
When employees paste customer data, contracts or source code into public AI chat tools with personal accounts (“shadow AI”), this can amount to uncontrolled processing of personal data and, in most cases, a cross-border transfer on behalf of your company, which is the data controller under KVKK, Türkiye’s Personal Data Protection Law. The answer is not a blanket ban but a control plan: an approved business tool, a written policy, DLP, SSO, logging and training. This article is not legal advice; have your specific situation reviewed by legal counsel.
Symptoms: what shadow AI looks like
- Firewall or DNS logs show regular traffic from personal accounts to AI chat sites with which the company has no contract.
- Employees paste an email containing a customer’s name, phone number and order details verbatim with a request like “Write a polite reply to this customer complaint”.
- Draft contracts, quotations and staff appraisals are uploaded with the instruction “summarise this”.
- Developers paste code containing a database connection string or API key to debug it.
- When asked “What are our rules for using AI?”, nobody in the company has a clear answer.
- The situation matches the observation in the Turkish Data Protection Authority’s announcement of 5 March 2026: these tools have spread through individual employee choices rather than a corporate strategy and have become hard to monitor.
Root cause: the legal and technical picture
KVKK (Law No. 6698 on the Protection of Personal Data) is Türkiye’s general data protection law. It was originally modelled on the former EU Data Protection Directive and shares many concepts with the GDPR, such as data controller, data processor and explicit consent. Under Article 12, the data controller must take all necessary technical and organisational measures to ensure an appropriate level of security in order to prevent unlawful processing of and unlawful access to personal data and to safeguard the data. If data is processed by another person on its behalf, the controller is jointly responsible with that person for these measures. Paragraph 5 of the same article requires the controller to notify the data subject and the Board as soon as possible if processed personal data is obtained by others through unlawful means. The GDPR sets a more specific deadline of 72 hours for notifying the supervisory authority. An employee pasting customer data into a personal AI account raises directly the question of whether the company is meeting these obligations.
The second dimension is cross-border transfer. Article 9 of the law was amended by Law No. 7499 of 2 March 2024, bringing it considerably closer to Chapter V of the GDPR. Transfers now rely first on an adequacy decision by the Board; without one, on appropriate safeguards such as binding corporate rules approved by the Board, standard contracts published by the Board, or a written undertaking authorised by the Board. One notable difference from the GDPR: a standard contract must be notified to the Authority within five business days of signature, whereas GDPR standard contractual clauses require no such filing. If none of these can be provided, transfers are permitted only on an incidental basis in the limited cases listed in the law, comparable to the derogations in Article 49 GDPR. With an employee’s personal account, there is no contract, no safeguard and no record of the transfer between the company and the service provider.
On 5 March 2026 the Personal Data Protection Authority published an announcement titled “Use of Generative AI Tools in Workplaces”. It highlights the risk of sharing customer information, trade secrets or personal data with third-party platforms and recommends using anonymised or generalised data where possible, a written usage policy, employee awareness, rule-based access control and human review of outputs.
The difference between consumer and business plans
- OpenAI: according to its enterprise privacy page, data from ChatGPT Business, Enterprise, Edu and the API Platform is not used for model training by default unless the customer explicitly opts in. The same page states that data from ChatGPT versions for individuals is also used for training.
- Anthropic: on consumer Claude plans (Free, Pro, Max), chats may be used for training if the user has allowed this in the privacy settings or if they are flagged for safety review. Anthropic explains in a separate help article that it does not use inputs from commercial products such as Claude for Work and the API for training by default.
- Microsoft: Microsoft Copilot and Copilot Chat (formerly Microsoft 365 Copilot and Copilot Chat) used with a work account are covered by enterprise data protection; Microsoft acts as a data processor under the Data Protection Addendum (DPA), and prompts and responses are not used to train foundation models.
This information is taken from the vendors’ official pages as of October 2026 and may change. A “not used for training” commitment does not equal KVKK or GDPR compliance: the legal basis for the transfer, transparency obligations, retention periods and access control remain your responsibility.
