KVKK and AI: What Company Data Can Employees Share with ChatGPT, Claude and Gemini? (2026)

The question I hear most in training sessions is: "Can I paste our customer list into ChatGPT?" In most companies the honest answer is unclear. Employees do not know the rule, and managers have not said clearly what is off limits. So either everyone pastes everything, or the tools get banned and usage disappears from view. In this article I explain, from a Turkish data protection (KVKK) angle, which data can go into AI tools, which cannot, and how I handle the gray area in between. This is not legal advice. For decisions specific to your company, ask your legal counsel.

If you want to write the full rulebook, see my article on shadow AI to understand how uncontrolled usage appears in the first place. This piece stays focused on data and KVKK.

What data can employees share with ChatGPT, Claude and Gemini?

Employees can share anything that contains no personal data and carries no confidentiality for the company: public information, published content, general questions and drafts that include no names, phone numbers or ID numbers. Text that does contain personal data should go in only if one of two conditions is met: the data has been anonymized, or the company has an enterprise contract, a legal basis and an approved tool. Trade secrets, contract texts and unreleased financial statements also should not be shared without enterprise safeguards. The simple test I use is this: if a stranger read this text, could the company or a person be harmed? If yes, the text should not enter a chat box without an approved tool and a written rule.

This test does not depend on the brand of the tool. ChatGPT, Claude and Gemini all have data settings that vary by account type, but the rule is the same: look at the data first, then at the tool.

What counts as personal data and special category data under KVKK?

Under Turkey's Personal Data Protection Law No. 6698 (KVKK), personal data is any information relating to an identified or identifiable natural person. Names, phone numbers, email and home addresses, ID numbers, IP addresses and a customer note that can be tied to one person are all examples. Special category data is protected more strictly: health information, biometric and genetic data, religion, union membership and criminal convictions fall into this group. A common misconception I see in training is treating personal data as just a name. A few details combined, such as "the 42 year old patient at our Kadıköy branch," can make a person identifiable. The rule therefore looks not at the type of information but at whether it can be tied to a person.

I teach employees two questions. First, "Is this information linked to a person?" If yes, the second is "Does the company have a basis to give this data to this tool?" An employee who cannot answer the second question should never decide alone.

What does the KVKK Authority's generative AI guide say?

In November 2025 Turkey's Personal Data Protection Authority published a guide titled "Generative AI and the Protection of Personal Data (in 15 Questions)." The guide explains how generative AI systems work and what risks they carry, assesses personal data processing in these systems under Law No. 6698, and includes recommendations for individuals and parents. It also covers how to determine the data controller and the data processor, compliance with the Law's principles and the conditions for lawful processing. I do not restate it article by article here because the text is published on the Authority's website and is worth reading in full: the KVKK generative AI guide (in Turkish). Your job as a company is to apply it, together with your legal counsel, to your own use cases.

Why is putting data into an AI tool a transfer abroad question?

Most free and individual chat tools run on infrastructure outside Turkey. When an employee pastes customer data into one of them, personal data may have been transferred abroad on behalf of the company, which is the data controller. Under KVKK, transfers abroad are subject to their own conditions, and the Authority's guide addresses this topic too. Where the provider processes the data, which contractual safeguards it offers and what legal basis the transfer rests on are questions the company must answer at the organizational level. Leaving this to each employee's judgment is not fair to them. The company should decide which tool may handle which data and turn that decision into a written rule.

What is the difference between an enterprise plan and a personal account?

Enterprise plans usually differ from personal accounts in three ways: administrator controls, contractual data commitments and how inputs are used for model training. Most providers state that enterprise plan inputs are not used for training by default. On individual plans this often depends on a setting, and an employee can forget to switch it off. My advice is not to rely on the provider's general statement but to check the following four items yourself. Details change with the plan and over time, so repeat the check periodically.

• Contract and data processing terms: Are input use, retention period, deletion requests and subprocessors written down?
• Account setting: Is use for model training turned off, and is it locked by an administrator or can each employee change it?
• Access and logs: Can you see who accessed which tool and when, and can you close the account of a departing employee?
• Data region: Does the provider state where data is processed, and does it offer a choice?

An enterprise plan does not make you KVKK compliant by itself. Safeguards improve on the tool's side, but the legal basis, the notice to individuals and the decision about what data goes in remain yours.

Does anonymizing data really protect you when you use AI?

Anonymization protects you when it is done properly, but it is more than deleting names. In true anonymization the data can no longer be tied to a person even when combined with other data. Writing "Customer A" is often only pseudonymization, because a table the company holds can bring the person back. In training I teach a four step routine that employees pick up within minutes. It reduces workload but does not eliminate it. For high risk data such as health records, not sharing at all is safer than trusting anonymization.

1. Remove identifiers: Delete names, phone numbers, emails, addresses, ID and account numbers.
2. Replace with a role or code: Use terms like "customer," "manager" or "Patient 1" and keep the mapping table away from the tool.
3. Generalize details: Use a month instead of an exact date, a region instead of an address, a general role instead of a rare title.
4. Read it as an outsider: Would a stranger reading the text understand who it is about?

How do you set up a traffic light data classification for employees?

A traffic light classification splits company data into three colors: green is free, yellow is conditional, red is off limits. Employees do not need to memorize rules. They only ask, "What color is this data?" Keep the system simple, because a table with ten categories helps no one and is forgotten after the first week. Use the table below as a starting point and swap the examples for your own sector. Put it on one page together with the list of approved tools, on the wall or on the intranet.

The yellow zone generates the most questions. For an employee in doubt the rule is clear: treat it as one color higher and ask.

What are the most common data mistakes in companies?

• Work data in a personal account: An employee opens a free account and pastes a customer email. The company has no view of the data or the settings.
• Treating file uploads as harmless: Uploading a spreadsheet or PDF has the same effect as pasting text. Any personal data inside goes along with it.
• Meeting recorder plugins: Tools that record and transcribe calls without participants knowing create both personal data and confidentiality risk.
• Sharing output unchecked: A tool can carry personal details from the input into the output. Read it before sending it outside.
• Managing by ban: A fully banned tool leaves through the door and reappears on personal phones and home computers. An approved path is safer. For profession specific examples, read my articles on AI in law firms and AI for HR teams, since these teams handle the most sensitive data.

Where should a company start?

This is the order I follow, and no step needs a legal document, only a decision:

1. Take inventory. Which team uses which tool with which data? Ask about personal accounts as well as official licenses.
2. Sort data into three colors. Fill in the table above with your own examples.
3. Pick an approved tool. Check the enterprise contract and settings together with legal and information security.
4. Put the rule in writing. Keep it under two pages, with the table on page one.
5. Train with real data. When employees practice removing sensitive details from their own files, the rule sticks. This is the data and responsibility part of AI literacy, which I cover in my article on what AI literacy is.

If you would like me to teach these rules hands on, using your team's own workflows, you can request a proposal tailored to your team on the corporate AI training page, or we can map your current situation together through my consulting work. Wherever a legal interpretation is needed, I recommend working with your company's legal counsel.