A familiar scenario: a salesperson under deadline pressure pastes a customer list into a public AI chatbot and asks for a quick analysis. The answer is useful and arrives in seconds. Without realising it, the employee has just sent confidential data to an outside service, beyond any control of the company. This is shadow AI, and the AI Act turns it from a security issue into a compliance one as well.
What shadow AI is
Shadow AI is the use of artificial intelligence tools within an organisation without the knowledge, approval or oversight of IT, legal or compliance: assistants, text and image generators, data analysis tools used by staff on their own, outside the official policies.
The term comes from shadow IT — unauthorised software in general, such as a cloud storage service installed without approval. The difference is that AI tools do not only store the data entered; they process it, and on some plans the provider may use it to train its models. Control is lost not only over where the data goes but over how it is used.
Why it spreads
- Accessibility. The most popular tools are free or cheap and run in a browser, with no installation and no technical knowledge needed.
- Productivity pressure. A tool that drafts an email or summarises a document in seconds is immediately attractive.
- No approved alternative. When the company offers no approved, secure tools, staff use the public ones.
- Unawareness of the risks. Many users do not realise that what they type into a public app can leave the company and be retained by the provider.
- No clear rules. Without a policy and training, nobody knows what is allowed and what is not.
The risks
Data security and confidentiality
- Confidential company or customer data sent to external servers with no assurance about its protection.
- Information retained by the tool’s provider and, under some terms, used to train models.
- Leaks and breaches that are hard to detect precisely because the activity is invisible.
Legal compliance
- Personal data entered into unapproved tools is processing without a documented basis, information to the people concerned or a processing agreement — a GDPR breach.
- Unapproved AI used in a process that falls under Annex III of the AI Act — recruitment, credit, access to services — makes the company a deployer of a high-risk system without knowing it.
- Without documentation, compliance cannot be demonstrated to an authority.
Operational and quality risks
- Output with errors, invented facts or bias, taken at face value, leads to wrong decisions.
- Reliance on uncontrolled tools makes processes inconsistent.
- Without human review, errors are amplified rather than caught.
What the AI Act requires
The AI Act (Regulation (EU) 2024/1689) applies to providers and to deployers — companies that use AI systems made by others. A company that develops no AI but whose staff use it still has obligations. The relevant ones here:
- AI literacy (Article 4)
Since 2 February 2025, and as reworded from 27 July 2026, the company must take measures supporting the AI literacy of the people who use AI on its behalf. It cannot do that for tools it does not know about.
- Prohibited practices (Article 5)
Apply to deployers too. A tool adopted informally — emotion recognition applied to call centre agents, for example — can fall straight into a prohibition.
- Transparency (Article 50)
Since 2 August 2026: disclosing deepfakes and certain AI-generated texts the company publishes. Content produced with unapproved tools escapes the check.
- High-risk systems (Article 26)
From 2 December 2027 for Annex III systems: use according to instructions, human oversight, monitoring, logs, informing workers. None of this is possible for a system nobody approved.
The dates are those of Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026. Fines reach 35 million EUR or 7% of worldwide turnover for prohibited practices and 15 million EUR or 3% for most other obligations. The postponement of the high-risk deadline is not a pause: an inventory is the first step either way.
Finding shadow AI — lawfully
- An inventory of the AI tools actually in use, through internal questionnaires and open conversations with teams.
- A climate in which staff can say which tools they use without fear of sanctions; concealment is the main obstacle.
- Technical detection of traffic to known AI services, at aggregate level — which services, how often — rather than the content of what individuals type.
In Romania, the last point is monitoring of employees’ electronic communications. When it rests on legitimate interest, Article 5 of Law no. 190/2018 requires a duly justified interest, prior, complete and explicit information to employees, consultation of the union or employee representatives, less intrusive means tried first, and storage in principle no longer than 30 days. Large-scale automated monitoring of employees’ behaviour is also on the ANSPDCP list of operations requiring a DPIA.
Governing it
Policy, training, roles
- An internal AI use policy: what is allowed, what is prohibited, and under what conditions.
- Training, so staff understand both the benefits and the risks — and records of it, as evidence for Article 4.
- Defined roles: who approves tools, who monitors use, who updates the policy.
Technical and process measures
- Approved, secure AI tools offered as the alternative to public ones — business plans with a processing agreement and no training on company data.
- Access control and data loss prevention measures.
- Human review built into the processes where AI supports decisions.
The structure of an AI use policy is set out in ChatGPT at work — GDPR and AI Act obligations; for training, see training.
Conclusion
Shadow AI is invisible but real, and it spreads as AI tools become easier to use. Under the AI Act it is no longer only a security problem but a compliance one: a company cannot meet obligations for systems it does not know it uses. Visibility comes first; policy, training and approved tools follow.
Frequently asked questions
What is shadow AI?
The use of artificial intelligence tools within an organisation without the knowledge, approval or oversight of IT, legal or compliance — employees using AI apps on their own, outside the company’s official policies.
How is shadow AI different from shadow IT?
Shadow IT is the unauthorised use of software or services in general. Shadow AI is a more specific risk: AI tools do not only store the data entered, they process it, and on some plans the provider may use it to train its models.
How does shadow AI relate to the AI Act?
The AI Act assumes the company knows which AI systems it uses. Shadow AI means it does not. If an unapproved tool is used in a process that would make it high-risk — screening job candidates, for example — the company cannot document the system, organise human oversight or answer the authorities, and does not even know it should.
Can we monitor network traffic to detect AI tools?
In Romania, monitoring employees’ electronic communications on the basis of legitimate interest must meet Article 5 of Law no. 190/2018: a duly justified interest, prior and complete information to employees, consultation of the union or employee representatives, less intrusive means tried first, and storage in principle no longer than 30 days. Aggregated detection of which services are used, without reading content, is easier to justify than inspecting what individuals type.
How can shadow AI be managed?
An inventory of the AI tools actually in use, a clear AI use policy, training, defined roles for AI governance, approved tools offered as an alternative, and human review in processes where AI supports decisions.