Can you run ChatGPT in-house? A decision guide to KVKK and on-premise LLMs

Short answer: ChatGPT itself cannot be installed on a company server; "an in-house ChatGPT" usually means one of three different options: an enterprise service from a provider abroad, a model API in the cloud, or an open-weight model installed on the company's own infrastructure. From the perspective of the Turkish data protection law (KVKK), the deciding question is whether personal data leaves the country. If it does, the transfer rules in Article 9 of the law apply; if it does not, the burden shifts to the company that installs and runs the system.

The question is not asked for nothing. According to TÜİK's AI Statistics published on 2 October 2026, among enterprises that are considering AI but do not yet use it, 66.4% cite legal uncertainty and 65.4% cite data protection and privacy concerns as reasons. This post is a framework for settling the decision around those two concerns. It is not legal advice; assess your own situation with your KVKK counsel.

First step: which data will be sent?

The decision starts from the data, not the model. Splitting what will be sent to the assistant into three classes shortens most discussions:

• Personal data: Information about customers, employees or individuals at suppliers. KVKK applies directly.
• Trade secrets and confidential information: Formulations, prices, contracts, tender files. Not the subject of KVKK, but a leak is at least as costly.
• General technical content: Public standards, the product catalogue, general correspondence drafts.

For the third class a cloud service is usually fine. If the first two classes are the assistant's main use, the differences between the options below become clear.

Three options

	Enterprise service (abroad)	Cloud model API	In-house installation
Where data is processed	On the provider's infrastructure	In the chosen cloud region	On the company's own infrastructure
KVKK cross-border transfer	Usually yes	Depends on region	No
Model quality	Latest commercial models	Latest commercial models	Open-weight models
Set-up and maintenance load	Low	Medium (integration in-house)	High
Cost structure	Subscription per user	Pay per use	Hardware and operations

In practice most organisations do not stay with a single option: a cloud service for general content, an in-house installation for the assistant that works on personal data and confidential documents.

If data leaves the country: the order under KVKK

The Generative AI guide published by the Turkish Personal Data Protection Authority on 24 November 2025 addresses this case directly: if the service of a provider established abroad is used and data is transferred abroad, Article 9 of the law and the transfer regulation of 10 July 2024 must be complied with. Following the amendment in force since 1 June 2024, the order in that article is:

1. Adequacy decision: A decision by the Board for the country, sector or international organisation the data is transferred to. As of November 2026 we could not see any adequacy decision in the Board's published decision lists.
2. Appropriate safeguards: Where there is no adequacy decision, in order: agreements between public bodies, binding corporate rules, standard contracts and written undertakings. For the private sector the most practical route is the standard contract.
3. Exceptions for incidental transfers: Only if no appropriate safeguard can be provided either, and the transfer is incidental. Explicit consent is one of these exceptions; it is not a transfer route on its own.

Two details are often missed. First, a standard contract must be notified to the Authority within five business days of signing; the administrative fine range for failing to notify in 2026 is TRY 90,308 to TRY 1,806,177. Second, the regulation defines an "incidental" transfer as one that is not regular, takes place once or a few times, and is not part of the ordinary course of business. An assistant that employees use every day hardly fits that definition, which is why solving the problem with explicit consent usually does not work.

Transfer is not the only issue

The guide covers obligations across the life cycle that apply even if the data stays in the country. The ones that create the most work in practice:

• Determining roles: For each processing activity, it should be written down where the company is the data controller and where the provider is the data processor.
• Notice: Processing for using the system and processing for training the model should be explained separately. Users should be told that they are interacting with an AI.
• Data subject rights: According to the guide, rights also extend to training data, user prompts and outputs. That is why it matters where prompts are stored, and for how long.
• Security: The technical and administrative measures in Article 12 of the law apply to these systems as well.

Technical components of an in-house installation

Keeping data inside is half the problem. The other half is building a system that is actually usable and secure:

• Model selection and Turkish evaluation: Open-weight models vary widely in how well they handle Turkish. We choose based not on general benchmark tables but on an evaluation set built from the company's own documents; we explained why a separate set is needed in our post on Turkish evaluation sets.
• Hardware sizing: Model size, quantisation level, the number of concurrent users expected and the acceptable response time determine the GPU memory required. This calculation is done before installation.
• Document permissions: If the assistant works on company documents (RAG), a user should only get answers from documents they can already access. If permission checks are not done in the retrieval layer, the assistant effectively removes access restrictions.
• Logging and retention: Who asked which question, when, and from which sources the answer came. These records are needed both for security and for data subject requests; the retention period should be set from the start.
• Updates: A new model version should not go live before being measured on the same evaluation set.

We describe our work on these topics on the Generative AI and LLM and AI Security pages.

Why a ban does not work

In organisations that ban AI use without offering a controlled channel, use often does not stop; it moves to employees' personal accounts. That is the worst scenario, with no contract, no records and no access control on the company's side. The realistic approach is to write down which data may be processed through which channel and to offer an approved tool for each class.

Frequently asked questions

Can we install ChatGPT on a company server? No, the weights of the models behind ChatGPT are not distributed. What can be installed on company infrastructure are open-weight models released by various developers, OpenAI among them. Their performance in Turkish varies by model, so they need to be measured on your own data.

We use an enterprise plan and our data is not used for training. Does that solve the KVKK issue? Not being used for training is an important safeguard, but it does not remove the cross-border transfer question. If data is processed abroad, a transfer route under Article 9 is needed.

Is explicit consent from employees enough? Usually not. Explicit consent is provided for incidental transfers only; a regularly used service does not fit that definition.

Is an in-house installation always more secure? The data does not leave the country; but responsibility for access control, logging and updates moves entirely to the company. If these are not set up, an in-house system carries risk too.

Sources

• KVKK, Generative AI and the Protection of Personal Data Guide (in 15 Questions), 24 November 2025 (Turkish)
• Law No. 6698 on the Protection of Personal Data (Turkish)
• Regulation on the Procedures and Principles for Transferring Personal Data Abroad, Official Gazette 10 July 2024, No. 32598 (Turkish)
• KVKK, administrative fine amounts for 2026 (Turkish)
• TÜİK, AI Statistics 2026, 2 October 2026 (Turkish)