Shadow AI is the AI variant of the well-known shadow IT. It rarely arises from bad intent: employees want to work faster, free chat assistants are just a click away, and many programs are constantly receiving new AI features. Microsoft describes shadow AI as the use of AI tools, especially generative AI, without the knowledge, approval, or control of IT and security teams, and names data leakage, regulatory violations, and reputational damage as consequences.
How shadow AI arises
- Customer inquiries, contracts, or HR data are copied into a freely available chat assistant to get an answer or summary faster.
- A transcription app runs during meetings without anyone having reviewed the data processing.
- A browser extension with an AI feature reads every visited page, including internal applications.
- Developers use an AI coding assistant with a personal account for company code.
- Software that is already in use activates new AI features that transfer data to additional service providers.
- A business department builds an automation that sends data to an AI model using a private API key.
Risks
- Data leakage: Whether inputs are stored or used for training depends on the provider, plan, and settings. With consumer offerings, users themselves often decide on this setting; with enterprise contracts, it is usually regulated differently.
- Data protection: Without a data processing agreement (DPA) and without checking the processing location, disclosing personal data is regularly problematic under the GDPR.
- Trade secrets: Quotes, calculations, or source code leave the company uncontrolled.
- Wrong results: AI answers are adopted unchecked, see hallucination.
- Obligations under the AI Act: Under Art. 4, companies that deploy AI systems must take measures to promote the AI literacy of their staff. The rule has applied since February 2, 2025, and was reworded by the 2026 Digital Omnibus Regulation. For high-risk applications such as personnel selection, additional obligations apply from December 2, 2027. Unknown tools can neither be covered by training nor classified.
- Security: Extensions and apps can contain malicious functions or be vulnerable to prompt injection.
Detecting shadow AI
- Ask: A short survey without sanctions shows which tools are used for what. This is often more productive than any technology.
- Cloud discovery: Microsoft Defender for Cloud Apps maintains a generative AI category in its app catalog. Discovered apps can be displayed with a risk assessment and tagged as sanctioned or unsanctioned.
- Network analysis: Microsoft Entra Global Secure Access detects access to AI services, AI APIs, and MCP servers in network traffic. Logs from a proxy or DNS filter also provide clues.
- Permissions and extensions: Regularly review consents to third-party apps in Microsoft Entra ID and installed browser extensions.
What helps against shadow AI
- Offer a good alternative: An approved tool with an enterprise contract, such as Microsoft Copilot in your own Microsoft 365 environment, removes the reason for workarounds.
- Clear rules: A short AI usage policy defines which tools are permitted and which types of data may be entered into them.
- Train: Employees understand why rules apply and recognize risky situations.
- Secure technically: Block unsanctioned AI apps or add a warning, and protect sensitive content with data loss prevention policies.
- Approve quickly: A simple review path for new tools prevents employees from waiting months for a decision and acting on their own.
- Keep a register: Record all AI systems in use with their purpose, data types, and responsible persons, as part of AI governance.
How it differs
Shadow IT covers any unapproved software and any unapproved cloud service. Shadow AI is the part of it in which AI models process data. There, the risk additionally lies in the further use of inputs and in results that cannot be traced. AI governance is the framework that turns shadow AI into regulated use.
How TOMORIS relates
The TOMORIS UNIVERSITY training AI in Everyday Work covers how teams use AI tools productively while taking confidential data, approvals, and internal rules into account.
Sources
- Microsoft Learn: Prevent data leak to shadow AI
- Microsoft Learn: Shadow AI discovery in Global Secure Access
- Microsoft Learn: Manage generative AI apps for your organization
- EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence, Art. 4 and Art. 113
- EUR-Lex: Regulation (EU) 2026/1744, Digital Omnibus Regulation on AI, OJ L of July 24, 2026
- BSI: Generative AI models, opportunities and risks for industry and public authorities
- Anthropic Privacy Center: Training data for consumer plans (example of a user setting)