What is shadow AI?
A contract goes into a free summariser so somebody can finish by six. The clause about confidentiality is now on a server your legal team has never heard of.
Shadow AI
Shadow IT for AIUnsanctioned AI
Shadow AI is the use of AI tools and assistants by employees without approval or oversight from their company's IT, security or legal teams. The pattern repeats shadow IT, with one difference: the data pasted into the tool may be used, stored or reviewed under terms nobody at the company has read.
Almost nobody does this to cause harm. Staff reach for a tool because it removes an hour of work and the approved route either does not exist or takes six weeks.
Three things typically leak. Customer records pasted in for summarising. Source code pasted in for debugging. And internal documents, which carry the strategy and the numbers a company would never email out.
The exposure is contractual before it is technical
A consumer AI account is governed by consumer terms. Your agreements with customers usually promise something stricter: named subprocessors, a defined retention period, and a location for the data. A tool nobody approved satisfies none of those, and the breach exists whether or not anything bad happens.
In regulated work it lands harder. Health data pasted into an unapproved assistant is a disclosure to a party with no agreement in place. Legal teams face the same shape of problem, where sharing material outside a protected relationship can cost the protection.
Then there is the quieter cost. Decisions start resting on output nobody logged, from a tool nobody can name. Ask later how a number was reached and the trail stops at a browser tab.
- 01Ask what people are already using before writing a policy. The answer sets the scope.
- 02A ban with no sanctioned alternative moves the same activity onto personal devices.
- 03Free tiers and paid business tiers of the same product often carry different data terms.
- GapApproved tooling is missing or slow.
- ReachSomeone finds a free tool.
- SpreadIt works, so the team copies it.
- PasteReal data goes in.
- AskAn auditor wants the trail.
The gap at the first station is the cause. Everything after it is a symptom, and policy aimed at the symptoms fails.
Approved AI, with roles and a record
Common questions
01How is shadow AI different from shadow IT?
Shadow AI is shadow IT where the payload is your data rather than your workflow. An unapproved project tracker mostly creates a licensing mess. An unapproved assistant takes in whatever a person pastes, which is why the same habit costs more.
02How do we find out what is already being used?
Ask, before you inspect. A short amnesty survey gets better coverage than network monitoring, because much of this happens on personal accounts and phones that your logs never see. Pair it with expense reports, where individual subscriptions show up.
03Is it still shadow AI if the tool is free?
Yes, and free tools are the larger share of the problem. No purchase order means no security review, no contract and no record that the tool exists. Cost was never what made it governed.
04What should a policy actually say?
Which tools are approved, what data may go into each, and who to ask when something is not on the list. A policy fails when it is all bans and no route to yes. The person facing a deadline needs an answer today, not a form.
Related
- protected health information →The data class where an unapproved tool costs the most.
- guardrails →Checks that run around a model in a sanctioned system.
- work product doctrine →Where sharing outside the relationship costs the protection.
- AI consulting →Deciding what to sanction, and what to build instead.
