“We are GDPR compliant” is a claim. A DPA is the document that turns it into obligations you can enforce.
How it works
GDPR Article 28(3) says processing by a processor “shall be governed by a contract or other legal act” that is binding on the processor and sets out the subject matter, duration, nature and purpose of the processing, the type of personal data and the categories of data subjects. The contract must stipulate that the processor:
| Point | Obligation |
|---|---|
| (a) | Processes personal data only on documented instructions, including on transfers to third countries |
| (b) | Ensures authorised people are bound by confidentiality |
| (c) | Takes the security measures required by Article 32 |
| (d) | Follows the rules for engaging a subprocessor |
| (e) | Helps the controller answer data subject requests |
| (f) | Helps with Articles 32 to 36 (security, breach notice, impact assessments) |
| (g) | Deletes or returns all personal data at the end of the service, at the controller’s choice |
| (h) | Provides the information needed to demonstrate compliance and allows audits |
Vendors usually publish a DPA that is part of their terms. Mistral’s, for example, says it “is incorporated in the Agreement by reference”, names the customer as controller and Mistral as processor, gives ten days to object to a new subprocessor, makes personal data inaccessible 30 days after access ends, and promises breach notice “without undue delay”. Each vendor’s numbers differ, so read the one you sign.
If the provider is outside the EEA, the DPA often brings in the Commission’s standard contractual clauses. The Commission describes them as pre-approved clauses for transfers to third countries, modernised on 4 June 2021.
Why it matters
- Retention and training in writing. A help-center page can change. The DPA binds the provider; see LLM API data retention compared.
- Deletion at the end. Point (g) is how you get data back or destroyed when you switch providers.
- It does not replace minimisation. Strip PII the model does not need before the request.
The GDPR checklist for LLM APIs maps 15 questions to these points.
On BroadNet Inference
BroadNet does not publish a DPA. Its Terms and privacy statement do not mention one, and neither names the GDPR. What they do state: the gateway keeps no prompt or output content, inference data is not used to train, fine-tune or evaluate models, a current subprocessor list is available to customers on request, and processing locations for residency requirements are agreed before data is sent. Private endpoints and reserved capacity follow a separate written agreement.
If you process EU personal data, ask BroadNet for a DPA through the contact page before you go live, as you would with any provider. This page is not legal advice.