Shadow AI refers to employees using AI tools without the company's knowledge or approval – usually via private accounts, free versions or their private phone. The term is based on the well-known “shadow IT”, meaning unapproved software within a company.

Is it forbidden for employees to use ChatGPT for work?

There is no general legal ban. It becomes problematic when personal data or trade secrets end up in tools for which there is no contractual basis – the company then risks data protection breaches and the loss of trade secret protection. Clear company rules are therefore what matters.

How do I recognise shadow AI in my company?

Typical signs are strikingly uniform texts, unusually fast translations or wording that does not fit the person. The most honest way, however, is not monitoring but an open conversation: if you ask anonymously or without sanctions who uses which tools, you usually get a realistic picture.

Why is banning AI tools not enough?

Because it does not end the use but shifts it – to private devices and accounts, where the company can neither see nor control what happens to the data. A ban without an alternative creates exactly the lack of transparency it is meant to prevent.

What is a safe alternative to public AI tools?

A tool with a verified contractual basis: a data processing agreement (DPA), data storage on EU servers, exclusion of your data from training third-party models, and access control. For knowledge questions within the company, an AI knowledge base that answers from your own documents and names the source is a good fit.

Shadow AI: when employees secretly use ChatGPT

Shadow AI is the use of AI tools in a company without the knowledge or approval of management – the colleague who secretly has ChatGPT draft customer emails, the employee who has contract passages translated via a private account. In many companies this has long been reality and is rarely ill intent: the tools are noticeably helpful, and where the company offers no official solution, the unofficial one emerges by itself. The effective answer is not a blanket ban but the combination of clear rules and a safe, approved alternative.

In brief

Shadow AI almost never arises from malice but from usefulness: employees reach for AI tools because they work faster with them – and do so secretly if there is no official route.
The main risks are data protection breaches, the leaking of trade secrets, errors adopted without checking and complete loss of control via private accounts.
The company remains responsible: if customer data ends up in an unvetted tool via a private account, that is the company’s problem, not just the employee’s.
Blanket bans shift use to the private phone and make it invisible – they make the problem worse instead of solving it.
The proven route: an inventory without blame, short clear rules, an approved safe alternative and training.

What is shadow AI and why does it arise?

The term is modelled on “shadow IT” – unapproved software that employees obtain themselves because the official tools are not enough. With AI the same thing happens, only faster: access costs nothing, signing up takes two minutes, and the benefit is felt immediately. A quotation text that would otherwise take an hour is ready much faster; an English complaint is quickly translated into something understandable.

People who use it do not see themselves as rule-breakers but as efficient. That is exactly the dynamic: as long as the company takes no position and offers no approved solution, the gap fills itself – invisibly, via private accounts, without any contractual basis. For most companies, the question is therefore not whether AI is being used, but whether its use is visible and regulated or invisible and unregulated.

What risks does shadow AI create?

The risks are spread across several levels that should be kept apart:

Risk	What can actually happen
Data protection	Personal data of customers or employees ends up in a tool without a data processing agreement (DPA) – a GDPR breach by the company, regardless of who entered the data.
Trade secrets	Cost calculations, formulations or contract details are entered into third-party systems. Legal protection of trade secrets requires appropriate confidentiality measures – careless handling can put it at risk.
Training third-party models	With free and private accounts, depending on the provider and settings, it cannot be ruled out that inputs are used to improve the models.
Incorrect results	AI answers sound convincing even when they are wrong. Without rules for checking, invented figures or half-truths find their way unchecked into quotations and customer emails.
Loss of control	The private account belongs to the employee. If they leave the company, the history and the company data entered remain in their access – the company has no recourse whatsoever.

Important for context: responsibility under data protection law lies with the company as the controller within the meaning of the GDPR. In the event of complaints, the Austrian Data Protection Authority (dsb.gv.at) is the point of contact for data subjects – and it contacts the company, not the individual employee. This article provides general orientation and does not replace legal advice in individual cases. What requirements a GDPR-compliant use of AI has to meet overall – server location, contracts, training exclusion – is summarised in the overview article AI and the GDPR in Austria.

Why don’t bans alone work?

A blanket ban looks consistent on paper but fails on three counts. First, it is hardly enforceable: no company filter can reach the private phone in someone’s pocket. Second, it punishes the wrong people: precisely the committed employees who look for ways to work faster are pushed into secrecy. Third, it destroys visibility: after the ban, management is even less likely to learn who does what with which tool – the risk remains, only the knowledge of it disappears.

That does not mean everything should be allowed. It means a ban should only apply to what cannot be controlled – unvetted tools, private accounts, sensitive data – and alongside it there must be a good official route. People take the most convenient route; the approved one has to be the most convenient.

How do you bring AI use out of the shadows? Five steps

Inventory without blame. Ask openly who uses which tools for what – expressly without sanctions. The aim is an honest picture, not a register of culprits. It usually turns out that the use cases are sensible; only the tools are not.
Set short, clear rules. Two to four pages are enough: permitted tools, forbidden inputs (personal data, trade secrets, login credentials), an approval route for new requests, and a duty to check AI results. Written to be understood, not in legalese.
Provide a safe alternative. The decisive step: an approved tool that covers the real use cases and is contractually sound. Only with an alternative does the list of prohibitions become something people can follow.
Train rather than announce. A short training session with examples from your own everyday work – this request yes, that one no, and why – has more effect than any circular email. Include new employees from the start.
Keep at it. The tool landscape changes constantly. A fixed date each year for reviewing the rules and the tool list, plus an easy-to-reach contact point for new requests, keeps the system up to date.

What makes a safe alternative?

Four criteria separate a tool that can be approved from a risky one: a contractual basis (a data processing agreement (DPA) with the provider), data storage on servers in the EU, an explicit exclusion of your content being used to train third-party models, and access control – meaning control over who in the company may see what.

For the most common business use case – questions about your own processes, products, instructions and service knowledge – a fifth criterion is added: verifiability. An AI knowledge base that bases its answers on your own documents and names the source document for every answer has a twofold advantage over the public chatbot: the answers are based on your actual company knowledge rather than the average of the internet, and every employee can open the source and check whether the answer is correct. AI can be wrong – a tool that discloses where its answers come from makes this manageable. The secret shortcut thus becomes an official working tool that also delivers better answers to business questions than any general-purpose tool.

The next step is a conversation, not a project: ask your team this week what they already use AI for today – without sanctions and with curiosity. Your list of requirements will almost write itself from the answers. How you then introduce a safe, approved knowledge base is shown step by step on the implementation process page.