Contents
Shadow AI is the use of artificial intelligence tools inside the organization without approval and without the knowledge of the IT function. It is a direct evolution of Shadow IT, with one feature that makes it worse: the information does not just go to a third party — it may be used to train a model.
IBM's 2026 report found that more than 20% of organizations reported a breach targeting AI models or applications. The common causes were not in the model but in the environment: compromised APIs and cloud misconfigurations, 27% each.
Where It Actually Happens
- Developers paste code and API keys into a chatbot to find a bug.
- Support and sales staff summarize inquiries and contact lists.
- Analysts run financial forecasts and contracts through free assistants.
- Managers transcribe meetings with a free tool installed without approval.
Note that none of these actions is malicious. All of them are honest attempts to work faster. That is exactly why a disciplinary approach does not work.
Why It Is Hard to Detect
The hardest part to detect is use through a personal free-tier account, from the browser. There is no organizational record, no contract, and no way to request deletion. From the standpoint of Amendment 13 to Israel's Privacy Protection Law, this is a transfer of personal data to a third party without a legal basis.
Action Plan
Measure before you ban
An anonymous employee survey and, in parallel, a review of network traffic to the domains of known AI vendors. Together, the two sources give a reliable picture.
Classify by risk, not by tool
The question is not "is ChatGPT allowed" but "what kind of information goes into it". The same tool can be allowed for drafting a marketing email and forbidden for analyzing a contract.
Provide an approved alternative
An enterprise version of one or two tools, with an agreement stating that data is not used for training. This brings the incentive for personal use down to almost zero.
Close what actually gets breached
According to IBM, the way in is through APIs and cloud permissions — not through the model. Revoke exposed keys, review permissions, and close open cloud configurations. This is ordinary IT work, not AI expertise.
Train
Most employees do not know there is a difference between a personal account and an enterprise version. Half an hour of explanation prevents more than any block.
If you run AI agents — not just chatbots — map their permissions. An agent acts under its own identity, and has usually been granted broad permissions "so it can work". That is a superuser nobody is monitoring.
What This Guide Does Not Cover
- Fully detecting use through personal devices outside the corporate network is not technically possible. This guide focuses on reducing the incentive, not on total blocking.
- The figures from the IBM report refer to a global average across all sectors. The implications for a specific organization vary.
Sources
Want to Go Through It Together?
A guide is a starting point. An assessment of your environment will give you a precise answer.