Step-by-step fix: a control plan
- Measure the current state: use firewall, proxy and DNS logs to find out which AI services are in use. The goal is an inventory, not punishment; remember that monitoring employees also requires transparency under KVKK.
- Classify data: create a simple scheme such as red (personal data, special categories of personal data, contracts, source code, passwords and keys), amber (internal correspondence) and green (public information), and state for each class what may be done with AI.
- Choose an approved tool: select a business plan that is managed through a company account, comes with a data processing agreement, documents that data is not used for training by default and lets you manage retention settings. Determine the legal basis for the cross-border transfer (for example, a standard contract and notification within five business days) together with your legal counsel.
- Publish a written policy: the list of approved tools, which data may be entered, the anonymisation rule, human review of outputs and whom to notify in case of a possible breach should fit clearly on one page.
- Provide access via SSO: enable the approved tool only with the corporate identity (for example Microsoft Entra ID), so that a leaver’s access is closed automatically together with their identity account.
- Apply DLP: use endpoint DLP and browser controls to audit pasting or uploading sensitive information types such as Turkish ID numbers, IBANs and card numbers to unapproved AI sites. Microsoft Purview endpoint DLP, for example, can monitor and block such activities on managed devices. Run it in audit-only mode first and weed out false positives.
- Redirect unapproved tools: restrict unapproved AI sites with a web filter, but show a link to the approved tool on the block page. A ban without an alternative moves shadow use onto personal phones.
- Keep logs: enable audit logs and retention policies in the approved tool so that you can later show who used which tool.
- Train staff: run a short training session with real examples and build the anonymisation habit, such as “write Customer A instead of the customer’s name”.
- Define incident response: if sensitive data was entered into an unapproved tool, record the incident, document the state of the chat and account settings, and assess with your legal counsel whether notification is required under Article 12(5) of the law.
This article is for general information only and is not legal advice. Work with a lawyer specialising in KVKK for decisions on the transfer basis, privacy notices and breach notification.
Lasting prevention
- Add AI tools as separate processes to your personal data processing inventory and update your privacy notices accordingly.
- Reread the providers’ privacy and data usage terms at least once a year and with every major version change.
- Review DLP and web filter reports monthly; recurring violations point to gaps in training or tooling.
- Add the policy to onboarding for new joiners and renew it once a year.
- When a new AI tool is requested, run a short approval process in which security, legal and IT assess it together.
When should you call in an expert?
If your team already uses AI heavily, if you process sensitive data such as customer or health data, or if there is a suspected data leak, technical controls need to be handled together with a legal assessment. What we usually see in the field at businesses in Denizli is a situation stuck between a ban and a free-for-all. As part of its AI-based business solutions, Workas provides technical support with selecting approved tools, configuring SSO and DLP, logging infrastructure and employee training; for the legal assessment we recommend working with your legal counsel.
Frequently Asked Questions
Is banning ChatGPT completely the solution?
Usually not. A ban without an alternative pushes use onto personal devices and invisible channels. An approved business tool, a clear policy and DLP together give better control.
If a business plan does not use data for training, is the KVKK issue solved?
No. The training commitment is only one aspect; controller obligations such as the legal basis for cross-border transfer, transparency, retention and access control still apply. Review this with your legal counsel.
Does anonymising the data remove the problem entirely?
Truly anonymous data, which cannot be linked to a person, falls outside KVKK, as it does under the GDPR. But deleting the name alone is often not enough; details such as an order number or address can still make the person identifiable.
What should we do if an employee has already pasted customer data?
Record the incident with date, tool and data type, document the state of the chat and account settings, and assess with your legal counsel without delay whether notification to the Board and the data subject is required under Article 12 of the law.
Sources
- 6698 sayılı Kişisel Verilerin Korunması Kanunu (mevzuat.gov.tr)
- KVKK – İş Yerlerinde Üretken Yapay Zekâ Araçlarının Kullanımı (5 Mart 2026)
- OpenAI – Enterprise privacy
- Anthropic Privacy Center – Is my data used for model training? (consumer)
- Microsoft Learn – Enterprise data protection in Microsoft Copilot and Copilot Chat


