An employee pastes confidential customer data into ChatGPT to quickly get a summary. Another uploads snippets of proprietary source code into a personal AI tool because the internal solution feels too slow. Neither tool was approved by IT, neither has been reviewed for data protection, neither appears in any inventory. Welcome to shadow AI — by current assessments, one of the fastest-growing corporate risks of 2026.

More Than Just the Next Shadow IT

The term deliberately echoes "shadow IT" — unapproved software like personal cloud storage — but goes considerably further. Every input into an unapproved AI tool is potentially a data processing activity under GDPR, and with US-based providers, a third-country data transfer on top. Many popular AI services also permit, under their own terms of use, submitted content to be used for model improvement unless users actively opt out — something almost nobody does in practice.

What an Incident Actually Costs

Current analyses put the average cost of a shadow-AI-related data breach at roughly $4.63 million — noticeably higher than a comparable conventional breach. Well over 80% of companies examined already showed concrete signs of shadow AI activity. The risk is twofold: data entered this way sits outside regular backup and control structures, potentially giving attackers extra leverage — up to a double-extortion scenario where, alongside system encryption, the exposure of previously unknown data leaks becomes an additional threat.

Shadow AI is the blind spot that can undermine your entire risk management — it lives in the gaps between official controls: in the personal account, the quietly enabled feature, the integration nobody approved.

Three Regulatory Frameworks, One Blind Spot

With the EU AI Act, NIS2, and GDPR all in play, three regulatory frameworks now meet a reality where AI adoption grows faster than any governance structure can keep up with. The EU AI Act in particular requires a risk-based inventory of all AI systems in use — a requirement that's practically impossible to meet without knowing which tools are actually being used.

What Actually Helps

Security experts broadly agree that outright bans don't work — employees simply shift to less visible workarounds. A pragmatic, staged approach is more effective:

  • Inventory: which AI tools are already in use, by whom, with what data?
  • Categorize risk: which uses are harmless, which are critical?
  • Clear AI usage policy: what's allowed, and which data must never be entered?
  • Provide an official, genuinely usable alternative — the most effective lever against shadow usage is a good approved option
  • Defined approval process for new tools

Conclusion

Your employees are already using AI — that question has been settled. What's still open is whether your company is steering that usage or whether it's happening unchecked, out of sight. A governance framework doesn't need to be perfect to be effective; above all, it needs to exist.