Shadow AI
The use of AI tools, models, or services inside an organization without approval, inventory, or governance — the AI-era descendant of shadow IT. It spans a spectrum: at the low-risk end, an employee pasting data into a personal ChatGPT account; at the high-risk end, a shadow agent wired into real systems with standing credentials. Shadow AI proliferates because sanctioned tooling is slower than building your own, and because the value is real — people route around procurement precisely when the unofficial option works better. The danger isn't the tool; it's the absence of visibility: you can't govern, log, secure, or attest to AI you don't know is running. The fix is not a ban (which drives it underground) but a fast sanction path plus discovery — see Agent Registry and Sanctioned Agent.
Example
During an amnesty survey, a fintech discovers 34 unsanctioned AI workflows: a support lead forwarding tickets to a third-party summarizer, an analyst running month-end numbers through an AI spreadsheet plugin, and a growth engineer's n8n workflow auto-replying to leads from the shared inbox. None had gone through security review; two were processing customer PII through personal accounts.
Frequently asked questions
- What's the difference between shadow AI and shadow IT?
- Shadow IT is unsanctioned software and services generally; shadow AI is the AI-specific slice of it. The reason shadow AI gets its own name is that AI — especially agents — can take autonomous action on data, not just store it, so the risk profile is sharper. Shadow SaaS was a data-residency problem; shadow agents are a data-agency problem.
- Should we ban shadow AI?
- Banning rarely works — it removes your visibility, not the incentive that created the shadow tool. The value that drove someone to build it is still there, so they hide it better. Mature programs 'sanction, don't ban': publish a paved road faster than going around it, so the compliant path is also the easy one. See Sanctioned Agent.