provider, because the agent calls the model with your own key. This page isn't legal advice: your data protection officer or counsel decides how the rules apply to your processing. ## When does Article 28 apply to AI agents? Article 28 applies when a service provider processes personal data on your behalf and on your instructions. Agents touch personal data more often than expected: names and email addresses in a repository's commit history, customer tickets, CRM exports, mailboxes, logged-in browser sessions. If an agent processes no personal data at all, Article 28 doesn't apply, but with file and web access that is hard to rule out. ## Is an AI service a processor? Usually. The Bavarian data protection authority (BayLDA) says AI used as a cloud service will, as a rule, be processing on behalf under Article 28. What decides it is whether the provider acts only on your instructions or also uses the data for its own purposes. The German data protection conference (DSK) takes a similar view in its May 2024 AI guidance: an external cloud AI provider acts on the controller's behalf, often as an Article 28 processor. Roles shift in two cases: - **The provider uses inputs or outputs for its own purposes,** such as training. In the DSK's view, that use needs its own legal basis, and the people concerned must be informed. The DSK considers AI applications that don't train on inputs and outputs preferable. - **Two parties decide purposes and means together.** That can be joint controllership under Article 26, which needs its own arrangement. If you sell an agent product to your own customers, you are often their processor, and your vendors become sub-processors. ## What must a DPA contain? Article 28 asks for a written contract (electronic form counts) that sets out the subject matter, duration, nature and purpose of the processing, the types of personal data and the categories of data subjects. It must oblige the processor to: - process data only on your documented instructions, including for transfers outside the EU; - bind its staff to confidentiality and secure the data under Article 32; - use sub-processors only with your prior specific or general written authorisation, tell you about changes so you can object, and pass the same obligations down, while staying liable to you for them; - help you with data subject requests and with your duties under Articles 32 to 36; - delete or return the data when the service ends; - give you the information you need to show compliance, and allow audits. A processor that decides purposes and means itself counts as a controller for that processing (Article 28(10)). ## Which vendors in an agent setup need one? The chain follows the data: | Layer | Examples | Usual role | What to request | |---|---|---|---| | Agent infrastructure | Computers, files, run records, logs | Processor | DPA, sub-processor list, TOMs | | Model provider | Anthropic, OpenAI, Mistral | Processor, under its DPA for business customers | DPA, sub-processors, region, training and retention terms | | Tools the agent calls | CRM, ticketing, search APIs | Depends on the service | DPA where it processes personal data for you | | Sub-processors | Hosting and services behind each vendor | Covered by that vendor's DPA | List with locations | The large model providers build their DPAs into their business terms. Anthropic's DPA forms part of the agreements that reference it, such as its Commercial Terms. OpenAI's is part of its Services Agreement, and customers in the EEA or Switzerland conclude it with OpenAI Ireland Ltd. Mistral's is incorporated by reference. Read the role clauses: some providers act as a controller for certain purposes, such as training models on some plans, so check which purposes apply to yours. ## What should I ask every vendor? - The DPA, and which law governs it. - The sub-processor list with locations, and how changes are announced. - The technical and organisational measures (TOMs) under Article 32. For AI systems, the DSK published separate TOM guidance in June 2025. - For model providers: whether inputs and outputs are used for training, how long they are kept, and where inference runs. - The basis for any transfer outside the EU under Chapter V. BayLDA notes that transfers to the US usually rely on the EU-U.S. Data Privacy Framework. ## What does Sanlabs GmbH commit to in its DPA? Sanbox supports GDPR-compliant use: an Article 28 DPA, processing in EU data centres, and a sub-processor list on request. - **Contract.