Your team is pasting company data into ChatGPT: how to manage the Shadow AI risk

Shadow AI is when employees use AI tools that the company never approved (usually a personal ChatGPT account or a similar assistant) for work, and feed company data into them. The problem is not the tool itself but the lack of oversight: the moment someone pastes sensitive information to summarize a contract or fix a broken piece of code, that data is no longer under your control. The right response is not to ban AI but to govern it. That means seeing what is being used, setting clear rules, offering a safe tool, and training your people.

How big is the problem?

The scale is concrete. According to IBM's 2025 Cost of a Data Breach report, the global average cost of a breach is USD 4.44 million. Breaches that involved shadow AI came in higher, at USD 4.63 million on average, which puts the premium of unsanctioned AI use at roughly USD 670,000 per incident. The same report found that one in five studied organizations suffered a breach linked to shadow AI. Of those organizations, 97 percent had no proper AI access controls, and 63 percent had no AI usage policy at all.

What actually leaks?

Usually it is the ordinary stuff of daily work: a customer list, a proposal draft, an unsigned contract, a block of code that keeps throwing errors, an end-of-quarter financial statement. The best-known example happened at Samsung. In 2023, within about three weeks of the company allowing ChatGPT, engineers pasted confidential data into the tool in three separate incidents: source code tied to semiconductor manufacturing, code used to detect equipment defects, and a transcript of internal meeting notes. Samsung quickly banned generative AI tools and made a violation grounds for discipline up to termination.

Where does the data go?

The critical distinction is where the data ends up. On the free and individual versions of ChatGPT (Plus, Pro), your conversations can be used to train the model by default unless you turn that off in settings. To stop it, you go to Data Controls under Settings and disable the relevant option. On the Enterprise, Business, and Team versions, your data is not used for training by default and comes with a data processing agreement. Typing the same prompt into a personal account versus a company plan are two very different things for data security.

KVKK applies to you directly

Pasting a customer's ID details, an employee's HR record, or a health record into an unvetted third-party tool is, in most cases, unlawful processing. If the tool is hosted abroad, it also triggers the rules on cross-border data transfers. You are the data controller. Knowing where the data goes and keeping it secure is your obligation, and shadow AI breaks that chain invisibly. In an audit, saying "the employee did it from their own account" will not clear you of responsibility.

Open a safe path instead of banning

Banning is the first idea that comes to mind, but it rarely works. Cutting off the tool does not end the usage, it just hides it: the employee does the same task from a phone or a personal laptop, and this time leaves no trace at all. People use AI because it speeds up their work. Give them an approved tool on a company plan, one that fits the rules, and the pull toward the risky path largely disappears.

A few steps go a long way in practice:

• See current usage. Which tools, used by whom, through which accounts? You cannot manage what you cannot measure.
• Write a short usage policy. Spell out what must never be pasted: source code, passwords and access keys, customer personal data, undisclosed financials.
• Offer an approved tool on a company plan. Connect it through single sign-on (SSO) so accounts stay under control.
• Add technical controls. Data loss prevention (DLP) and domain-level blocking limit risky use from the start.
• Train your people. With real examples, in short and regular sessions. The goal is not punishment but building reflexes.

A one-question rule for staff

The most practical test you can give your team is this: before typing something into a tool, they should ask, "would I be uncomfortable if this ended up on a public forum?" If the answer is yes, that text does not belong in a consumer tool. For testing and experimentation, using sample or masked data instead of the real thing usually does the same job and keeps the risk near zero.

If the data is already out

If you discover that sensitive data has already been shared, work through it in order. Identify what was entered and into which tool. Remember that turning off training does not work retroactively, so data shared earlier may not be recoverable. If what leaked is a password or an access key, rotate it immediately. If personal data is involved, assess your KVKK notification duty. A simple, prepared response checklist makes it clear who does what in the moment.

Shadow AI is already happening in most companies, and rarely discussed. If you want to make usage visible, choose the right tool, and set up a KVKK-compliant usage policy, Wedevit can run an independent assessment to map where and how AI is used in your company and build a safe framework with you.