A data processing agreement (DPA, in German Auftragsverarbeitungsvertrag or AVV) is the contract under Article 28 GDPR that regulates how an AI vendor (the processor) handles personal data on behalf of your organization (the controller). For EU AI chat platforms, three clause families fail more often than the rest: audit rights, sub-processor disclosure, and deletion. A DPA missing one of these three is not Article 28 compliant, regardless of marketing language. This checklist gives your DPO 12 specific questions to score every vendor DPA before you sign.

Vendors with a strong DPA welcome the questions. Vendors with a weak DPA push back, redirect to marketing materials, or quote standard industry terms. Push-back is a signal. The strongest DPAs in our EU ChatGPT alternative for enterprise comparison (Teamo AI, LangDock, Aleph Alpha) hand over the full DPA text within 24 hours of request and walk you through Article 28 clause-by-clause on a procurement call. The weakest hide behind available under NDA or enterprise tier only. For the broader compliance frame, see the EU AI Act + GDPR small business playbook.

Article 28GDPR clause that defines DPA requirements
12must-have clauses every EU AI chat DPA should pass
3clauses where vendor DPAs most often fail (audit, sub-processor, deletion)
20M EUR or 4 %max GDPR fine for processor DPA non-compliance (Article 83)

Why the DPA Matters More for AI Chat Than Other SaaS

Three reasons AI chat DPAs need stricter scrutiny than ordinary SaaS DPAs. One: AI training as a hidden processing purpose. Customer prompts and responses can be used to train, fine-tune, or evaluate models. Without an explicit contractual prohibition, this is processing for the vendor's own purposes (Article 6 lawful basis required) and turns the vendor from processor into joint controller, which changes the entire DPA structure. Two: model-vendor sub-processing. Most EU AI chats are gateways routing to upstream model vendors (OpenAI, Anthropic, Google, Meta). Each upstream is a sub-processor under Article 28(2). The DPA must disclose them, give you veto rights on changes, and contractually flow your no-training requirement down to them. Three: log retention asymmetry. AI chat generates extensive logs (prompts, responses, metadata, usage analytics). Vendors often retain these for service improvement which becomes processing for their own purposes if not contractually capped.

An ordinary SaaS DPA covers data storage, access, and deletion. An AI chat DPA needs all of that plus explicit handling of training, sub-processor flow-down, log retention, and disclosure of which exact model handled which exact request (for Article 50 transparency). The 12 clauses in the next section cover all of these specifically.

The 12 DPA Clauses Every EU AI Chat Vendor Must Pass

1

1. Vendor is incorporated in the EU and contracts under EU jurisdiction

The legal entity that signs your DPA must be EU-headquartered and contractually subject to EU law. US-headquartered vendors with EU subsidiaries can still be compelled under the US CLOUD Act regardless of which entity signs. Verify the Handelsregister/Firmenbuch entry of the contracting party, not the marketing copy.

2

2. All processing happens within the EU/EEA, no third-country transfer

Request bodies, model inference, logs, metadata, and backups all stay in EU/EEA infrastructure. If the vendor uses a global CDN or backup region, those transfers must be specifically disclosed and either eliminated or covered under SCC plus appropriate safeguards. The vendor should publish a detailed data flow diagram in their trust center.

3

3. Customer data is contractually excluded from any model training, fine-tuning, or evaluation

Explicit clause that the vendor will not, and will not permit any sub-processor (including upstream model vendors) to, use customer prompts, responses, or derived metadata for training, fine-tuning, or human evaluation. Marketing language like we respect your data is insufficient, the contract must be specific.

4

4. All sub-processors are disclosed by name with their roles, including upstream model vendors

Article 28(2) GDPR requires general or specific authorization in writing. Strong DPAs publish a current list of all sub-processors (cloud provider, model vendors, monitoring tools, etc.) with their geographic location and processing purpose. Vendors that hide this list for competitive reasons are non-compliant.

