Someone in sales pastes an entire e-mail thread into a chatbot to draft a reply to a difficult customer. It takes thirty seconds, the reply is good, nobody notices. What actually happened in those thirty seconds?
What happened, technically
The customer's name, e-mail address, order history and the details of their complaint left the corporate boundary, most likely to a server abroad. Because the employee used a personal account, there is no contract behind it: the organisation does not know how long the data is kept, whether it is used for training, or who can access it, and it has no one to ask.
More importantly, the event left no trace in the organisation's systems. On the network it looks like an ordinary website visit; there is no log line that says "data sent to an AI service". This is what people mean by shadow AI: usage the organisation cannot see and therefore cannot govern.
What happened, under KVKK
Under Turkey's Personal Data Protection Law, responsibility does not sit with the employee who pasted the data; it sits with the organisation as the data controller. Sending customer data to a service abroad is a cross-border transfer, and a transfer needs a legal basis. An employee's personal account provides none, and the service is not a processor the organisation has approved.
Whether the event counts as a data breach depends on the nature of the data and where it went, and it has to be assessed. But "we had banned it" is not a defence on its own. What an audit looks at is the technical and administrative measures the controller took. A ban is an administrative measure; without a technical measure showing it works, it is incomplete.
Why the ban does not work
Because the tool is genuinely useful. Summarising a report, drafting an e-mail, turning a contract clause into plain language: these save hours of someone's day. A ban does not remove the need; it moves it to a personal phone.
The real cost of a ban is the loss of visibility. When the organisation provides the tool, it can see which team uses what and for what, classify the data, and stop risky use. When it bans the tool, all of that information drops to zero; usage continues, it just becomes invisible.
Usage you cannot see cannot be governed. Usage you cannot govern cannot be defended in an audit.
What a working alternative looks like
The test is simple: data must not leave the organisation's boundary. Today the clearest way to meet it is to run the model on the organisation's own server, or in an environment it controls. Open-weight models are mature enough for the job; deployment is no longer a research project but infrastructure work.
- The model runs inside. Questions and pasted documents never leave the server; the system keeps working if the internet connection does not.
- It speaks with the organisation's documents. Procedures, contracts and product documentation are placed in front of the model; answers come with their sources.
- It plugs into the existing identity infrastructure. Who asked what and which document they reached is logged and subject to the organisation's access policies.
- Policy becomes enforceable. Instead of "do not use", you can say "use this, and do not enter that", because now there is a tool to point to.
Where to start
- Inventory. Which teams use which tools, for what? Find out by asking, not by banning. People will tell you if they know they will not be punished.
- A sanctioned path. A pilot on a single workstation for a team of five to ten. No large hardware investment is needed; the goal is to show that an alternative exists which is no worse than the tool in the employee's pocket.
- A policy update. Replace the ban clause with a short one naming the approved tool and the data classes that must not be entered. It should not exceed a page.
When an employee pastes customer data into an external AI tool, the organisation becomes the controller of a cross-border transfer it cannot see. A ban does not fix that; it hides it. What fixes it is an alternative where data stays inside the boundary and usage is visible.
If you do not know today how many people in your organisation use which tools, that is exactly the starting point. We map it together in a discovery call.