GDPR-compliant AI automation: a checklist for B2B teams
03 November 2026
Most B2B automation touches personal data, even when nobody planned it. An inbound email has a sender’s name and signature. A CRM record has a phone number. A LinkedIn conversation has a profile. Once an AI model reads that data, GDPR applies to the whole chain.
This is a working checklist from the technical side, for teams that want to automate without creating a problem for their data protection officer. It is not legal advice. Your DPO or counsel decides what applies to your company.
How does GDPR affect AI systems?
GDPR does not have a separate chapter for AI. The same principles apply, and AI makes some of them harder to meet:
- Lawful basis (art. 6). You need a reason to process the data. For B2B outreach and email handling, that is often legitimate interest, which requires a documented balancing of interests. Cold B2B email is also governed by national ePrivacy rules (in Germany, UWG §7), which can require prior consent regardless of GDPR.
- Purpose and minimisation (art. 5). The system should receive only the data the task needs. An email classifier does not need the sender’s full history.
- Transparency (arts. 13 and 14). People should be told their data is processed, including by automated means.
- Storage limitation (art. 5). Logs, conversation histories and execution data need a retention period.
- Security (art. 32). Access control, encryption and logging in proportion to the risk.
- Automated decisions (art. 22). People have the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them. Drafting a reply that a person approves is not that. Automatically rejecting an application or refusing a contract would be.
How do I make AI GDPR compliant?
Work through eight points, in this order.
1. Map the data flow. For each automation, write down what personal data enters, where it goes, which systems and models touch it, and where it is stored. If you cannot draw it, you cannot defend it.
2. Decide where the data lives. Choose an EU region for hosting and storage. Where a component runs outside the EU, check the transfer mechanism under Chapter V (arts. 44 to 46) and document it.
3. List every processor. The automation platform, the hosting provider, the AI model provider and any vector database are all processors if they handle personal data on your behalf. Each needs a contract that meets article 28.
4. Limit what the model sees. Strip or mask what the task does not need before it reaches the model. Pass an order number, not the customer’s whole profile.
5. Set retention. Decide how long you keep workflow executions, chat histories and drafts. Delete on schedule, automatically.
6. Keep a human where it matters. Anything with legal or similarly significant effect on a person goes through a human who can change the outcome. Where the agent drafts and a person approves, article 22 is generally not engaged, but check your case.
7. Log every action. An audit trail of what the system read, decided and sent lets you answer a data subject request or an incident review. It also records who approved what.
8. Test before launch, and after every change. A fixed set of real cases, re-run on each release, catches the regressions that would otherwise reach live data. Document the results.
Add the records of processing (art. 30) and, for higher-risk uses, a data protection impact assessment (art. 35). Both are paperwork, and both are easier if the technical steps above exist first.
Is ChatGPT GDPR compliant for business data?
It depends on which product and which contract. The consumer version of a chatbot is not set up for business data: you have no processor agreement, and the terms may allow your inputs to be used for improvement. Business and API offerings from major providers typically include a data processing agreement, retention controls and an option not to train on your data, and some offer EU data residency.
Do not take that as a yes. Read the current terms of the exact plan you would use. Check four things: whether a DPA is available and signed, where data is processed and stored, how long inputs are retained, and whether your data is used for training. If the answer to any of these is unclear, do not send personal data to it.
The same questions apply to every provider, not only to ChatGPT.
Do we need a data processing agreement with the AI provider?
Yes, if the provider processes personal data on your behalf. Article 28 requires a contract between controller and processor that sets out the subject matter, duration, nature and purpose of the processing, the type of data, and the processor’s obligations, including security, sub-processors and deletion or return of data at the end. Most major providers publish a standard DPA you can sign. Look at the sub-processor list and the transfer clauses as well.
If you avoid sending personal data to the model at all, because you masked it first, the question gets smaller. It does not disappear, because the surrounding systems still process data.
When private deployment is worth it
Private deployment means the automation platform, the database and the application run on infrastructure you control, in an EU data centre, instead of in a shared cloud service. It costs more effort. It is worth considering when:
- the data is sensitive, such as contracts, health or financial details;
- you need full control over retention and deletion;
- a customer or public body requires it in procurement;
- volumes are high enough that the control pays for itself.
For lower-risk workflows, a hosted service with a signed DPA and an EU region is often enough. Our own approach is to offer both: private deployment where the data calls for it, and hosted components where it does not. For n8n in particular, self-hosted versus cloud lays out the trade-offs.
What this looks like in practice
When we built a knowledge assistant for a German home-security manufacturer, the retrieval layer worked only on the client’s own documents. A calibrated threshold sends any question the documents cannot answer to a person, with the conversation attached. Releases are tested on a fixed evaluation suite before they go live. Limited scope, human handover and documented testing are also the controls a DPO asks about.
If you want to know where your current or planned automations stand against this list, the free automation audit starts by mapping your workflows, which is also step one of this list. For the systems that run documents and back-office data, see the AI Operations Agent.