5

5. Sub-processor changes require advance notification with veto right

When the vendor wants to add or replace a sub-processor, you must be notified in advance (typically 30 days) with the right to object. If you object and the vendor proceeds anyway, you must have the right to terminate without penalty. Many DPAs have notification but no veto, this fails Article 28(2).

6

6. Audit rights are real (on-site, third-party, or detailed annual report acceptable)

Article 28(3)(h) requires audit rights. Strong DPAs offer three options: customer-led on-site audit (rare but legally required option), third-party audit report (e.g., SOC 2 Type II), or detailed annual self-assessment with right to deeper inspection on cause. Vendors that cap audits at remote questionnaire only fail the article.

7

7. Breach notification within 72 hours of vendor awareness

Article 33 GDPR requires breach notification to the supervisory authority within 72 hours of awareness. The DPA must mirror this with vendor-to-customer notification within the same window, including all information you need to make your own Article 33 notification (nature of breach, data categories, approximate number of records, likely consequences).

8

8. Deletion on contract end covers all copies including backups within defined window

Article 28(3)(g) requires deletion or return of all personal data at the end of services. Strong DPAs commit to deletion within 30-90 days, including all backups, with a written certification of deletion provided to you. Vendors with vague data may be retained for legal/operational reasons fail this article.

9

9. Log retention is contractually capped (typically 30 days for full content, 90 days for metadata)

AI chat generates extensive logs. The DPA must specify maximum retention periods for prompts, responses, metadata, and usage analytics. Strong DPAs cap full content at 30 days and metadata at 90 days. Vendors that retain indefinitely for service improvement need explicit contractual change before signing.

10

10. Article 50 EU AI Act transparency disclosure UI is built-in or contractually committed

From August 2026, users must be informed they are interacting with AI. Strong vendors ship this disclosure UI by default (EU-built vendors are typically Article 50 compliant from launch). Vendors that have not yet built it should commit to delivery before the August 2026 deadline as a contractual obligation, with breach consequences if missed.

11

11. Per-message audit log is accessible to your compliance team

For investigations, regulatory inquiries, or internal compliance audits, you need access to a per-message audit log showing who asked what, which model handled it, when, and (where applicable) which integrations the AI used. This goes beyond Article 28 minimums but is increasingly expected by procurement teams in regulated industries.

12

12. Liability cap is at least 12 months of fees (or higher for breach involving sensitive data)

Many SaaS DPAs cap vendor liability at 1 month of fees, which is meaningless for a serious data breach. Demand at least 12 months of fees as the cap, with higher cap or no cap for breaches involving special categories of personal data (Article 9 GDPR). Strong vendors agree to 24 months or more.

Three quietly missing clauses kill most vendor DPAs: weak audit rights (capped at remote questionnaire), incomplete sub-processor disclosure (model vendors hidden), and vague deletion language (no specific window or backup coverage). Cross-check these three first, before reading the rest. If any of the three fail, the DPA is non-compliant regardless of how good the other clauses look.

Score your AI vendor against the 12 DPA clauses in 7 minutes

Independent assessment maps your current vendor DPA against the 12 must-have clauses. Output is a procurement-grade scorecard your DPO can use directly. EU-hosted, free.

Try It Free

How to Run the DPA Review with Vendors

Three steps that work in practice. One: request the full DPA text up front, not after the pilot starts. If a vendor only releases the DPA after a signed pilot agreement, walk away. Strong vendors send the DPA within 24 hours of request, weak ones make you ask twice and then send a heavily-redacted version. Two: have your DPO score the DPA against the 12 clauses in this checklist, with traffic-light status (green for compliant, yellow for conditional, red for failing). Bring the scorecard to the next vendor call and walk through it together. Strong vendors engage clause-by-clause, weak vendors deflect or escalate. Three: never sign a DPA where any of clauses 1-3, 4-5, or 6-8 are red. Those three groups are the structural minimums. Other clauses can be negotiated post-signature, the structural ones cannot.

