Insights
Sending business data to an LLM under KVKK and GDPR
An LLM call is a data transfer to another company. Under KVKK and GDPR that makes it a decision about a recipient, a purpose, and a location before it is a decision about a prompt. This page is the engineering side of that decision. It is not legal advice.
Why is an LLM call a transfer?
When your application sends a prompt to a hosted model, the text leaves your infrastructure and is processed by the model provider. If the prompt contains personal data, that provider is processing personal data on your behalf, and if its servers are in another country, the data has crossed a border. KVKK and GDPR each regulate both facts. Which legal basis and transfer mechanism apply is a question for counsel.
What goes into a written data boundary?
The boundary is one page the product team, counsel, and the engineers can all read. It lists each field the feature sends, the provider and the region that receive it, the purpose, and the retention on both sides. Reorigin writes it before the first model call on an AI solutions engagement, because the answers change the architecture.
The most useful line is the list of fields that are not sent. A support summary rarely needs a national ID number or a full address. Removing or masking fields before the call is cheaper and easier to defend than explaining afterwards why they were included.
| Item | Question it answers | Where it is enforced |
|---|---|---|
| Fields sent | Which fields may leave the system | Allow-list in the model gateway service |
| Fields withheld | Which personal data the task does not need | Masking before the call |
| Provider and region | Who processes the prompt, and where | Provider configuration and contract |
| Purpose | What the output is used for | Feature scope and access roles |
| Retention | How long prompts and outputs are kept | Log retention and provider settings |
| Tenant scope | Whose documents can inform an answer | Tenant filter on retrieval |
What does engineering have to enforce?
A boundary on paper does nothing if the code can ignore it. The call to the model goes through one service that applies the field allow-list, so no screen can send a raw record. Tenant scoping applies to retrieval as it does to any other query: one tenant's documents are never used to answer another tenant's question. Prompts and outputs are logged with the same access controls and retention as the data they contain.
Provider settings are part of the system. Whether the provider stores prompts, for how long, and whether it uses them to train models are contract and configuration facts that should be recorded next to the boundary, with the date they were checked.
What will we not claim?
We do not claim that a hosted region, a signed data processing agreement, or a masking step makes a feature compliant on its own. We do not tell you which legal basis applies. We build the controls that let your counsel's decision be carried out and shown to work. The wider picture for a SaaS product is in KVKK and GDPR for SaaS built in Turkey.
See AI solutions for how an AI engagement is scoped.
Sources and further reading
- Regulation (EU) 2016/679 (GDPR) — official text on EUR-Lex.
- Law No. 6698 on the Protection of Personal Data (KVKK) — official English text, Personal Data Protection Authority of Türkiye.
- Standard contractual clauses — European Commission, one mechanism for transfers outside the EU.
FAQ
Can a business send personal data to an LLM under GDPR or KVKK?
That depends on the legal basis, the provider's role, and where the data is processed, which your counsel decides. This page is not legal advice. Engineering can limit what is sent and make that limit enforceable.
What is a data boundary for an AI feature?
A written record of which fields may be sent to a model, to which provider and region, for what purpose, and how long prompts and outputs are kept. Reorigin writes it before the first model call.
Does choosing an EU region make an LLM feature compliant?
No. Region is one fact among several. The provider's role, the contract, retention, and what the prompt contains all matter. A region setting does not replace a legal review.
How do you stop one customer's data reaching another through an AI feature?
Retrieval is scoped to the tenant the same way every other query is. One tenant's documents are never used to answer another tenant's question, and a test proves it.