Ask your security team how many AI tools are in use across the company. Whatever number they give you, the real one is higher. That is not an insult, it is the defining feature of the problem.
The Verizon 2026 Data Breach Investigations Report found unsanctioned AI tool use tripled in twelve months, rising from 15% to 45% of the workforce. IBM research attributes around 43% of AI-related breaches to shadow AI. And research from the Cloud Security Alliance found that roughly 89% of enterprise AI usage is invisible to security teams, with unauthorised tools remaining active for a median of 403 days before detection.
Why Sensible People Do This
Nobody wakes up planning to leak data. They have a deadline, an approved tool that cannot do the thing, and a free one that can.
Reported figures suggest a majority of employees have pasted sensitive information into personal chatbots. That behaviour is not a training failure, it is a friction signal. Every instance of shadow AI is a note telling you your approved stack is missing something.
Where to Actually Look
- Your expense reports. The cheapest audit available. Search reimbursements for AI vendor names and subscription charges under twenty pounds.
- Browser and DNS logs. Domain-level visibility catches the majority of web-based tools without deploying anything new.
- OAuth grants in your identity provider. This is the one people miss. Every app someone authorised with a work account is listed there, including ones that read mail and files.
- Browser extensions. Under-audited and frequently over-permissioned.
- Your own product analytics. AI-generated content often carries recognisable patterns in support replies and documentation.
Start with OAuth grants and expense reports. Between them you will find more in an afternoon than a scanning tool finds in a month.
What to Do With What You Find
Resist the urge to send a stern all-hands email. It generates a week of compliance theatre and drives the behaviour further underground.
| What you found | Response |
|---|---|
| A tool many people use for real work | Approve it properly or buy the equivalent |
| A tool with broad OAuth scopes | Revoke today, then ask who needed it |
| One person, one tool, sensitive data | A quiet conversation, not a policy |
| A tool nobody remembers authorising | Revoke and add to the monthly review |
The Prevention That Works
Two things reduce shadow AI more than any control: a genuinely useful approved tool, and an approval process measured in days rather than months.
If your approval takes six weeks, you do not have a policy, you have a queue that people route around. Publish the SLA, then meet it. That single operational commitment does more than any detection product you can buy.
Conclusion
Assume shadow AI exists and go looking, starting with OAuth grants and expense reports because they are free and fast. Treat every discovery as a product gap rather than a discipline problem, approve the useful tools quickly, and revoke over-permissioned access the same day. Detection alone never fixes this. Reducing the reason people go around you does.
Frequently Asked Questions
Should we block AI tools at the network level?
Blanket blocking pushes usage onto personal devices, where you have no visibility at all. Block genuinely risky tools, approve good alternatives, and monitor the middle.
What is the single highest-risk pattern?
An AI tool granted broad OAuth access to email or file storage. It is quiet, persistent, and it reads far more than anyone intended.
How do we handle AI features inside tools we already bought?
Review them at renewal and check what data trains what. Embedded features are the most common blind spot, because nobody thinks of them as adopting a new AI tool.