** Sanlabs GmbH, a company under German law, acts as your processor under a DPA governed by German law, and processes your data only on your documented instructions. - **Location.** Computers, disks, snapshots, files, run records, accounts, logs and monitoring are stored and processed in EU data centres. - **Sub-processors.** Listed, with at least 14 days' notice by email before one is added or replaced; you can object. Some have parent companies outside the EU; transfers outside the EU go only through listed sub-processors, under Chapter V safeguards such as the Standard Contractual Clauses. - **Breaches.** Notification within 48 hours at most, with a target of 24 hours. - **Deletion.** The data processing agreement sets out how quickly data is deleted from live systems, backups and logs after your instruction. - **Help.** Free of charge within the normal scope for data subject requests, security, breach notifications and data protection impact assessments. - **Documents.** The DPA, the sub-processor list and the TOMs come on request from [support@sanlabs.ai](mailto:support@sanlabs.ai), the DPA and the list even before you sign up. Model calls run under your own agreement with your model provider, so that DPA is yours to conclude. [Where do my AI agent's model calls go?](https://sanbox.cloud/kb/eu/where-do-agent-model-calls-go/) lists the EU options. See also the [EU Data Act page](https://sanbox.cloud/data-act/) and the [EU and GDPR post](https://sanbox.cloud/blog/eu-hosted-gdpr-compliant-ai-agents/). ## When Sanbox is not the right fit - **You want a vendor to carry your controller duties.** Which personal data your agents process, the legal basis, which websites and services they may reach, which model they call and whether you need an impact assessment stay with you. - **You want the model and the infrastructure under one contract.** Sanbox doesn't sell or host models, so you need a second DPA with your model provider. ## Related - [Where do my AI agent's model calls go? Claude, OpenAI and Mistral EU options](https://sanbox.cloud/kb/eu/where-do-agent-model-calls-go/) - [Best EU-hosted sandboxes for AI agents in 2026](https://sanbox.cloud/compare/best-eu-hosted-sandboxes-for-ai-agents-2026/) - [How do I give every employee their own AI agent computer?](https://sanbox.cloud/kb/teams/agent-computer-for-every-employee/) - [How do I give each of my users their own persistent agent computer?](https://sanbox.cloud/kb/build/one-persistent-computer-per-user/) - [How do I give an AI agent API keys without exposing them?](https://sanbox.cloud/kb/security/api-keys-your-agent-cannot-see/) ## Sources - [Regulation (EU) 2016/679 (GDPR), Articles 26, 28 and 32](https://eur-lex.europa.eu/eli/reg/2016/679/oj) (checked 2026-11-06) - [BayLDA: KI und Datenschutz](https://lda.bayern.de/de/ki.html), section on AI as a service, in German (checked 2026-11-06) - [DSK: Orientierungshilfe Künstliche Intelligenz und Datenschutz, 6 May 2024](https://www.datenschutzkonferenz-online.de/media/oh/20240506_DSK_Orientierungshilfe_KI_und_Datenschutz.pdf), sections 1.9 and 2.1, in German (checked 2026-11-06) - [DSK: Orientierungshilfe zu empfohlenen technischen und organisatorischen Maßnahmen bei der Entwicklung und beim Betrieb von KI-Systemen, June 2025](https://datenschutzkonferenz-online.de/media/oh/DSK-OH_KI-Systeme.pdf), in German (checked 2026-11-06) - [Anthropic: Data Processing Addendum](https://www.anthropic.com/legal/data-processing-addendum) (checked 2026-11-06) - [OpenAI: Data Processing Addendum, v.010126](https://cdn.openai.com/pdf/openai-data-processing-addendum.pdf) (checked 2026-11-06) - [Mistral AI: Data Processing Addendum](https://legal.mistral.ai/terms/data-processing-addendum) (checked 2026-11-06) ## Next step Talk to us at sales@sanlabs.ai. Sanlabs' forward-deployed engineers build agent use cases with your team, with experience in aviation, pharmaceuticals, supply chain and logistics, and internal admin workflows. How we work: https://sanbox.cloud/kb/teams/forward-deployed-engineering.md Existing customers can use Sanbox from a coding agent: install the CLI with `npm install -g @sanlabs/sanbox-cli@latest`, sign in with `sanbox login`, then run `sanbox skill` for the full agent guide.