For the broader compliance picture (DPA plus AI Act plus Betriebsrat), see the GDPR + AI Act compliance checklist and the GDPR + AI Act compliance software comparison for tooling that monitors DPA adherence over time.

Bonus: ask about concrete safety mechanisms beyond the DPA. Strong vendors implement runtime guards that go beyond contract clauses. Example: an AI hallucination guard rejects bulk-update tool calls where the AI's claimed prior value does not match the actual stored value, preventing the AI from hallucinating updates across long agent loops. Vendors with this kind of mechanism in place have thought about the real failure modes of AI in production, vendors that only point to their DPA have not. Ask specifically: what runtime guards prevent AI hallucination on bulk operations?

Run the AI Readiness Assessment alongside the DPA review

Independent baseline of which teams are ready for the AI rollout. Lets you scope the DPA negotiation specifically: which data categories, which sub-processors, which retention windows are actually needed.

Try It Free

The DPA is the contract that makes EU AI sovereignty real

Marketing language is not contractual. A vendor saying GDPR-compliant or EU-sovereign on their landing page does not bind them to anything. The DPA is the contract. If the DPA does not contain the 12 clauses in this checklist, the marketing language is unenforceable.

Three clause families fail most often: audit rights, sub-processor disclosure, deletion. Cross-check these three first. If any fails, the DPA is structurally non-compliant.

Strong vendors welcome the scrutiny. Teamo AI, LangDock, Aleph Alpha PhariaAssistant, Mistral Le Chat Enterprise, and DeutschlandGPT all engage clause-by-clause on procurement calls. Vendors that deflect standard industry terms are signaling weakness, not strength.

DPA per AI vendor: where you get it and what it says

The table below answers the question the checklist above cannot: for each vendor companies actually ask about, is there a data processing agreement at all, for which tier, how it is executed, where the data is processed, on what transfer basis, and whether training on your inputs is excluded. Rows marked "verify" mean the vendor publishes the document but not every detail; request the full text before signing. If the DSB is the one asking, the 12 questions your DSB asks before approving AI turns this table into the approval file. For the most common case, a ChatGPT Business workspace without EU residency, see ChatGPT Business and GDPR.

VendorDPA for which tierHow it is executedEU processingTransfer basisTraining exclusion
OpenAI (ChatGPT Business, Enterprise, Edu, API)Business, Enterprise, Edu, API; none for Free and Plusself-serve online form on openai.comEU data residency on Enterprise and API projects; not on BusinessDPF certification and SCCsyes, contractual for all business tiers
Anthropic (Claude Team, Enterprise, API)Team, Enterprise, APIincorporated into the Commercial Terms with SCCs; download a copy for the fileUS processing; EU hosting only via AWS Bedrock or Google Cloud, where the hyperscaler DPA governsSCCs in the terms; TIA required Yes
Microsoft 365 Copilotall commercial tenantsautomatic via the Products and Services DPA, nothing to signEU Data Boundary; Anthropic models inside Copilot currently excludedDPF and SCCsyes; prompts, responses and Graph data are not used to train foundation models
Google Gemini for Workspaceall Workspace business editionsautomatic via the Cloud Data Processing Addendum, incorporated into the agreementEU data regions selectable in the Workspace contractDPF and SCCsyes for business customers
Mistral (Le Chat Team, Enterprise, La Plateforme)commercial customers; published in the legal centreon acceptance of the terms; verify product scope in the DPA textFrance and EU hostingno third-country transfer for EU hostingverify per tier; zero-retention options exist
Langdockall paid workspacesdownload from the trust centre, countersignEU hosting; model providers configurable, EU routing availabledepends on the model provider chosenyes, customer data never used for training
meinGPTall business planson request from the vendorGerman and EU hosting; OpenAI via EU endpoint with storage offSCCs and DPF for US sub-processors Yes
Teamo AIevery workspace, including the 14-day trialEU AVV included at signup; countersigned copy on requestEU hosting, self-hosting optionalper model provider, EU providers selectableyes, contractual per provider