Data protection in AI: what you can put in the prompt

The most common question in every ahead-future-labs workshop: “Am I even allowed to do this? I dump customer data into ChatGPT, that can’t possibly be legal.”

The short answer: it depends. On which tier, which data, which legal basis. The detailed answer follows now, in a format that lets you decide after 10 minutes what works for you.

The GDPR basics (in 5 sentences)

1. Personal data may only be processed if a legal basis exists (Art. 6 GDPR).
2. For external processing, a Data Processing Agreement (DPA) is required.
3. Without a DPA, transferring personal data to external providers is unlawful, including into “the cloud”, including to the USA.
4. What counts as personal data is broader than most people think: name, email, IP address, device IDs, often also phone number plus industry.
5. For sensitive data (health, religion, trade union membership, sexual orientation), Art. 9 applies, which is stricter.

Which tiers give you GDPR compliance?

Provider	Tier	GDPR-ready?	DPA available?	Data location
ChatGPT (OpenAI)	Free	No	No	USA
ChatGPT Plus	Personal	No	No	USA
ChatGPT Team	B2B	Yes*	Yes	USA + EU options
ChatGPT Enterprise	Enterprise	Yes	Yes	EU region selectable
OpenAI API	API	Yes*	Yes (DPA)	depending on region
Claude (Anthropic)	Free / Pro	Personal: no	No	USA
Claude for Work	B2B	Yes	Yes	USA + EU
Microsoft Copilot (Office 365)	Business	Yes	Yes (Microsoft DPA)	EU region selectable
Google Gemini Workspace	Workspace	Yes	Yes	EU region selectable
Mistral / Aleph Alpha	API	Yes	Yes	EU

“Yes” with an asterisk: additionally switch on privacy settings, disable “Train on my data”.

The simple decision tree

1. Is a DPA in place with the provider?
   ├── No → Don't paste in any personal data. No clauses from contracts, no customer emails.
   └── Yes → continue to 2

2. Is sensitive data involved (health, religion, etc.)?
   ├── Yes → Additional review of the legal basis (Art. 9 GDPR). When in doubt: don't.
   └── No → continue to 3

3. Is the processing compatible with the original purpose of collection?
   ├── No → Obtain the consent of the data subjects or review another legal basis.
   └── Yes → OK, into the prompt with it. But: mind logging and retention periods.

What actually happens in daily practice

In most SMEs, daily practice looks like this:

• Employees use private ChatGPT accounts (“it’s just faster”)
• Customer data ends up in there
• No one has a DPA
• GDPR violation: yes
• Probability that anyone notices: low
• Probability that there are problems if someone does notice: high

The solution is not “ban it and threaten penalties”, that drives shadow IT. The solution is: procure an approved tier, train people, document it.

Three rules of thumb for daily use

1. Anonymize before pasting. Instead of “Maria Huber, mariahuber@firma.at, complained about…” → “A customer complained about…”. Works in 80% of cases.

2. For sensitive material: local model or Enterprise. Personnel files, health data, contracts with clauses → either local (Llama, Mistral self-hosted) or Enterprise with an EU region.

3. Check the output. If the model answers you with “Anna K. lives at Hauptstr. 12”, even when you didn’t enter it, AI can also hallucinate personal data. Never pass it on one-to-one.

Contractual steps for your company

Step	What happens?
Conclude a DPA	With every AI provider you use
Record of Processing Activities (RoPA)	Mandatory document; enter AI tools there
TIA / Data Protection Impact Assessment	For high-risk use cases (HR, applicants)
Document training	Who was trained on which tools and when
Retention periods	Which prompts/responses are stored and for how long

Common misconceptions

“We use Microsoft 365 Copilot, so everything is GDPR-compliant.” Wrong. Microsoft has the DPA, but you have to have concluded it. Also: set the region in the tenant config to EU.

“The AI doesn’t learn from my inputs.” It depends. With most Enterprise tiers: correct. With Free / Plus / Pro: often it does after all, until you manually disable it.

“Pseudonymizing is enough.” Often it isn’t. Pseudonymized data remains personal data if it can be re-attributed. Strict anonymization is the only safe option.

“We host Llama ourselves, so data protection isn’t an issue.” It is. GDPR applies regardless of hosting. What changes: no DPA with an external party is needed. But: RoPA, legal basis, impact assessment, all remain.

---

Disclaimer: This article is editorial and does not replace legal advice. For concrete GDPR compliance in your own company: involve a specialized lawyer or data protection officer.