August 19, 2026 · Triox Informatika · 5 min read
The question isn't whether your employees use AI
It's which tools, and what they paste into them. Free AI tools are in use even at companies that officially have no AI—and that's not a disciplinary issue, it's a leadership one.
In brief
- Even where a company officially has no AI, people use it—in the browser, on their own phones, while they work.
- This isn’t rule-breaking; it’s filling a gap: the employee wants to get their work done and found a tool for it.
- The risk isn’t AI. It’s that company data ends up somewhere with no contract, no log and no control.
- Banning it doesn’t work. What works is an approved route and three simple rules.
An everyday scene. An employee receives a twenty-page draft contract in English and has to summarize the key points by tomorrow. They open a free AI site, paste in the text, and in fifteen minutes the task is done that would otherwise have taken half a day.
They didn’t do anything wrong. They didn’t mean to harm anyone. But at that moment, the company’s contract went to a provider the company has no agreement with—and nobody at the company knows about it.
This isn’t a disciplinary issue
The phenomenon has a name: shadow AI. It’s the same pattern we saw earlier with shadow IT, when employees started using their own file-sharing service or messaging app because the official tool was cumbersome.
Let’s be clear right away: this isn’t negligence. People who work this way are typically among the most motivated employees—the ones who want to move faster. If the company treats it as a disciplinary matter, it achieves two things: the use doesn’t stop, it just gets better hidden, and the next similar idea never reaches management either.
If the official route is slower than the unofficial one, what you need isn’t a policy—it’s a better official route.
What happens to the data?
This is where the conversation usually goes off track, because people either underestimate the risk (“nobody reads it anyway”) or exaggerate it (“now everyone can see it”). The reality is more mundane, and it comes down to three questions.
Where is it stored? With a free account, the conversation stays with the provider, typically tied to the account. Which country, for how long and who can access it is set by the terms of use—not by a contract signed with you.
Do they learn from it? With several free services, the default is that what you type in may be used to further develop the system. In paid business plans this is typically not the case—but that is exactly the difference that disappears when someone works from a personal account.
Who knows about it? This is the most important one. If data leaves through a personal account, the company has no log, no way to trace it, and in a data protection incident it can’t answer the simple question of what data was affected.
Why banning it doesn’t work
The obvious reaction is a company-wide memo: “The use of AI tools is prohibited.” It fails for three reasons.
First, it can’t be enforced. Personal phones and personal accounts are outside the company network.
Second, it puts you at a competitive disadvantage. If the competition prepares a proposal in a third of the time, the company will feel it sooner or later.
Third, it takes away visibility. Someone who keeps using it despite the ban won’t speak up if something goes wrong—and that’s exactly what you need to find out.
What works instead
You don’t need a thicker policy; you need a usable route and a few sentences everyone remembers.
The first is the official tool. If there is a company subscription that employees access with their company identity, the reason for using personal accounts goes away. We covered this and the pitfalls of rolling it out in a separate post earlier: the biggest surprise is typically not what the AI can do, but the state of the permissions.
The second is the three categories. You don’t need a detailed list—this is enough:
| Fine to use freely | Company tools only | Never, anywhere |
|---|---|---|
| Public information, general professional questions | Internal material, proposals, client names, contracts | Passwords, access keys |
| Your own draft with no client data in it | Financial statements, business plans | Personal data in bulk (payroll, health data) |
| Templates, samples, example text | Internal documentation, network descriptions | Anything covered by a nondisclosure agreement |
The third is the option to report. Make it explicit that anyone who flags after the fact that they may have pasted in data they shouldn’t have won’t get in trouble. This is the only way the company will find out at all.
What isn’t approved can’t be monitored. And with what can’t be monitored, you only find out something went wrong when it’s already too late.
Five steps to get started
- Ask who uses what—with no repercussions. An honest survey reveals more than any technical assessment. The goal is the picture, not holding people to account.
- Provide an official tool. A company subscription is usually cheaper than investigating a single data protection incident—and it removes the reason for using personal accounts.
- Write down the three categories. One page is enough: what can go in, what can only go into company tools, and what can't go anywhere. If it's longer, nobody will read it.
- Make it explicit that the output has to be checked. AI sounds convincing even when it's wrong. Whoever sends it out is still responsible for the numbers, the deadlines and the legal content.
- Revisit it in six months. In this area, as much changes in six months as in three years elsewhere. Whatever is a good rule today, in six months it's worth deciding whether it still is.
Technical details
- A company subscription with sign-in through the company’s own identity (SSO), so that access also ends when someone leaves
- Clarifying the data handling terms with the provider: where it is stored, for how long, whether it is used for development—with business plans this is typically a contractual question, not a setting
- Data loss prevention (DLP) for email and cloud storage, so that bulk movement of sensitive data is at least visible
- Logging and reporting: who used it, when, and to what extent—this is the foundation of incident response
- Permissions review before rollout: AI reaches what the user can reach, so messy shares surface here as well
- Fitting it into internal policy: the privacy notice and the employee notice may also be affected
- Regular review: the tools and their terms change fast
The leadership decision
This topic doesn’t start as an IT question, and it isn’t decided there either. IT can procure the tool, set up sign-in and keep logs—but what can go into it and what can’t is for the company’s leadership to say.
Until that decision is made, people use it anyway. Just in a way that nobody knows about.
Let’s start by looking at where things stand at your company today. At the end of the assessment you get a documented picture of the current state and a list of risks—including where your permissions management stands, which is the foundation of any AI rollout. Request an assessment, or read about our approach to security.