Blog
Turkey's KVKK guides on generative and agentic AI: 7 checks before you share
What Turkey's KVKK generative AI guide (15 questions) and its agentic AI guide mean for teams pasting documents into models, plus a 7-point pre-share checklist.
Published 7 min read
Turkey's data protection authority, the Personal Data Protection Authority (KVKK, Kişisel Verileri Koruma Kurumu), has published two guides for organizations that use AI. The first is a question-and-answer document, the Guide on Generative AI and the Protection of Personal Data in 15 Questions (in Turkish). The second, published on 12 March 2026, is the Guide on Agentic AI Systems. This post explains, for international readers, what the two guides emphasize for teams that paste documents into models, and ends with a seven-point checklist you can apply before sharing any file with a hosted assistant.
A caveat first: the summary below is based on the published texts but does not replace them. The authority may update the guides; consult the current versions when making decisions. If your organization processes data of people in Turkey, or works with Turkish partners, the guides are relevant even if you are not based there.
Background: what KVKK is
Law No. 6698 on the Protection of Personal Data (KVKK) is Turkey's data protection statute. It predates the GDPR by a few months and follows the same architecture: lawful bases, principles (lawfulness and fairness, accuracy, purpose limitation, data minimization, storage limitation), data subject rights, controller and processor roles, and restrictions on international transfers. The authority publishes guidance to explain how these rules apply to new technologies; the two AI guides are the latest examples. For a comparison of KVKK and the GDPR in the context of prompts, see GDPR, KVKK and LLM prompts.
What the generative AI guide emphasizes
The 15-question guide addresses the use of generative tools (models that produce text, images or code) and how the law's principles apply. For teams sharing documents, these are the points that matter most:
- Data minimization. Only the data necessary for the purpose should be given to a model. If you want a contract summarized, the parties' real names, addresses and ID numbers are usually not necessary.
- Purpose limitation. Personal data collected for one purpose cannot simply be reused for a new purpose the data subject would not expect, such as training a model. A provider's use of prompt data for training is therefore something the controller must assess, not ignore.
- Transparency and information duties. Data subjects should be informed that their data is processed with an AI tool, and the controller should be able to explain how the tool works and where the data goes.
- Accuracy and the hallucination problem. The guide notes that generative models can produce content that is not true. If a model generates a false statement about a person, that touches the principle that data must be accurate and up to date; outputs should not be recorded without human review.
- International transfers. Using a hosted model in most cases means transferring personal data to servers abroad. The law's transfer rules (explicit consent, adequacy decisions, standard contractual safeguards and similar mechanisms) apply.
- Security measures. The controller must evaluate the tool's technical and organizational safeguards and keep access, retention and deletion under control.
What the agentic AI guide adds
Agentic systems are AI agents that pursue a goal through multi-step processing rather than a single answer: they call tools, read and write files and talk to other systems. The 12 March 2026 guide addresses the additional risks this creates. For document sharing, the relevant points are:
- Profiling risk in multi-step processing. An agent may combine the document you gave it with other sources and draw inferences about individuals. That is a much broader processing activity than a single summary request and may engage the rules on profiling.
- Control and traceability. It should be possible to record which data the agent processed at which step, and human oversight should not be removed from the chain.
- Purpose drift. Data provided for one task may end up serving another purpose in the agent's intermediate steps. The guide expects the controller to design the chain deliberately.
- Third-party tools and transfers. Every tool and API the agent calls is potentially a new processor or a new transfer point.
- Deletion and retention. Personal data can linger in the agent's intermediate outputs and memory; deletion procedures must reach those layers too.
The shared message of both guides is simple: personal data given to a model is still personal data, and every principle in the law continues to apply. The tool being "AI" does not reduce the obligations; with agents, obligations grow as the chain grows.
What this means for document sharing
In practice a team faces one question: "Can I give this file to the assistant?" The guides give two answers. First, yes, if your lawful basis, information duties and transfer safeguards are in place. Second, and easier: if you remove the personal data that the task does not need, most of the questions disappear.
| Principle | Document shared as is | Document pseudonymized first |
|---|---|---|
| Data minimization | unnecessary identifiers transferred | only the necessary content remains |
| Purpose limitation | provider's use must be assessed separately | lower risk, since no identifiers are present |
| International transfer | a transfer of personal data | if the mapping table stays local, the transferred content carries no identifiers |
| Accuracy | model may generate falsehoods about a real person | output is about "Person 1"; you control the mapping |
| Profiling (agents) | agent can link real people | agent works with labels |
| Deletion | depends on provider retention windows | no personal data in the file |
Pseudonymization does not take the data outside the law; as long as you hold the mapping table, the data is still personal data for you. But it substantially lowers the risk for the content that reaches the provider. For the terminology, see pseudonymization vs anonymization vs redaction.
The 7-point pre-share checklist
Answer these seven questions before giving a document to a hosted model or an agent:
- Does the task need real names? Summaries, translations, classification and drafting usually do not. If not, anonymize.
- Does the document contain special-category data? Health, criminal records, biometric data and similar content may need more than a label; consider removing that section entirely.
- Where does the data go? Are the provider's server location and transfer safeguards (standard contracts, undertakings, explicit consent) known? If not, assume an international transfer.
- Does the provider use this content for training? Check the current settings of the plan you use; for consumer plans the default is often yes. Business and API plans usually offer different commitments.
- Will the agent call other tools? If you use an agentic system, limit which tools and data it can reach, and know what it can do beyond the step where you handed over the document.
- Who validates the output? If the model's output will become a record about a person, human review is required. A wrong output can breach the accuracy principle.
- Where will the mapping table and the original be stored? If you pseudonymized, keep the table linking labels to real values only on your device, and never paste it into the same chat.
Do the anonymization on your device
For the first item on the checklist, Unformation anonymizes the document in your browser. The file does not go to any server; the tool shows "0 network requests" while processing and works with Wi‑Fi off. The flow is four steps: upload, choose a rule (labels, redaction, fake values, masks or hashes), review the findings, download. Turkish national ID numbers, tax numbers and IBANs are validated by checksum; for Turkish names the optional multilingual model runs in your browser. Formatting is preserved and the mapping table stays local as CSV or JSON. See how it works and security for details.
Do it now
Before your next document goes to an assistant, drop it into Unformation and run the seven questions. Scenarios and examples for legal teams are on the legal use case page.