Using OpenAI / ChatGPT in a GDPR-Compliant Way: Consumer vs. Enterprise vs. API

Using OpenAI in a GDPR-compliant way means deploying only ChatGPT Enterprise/Team or the OpenAI API with a signed data processing agreement (DPA) for personal corporate data, where customer data is not used for model training by default. Consumer ChatGPT is unsuitable for this because a binding DPA as a processor, guaranteed EU data residency and contractual zero-data-retention are missing.

The key question for DACH companies and agencies is not "ChatGPT yes or no", but "which procurement route under which contract". This is exactly where GDPR compliance is decided.

• Consumer ChatGPT (free, Plus): no DPA as a processor, no reliable training exclusion on a contractual basis, no guaranteed EU data residency - off-limits for personal business data.
• ChatGPT Enterprise/Team: operates under the OpenAI DPA, no training on customer data by default, OpenAI as a processor.
• OpenAI API: also under a DPA, max. 30 days retention by default, zero-data-retention (ZDR) available contractually.

Why Consumer ChatGPT Is Out of the Question in a Corporate Setting

For any processing of personal data via an AI system, OpenAI acts as a processor within the meaning of Art. 28 GDPR in a corporate context. An effective data processing agreement (DPA) is therefore mandatory. The consumer plans lack this contract in the form that a controller requires: no documented instruction bindings, no assured data residency, no configurable retention periods.

On top of this comes the technical risk of memorisation. Nasr et al. (2023) demonstrated the extraction of training data - including personal data - from production models such as ChatGPT at a query cost of around USD 200. Anyone who enters customer names, personnel data or health information unprotected into a consumer frontend cannot control what happens to that data.

Hallucinated personal data also remains personal data. This is confirmed equally by the Hamburg ChatGPT complaint and the Garante proceedings against OpenAI - an incorrect but plausible statement about an identifiable person triggers rectification and access obligations.

Consumer vs. Enterprise/Team vs. API in a GDPR Comparison

The following table summarises the differences that are decisive for compliance. All figures are based on internal research and must be re-validated at the time of signing, as provider terms change frequently.

Important: the training exclusion and ZDR are contractual commitments, not statutory minimum standards. Providers typically reserve narrow rights for abuse monitoring, safety review and aggregated analytics, in which they may act as their own controller. These carve-outs should be delineated with a bright line in the DPA.

What the Garante Decision Teaches Us

The leading enforcement example is the decision of the Italian Garante of 2 November 2024 (published 20 December 2024): a EUR 15 million fine against OpenAI. Breaches identified:

• Processing of personal data to train ChatGPT before a viable legal basis was identified - a breach of Art. 5(1)(a), Art. 5(2) and Art. 6 GDPR.
• Insufficient transparency towards users (Art. 12, 13).
• No adequate age verification (Art. 8, 25).
• Late notification of the chat history data breach of March 2023 (Art. 33).
• In addition: a mandatory six-month awareness campaign in the Italian media.

The GDPR fine framework under Art. 83 is up to EUR 20 million or 4 per cent of global annual turnover. By way of comparison: Clearview AI was fined EUR 20 million in Italy for scraping public data (10 February 2022) - publicly accessible does not mean "free to use".

Seven Steps to Legally Compliant Deployment

1. Determine the procurement route. For personal data, only ChatGPT Enterprise/Team or the API. Block consumer access organisationally (policy, and where appropriate technical blocking).
2. Sign and review the DPA. OpenAI DPA (effective from 1 January 2026) with OpenAI Ireland Ltd. as the EEA/CH contracting party; check the "no training" commitment, instruction binding and abuse-monitoring carve-out.
3. Map sub-processors. Add the published sub-processor list (incl. Microsoft Azure as hosting) to the record of processing activities (RoPA); ensure an unbroken DPA chain under Art. 28(4).
4. Configure data residency. Activate EU data residency for eligible Enterprise/Edu accounts or suitable API setups; the default is global routing.
5. Control retention. The API default is max. 30 days, Enterprise stores for the contract term. For sensitive data, agree ZDR contractually; pseudonymise/redact prompt logs before permanent storage.
6. Secure the transfer basis. Use the EU-US DPF, but document standard contractual clauses as a fallback and a transfer impact assessment (as of 2026: the DPF appeal Latombe has been pending before the CJEU since 31 October 2025, subject to change).
7. Document the DPIA and legal basis. Data protection impact assessment in line with the DSK guidance (6 May 2024); record the legal basis per use case - usually Art. 6(1)(f) with a documented legitimate interest, in customer service possibly Art. 6(1)(b).

What You Should NOT Do with Consumer ChatGPT

• Enter customer names, CRM extracts, applicant or personnel data.
• Process special categories under Art. 9 (health, religion, trade union, sex life).
• Upload confidential business documents or entire document sets.
• Prepare automated decisions with legal or significant effect (Art. 22). In addition: the right to an explanation of individual decisions under EU AI Act Art. 86 becomes effective from 2 August 2026 (legally binding date of application, as of 2026).

Worked Example: Retention as a Risk Lever

An agency runs a support assistant for a client via the OpenAI API with 5,000 prompts per day, of which around 8 per cent (400 prompts) contain customer names or contact data. With the API default retention of max. 30 days, at any given time up to 12,000 personal prompts (400 x 30) are held in the processor's storage - each of them within the scope of access and erasure requests. With contractually agreed ZDR plus upstream redaction, this permanent stock drops to nearly zero. This is the difference between an answerable and an unanswerable data subject access request (DSAR).

For Agencies and B2B

Marketing agencies and B2B decision-makers who deploy ChatGPT in client projects are usually controllers themselves, or processors for their clients - and must map the chain downwards cleanly: a DPA with OpenAI, documented data residency, ZDR for sensitive flows and a DPIA template that is reusable per use case. Anyone who sets this up cleanly once turns a compliance risk into a sellable quality feature. Blck Alpaca supports DPA review, sub-processor mapping and DPIAs for AI agents in the DACH region.

Note: This article serves as professional information and does not constitute legal advice. For a legally binding assessment of your specific deployment scenario, please consult a data protection officer or specialised legal counsel.