Skip to content
CPLT
Request a free call
Menu
All posts
gdprcomplianceopenaiself-hosted-ai

OpenAI DPA and GDPR Article 28: What to Check

A practical review of the contract, API retention, regional processing and the decisions your team still owns before sending personal data to an AI service.

The contract and the configured data flow need to describe the same system.

Stéphane Lepain··Updated ·7 min read

GDPR Article 28(3) says the contract with a processor must set out the subject-matter and duration of the processing, its nature and purpose, the type of personal data, the categories of data subjects and the obligations and rights of the controller. The OpenAI and Anthropic Data Processing Addendums (DPAs) put processing details in Schedule 1 and customer obligations in the main text.

Short answer

  • OpenAI's DPA is part of the OpenAI Services Agreement. Section 1.1 names OpenAI as the processor. Its Schedule 1 gives the duration as the term, plus the time needed afterwards for obligations such as data deletion.
  • Anthropic's DPA is part of its Commercial Terms of Service. Its Schedule 1 says processing runs for the duration of the agreement and leaves the categories of personal data and data subjects to the customer.
  • Signing it is not enough. You still need a lawful basis, the intended account and endpoint, checked retention and processing locations, and your own records.

This is an engineering checklist, not legal advice or a compliance certification. Your DPO or qualified adviser needs to assess the lawful basis, purpose and risks for your use. Sources were checked on 9 September 2026, and both DPA pages again on 20 November 2026. The OpenAI DPA I reviewed is effective 1 January 2026, the Anthropic DPA 24 February 2025; your order form and applicable agreement also matter.

What Article 28 and the DPA cover

GDPR Article 28 also requires the contract to cover documented instructions, confidentiality, security, subprocessors and assistance with the controller's obligations.

In OpenAI's public DPA, section 2.1 says instructions include the agreement, order form and configuration tools. The settings your engineers choose are part of the implementation you need to review.

CheckWhere to startEvidence to keep
Which service and contracting entity apply?Applicable agreement, order form and DPA scopeDated copies and the account using them
What instructions apply?DPA 2.1 and your configurationAllowed use, endpoint list and settings
Who else processes the data?DPA 2.9–2.10 and current subprocessor listRelevant providers and change notifications
How are access, deletion and incidents handled?DPA 2.3–2.8 and 2.11Your access policy, deletion test and incident procedure
What remains your responsibility?DPA 3.1–3.3Notices, authorisations and configuration decisions
Are international transfers involved?DPA section 4 and actual data flowApplicable transfer mechanism and assessment

The DPA describes safeguards. It does not prove that your application uses the intended account, region or endpoint.

No training is not the same as no retention

OpenAI's API data documentation states that API data is not used to train or improve models unless you explicitly opt in.

The default abuse-monitoring period is up to 30 days, with the documented legal and safety exceptions. Check files, conversations and vector stores against their own entries in the endpoint table; one retention number does not describe every feature.

Zero Data Retention and Modified Abuse Monitoring require approval and have limitations. Under ZDR, supported endpoints may change behaviour; other capabilities may still store application state. A store: false setting alone is not evidence that your account has ZDR.

EU storage and regional processing need separate checks

The API documentation distinguishes regional storage from regional processing. Where an eligible region supports processing, supported requests can perform inference there too.

Eligibility, region, project configuration, endpoint and model support still matter. OpenAI distinguishes customer content from system data; residency does not apply to all system data. Review the current limitations for tools and integrations rather than extending a regional promise to every component.

DPA section 4.1 provides for applicable safeguards when EEA or Swiss data is transferred outside those areas, including SCCs or an adequacy decision. SCCs are not a promise that a transfer never occurs. An EU contracting entity is not, on its own, proof of EU-only processing.

What changes if you self-host?

If inference, embeddings and document handling run locally, that route does not send the workload to an external model provider. But backups, monitoring, remote support and administrators may still introduce processors or transfers. Trace those paths before making an EU-only or local-only claim.

For scanned documents, I also check the OCR and document-processing path.

RouteWhat I would verify before use
Local models onlyNo unintended outbound route; operator access; local logs and backups; deletion and recovery tests
External API permittedContract and account; minimum necessary data; endpoint retention; eligible regional processing; tool destinations
Mixed local and externalExplicit allowed routes; sensitive work stays on its approved route even when a larger model might answer better

A practical review before deployment

Write down one representative task and follow its data: upload, extraction, embedding, retrieval, model request, response, logs and backup. Record which system receives each part, who can access it, how long it is retained and how deletion works. Test the configured route rather than relying only on an architecture diagram.

I explain how every result is checked on the proof page, and the acceptance checks for a knowledge assistant include a synthetic source-check example. If you need it implemented, private AI for your company sets out the scoped build and operator handover. A free scoping call is the place to establish whether that work is needed.

Questions buyers ask

Does signing the DPA make our AI application compliant?

No. Review the applicable agreement and your own use, notices, authorisations, configuration and data flow. OpenAI's DPA explicitly assigns configuration responsibilities to the customer in section 3.3.

Does self-hosting the chat interface keep everything local?

No. The model gateway, embedding service, tools and backups determine where data goes. A local interface can still call an external API.

How do I sign the OpenAI or Anthropic DPA?

Both public pages describe acceptance through the main business terms. OpenAI's DPA says the customer agrees "by clicking 'I agree,' accepting the Order Form, or using the Services", and customers based in the EEA or Switzerland contract with OpenAI Ireland Ltd. Anthropic's DPA forms part of its Commercial Terms of Service or another agreement that references it.

Sources and revision

Revision, 9 September 2026: clarified API versus subscription scope, corrected the earlier storage-only description of residency, and removed blanket claims that self-hosting eliminates processor or transfer obligations.

Revision, 15 October 2026: aligned the search listing with the Article 28 and DPA questions and opened with the direct answer. The DPA review and checks are unchanged.

Revision, 20 November 2026: added the Article 28(3) contents, the OpenAI and Anthropic Schedule 1 details and how each DPA is accepted; cut repeated passages.