In many companies, shadow AI has long been part of everyday life – only nobody sees it. A quote is quickly reworded in a personal AI chat, the meeting minutes are summarised by a free app, the customer list is uploaded for analysis. Nobody means any harm; everyone just wants to finish faster.
That is precisely the problem: the work gets done, but the company loses track. Ultimately, no one knows which data has ended up where. The answer is not a blanket ban, but a clear line that makes the right way easier than the wrong one.
We show what lies behind the phenomenon and what risks it entails. We also present five building blocks of an AI policy that turn uncontrolled sprawl into an orderly use of AI.
Key facts at a glance
- Shadow AI refers to AI tools that employees use for their work without the company’s knowledge or approval.
- According to an AXA study, only around a third of Swiss SMEs that use AI have set rules on which company data may be entered into AI tools.
- The risks mainly concern confidential information, personal data and a lack of traceability.
- A short AI policy covering data classes, permitted tools, an approval process, labelling and training creates clarity.
- Bans merely displace shadow AI; what works is an approved tool that is better in everyday use than the personal one.
What is shadow AI – and why does it emerge?
The term is new, the pattern is not. Shadow IT refers to systems that employees or business units procure and operate without involving the IT department. Shadow AI is its latest variant.
From free tool to work tool
It used to be the Excel file full of macros that only one person understood. Today it is the AI chat in the browser: freely accessible, instantly available, no installation needed. This low threshold is exactly what makes personal AI tools so widespread.
According to the Federal Statistical Office, more than two in five people in Switzerland used AI in 2025 to create content such as texts or images. Among 15- to 24-year-olds, it was as many as four in five.
Use is growing in companies, too. According to AXA’s 2025 SME labour market study, 34 per cent of the SMEs surveyed deliberately integrate artificial intelligence into their work processes, up from 22 per cent a year earlier.
Why employees turn to their own tools
This is rarely rebellion. Usually, employees want to work more productively than the official tools allow. The Work Trend Index 2024 by Microsoft and LinkedIn surveyed 31,000 people in 31 countries worldwide. The result: 78 per cent of AI users bring their own AI tools to work.
Often there is no official tool, or it is cumbersome, or nobody has explained what is allowed. Where rules are missing, each person decides for themselves. And they do so anew every day.
The risks of shadow AI: data, confidentiality, traceability
A single prompt seems harmless. But it is the sum that counts: many small inputs, spread across dozens of personal accounts.
Confidential data leaves the company
Anyone who copies a draft contract, a customer list or salary details into a personal AI account passes data on to third parties. With international providers, the place of processing, the applicable law and the use of the content entered depend on the respective contractual terms. In the case of personal accounts, nobody in the company has ever checked these terms.
Where personal data is involved, the Swiss Federal Act on Data Protection (FADP) also comes into play. The Federal Data Protection and Information Commissioner (FDPIC) points out that the Act is technology-neutral and therefore directly applicable to AI-supported data processing. Read more in our article on AI and data protection.
Nobody knows what is stored where
Personal AI accounts leave hardly any trace in the company, but they do at the provider. When someone leaves the firm, chat histories containing company content remain in their personal account. Deleting data, providing information, investigating incidents: all this becomes difficult when the tools in use are unknown.
Results without review
Language models write convincingly, even when they are wrong. If unchecked AI texts end up in quotes, contracts or replies to customers, the company bears the consequences. It also remains unclear which content comes from an AI.
Five building blocks of an AI policy against shadow AI
Many businesses still have catching up to do here. In the same AXA study, only around a third of SMEs that use AI have set rules on which company data employees may enter into AI tools. Fifty-eight per cent have no such rules; among small SMEs with 5 to 9 employees, the figure is 68 per cent.
A good AI policy is not a legal code. It fits on two pages and answers the questions that arise in daily work.
1. Data classes: what can go into the AI?
Divide your information into a few clear classes, such as public, internal, confidential and personal data. For each class, specify which tools may process it. A press release is simply not a payslip.
2. Permitted tools: a short, clear list
Name the approved tools and what they are intended for. A positive list shows the right way and achieves more than any list of prohibitions. It is important that it grows with demand; otherwise an unofficial one will soon emerge alongside it.
3. Approval process: review new tools instead of banning them
Employees are constantly discovering new applications. Give them a simple way to propose a tool for review, with a clear contact person and a reply within a reasonable time. This turns uncontrolled sprawl into a channel for good ideas.
4. Labelling: disclose where AI is involved
Specify when AI assistance is disclosed, especially in texts for customers and in decisions about people. This creates transparency; a note on human review strengthens trust.
5. Training: understanding the rules, not just signing them
A policy only works once it is understood. Short, practical training shows how to check results and recognise sensitive data. People who know why a rule exists are more likely to follow it.
From ban to better alternative
The instinctive response to shadow AI is often a ban. That is understandable but rarely solves the problem: the need remains, so usage moves to the personal smartphone. The risk does not shrink as a result; it merely becomes less visible.
Urban planners know this as a desire path. Where people keep walking across the lawn, a path is obviously missing. Wise planners then pave the path instead of putting up a fence.
Applied to AI, the most effective measure is an approved tool that is at least as convenient in everyday use as the personal one and integrates company knowledge in a controlled way. Our comparison offers an overview of alternatives to well-known AI chats.
Shadow AI needs light, not prohibition signs
A proliferation of AI tools is not a sign of a lack of discipline, but of genuine need. Look closely and you will see where time is lost in everyday work. What matters is channelling this need in an orderly way: with clear rules and a good tool.
At BE BRAVE, we support companies on both levels. In our AI consulting, we assess AI readiness, data situation and risks and derive a prioritised roadmap from them.
With EagleCHAT, teams get a private Swiss AI chat that answers questions on the basis of approved documents. All AI processing takes place on our own servers in Switzerland. Your content is not used for training, fine-tuning or improving AI models. This is how you bring AI out of the shadows.
FAQ
What is meant by shadow AI?
It means AI tools that employees use for work without the company’s knowledge or approval. Typical examples are personal accounts with AI chats, translation or transcription services.
Is shadow AI prohibited?
Not in principle. It becomes problematic when confidential information or personal data reaches third parties on terms nobody has checked. Internal directives, confidentiality obligations and the FADP may then be affected.
How can you identify shadow AI in your own company?
The easiest way is to ask rather than monitor. An anonymous survey shows which tools are in use and what for.
What belongs in an AI policy?
Data classes, permitted tools, an approval process for new tools, rules on labelling and training opportunities. Short and easy to understand, it works better than a lengthy rulebook.
Does a ban help against shadow AI?
Usually not. A ban often just shifts usage to personal devices. An approved tool that meets employees’ needs better is more effective.
