Ask your team whether they use AI at work and most will say "a little." Verizon's 2026 Data Breach Investigations Report counted instead: 45% of employees in its data now use AI on work devices at least once every 15 days, up from 15% a year earlier. And 67% of the people reaching AI services on work devices log in with personal, non-company accounts. Those personal accounts are the heart of shadow AI, and in IBM's 2025 breach study only 37% of organizations had policies to manage AI or detect shadow AI.
What Shadow AI Is
Verizon defines shadow AI as "employees using unapproved AI tools at work." It is rarely malicious. It is a bookkeeper pasting a messy spreadsheet into a free chatbot, or a salesperson installing a browser add-on that drafts emails. The work gets done faster. The problem is where the data goes: into accounts the business cannot see, which leave when the employee does.
Why It Matters
IBM's yearly breach study puts numbers on the risk. In its 2025 edition, covering 600 organizations' breaches, one in five reported a breach due to shadow AI, and organizations with high levels of shadow AI saw breach costs about $670,000 higher than those with little or none. In 2026, IBM says, incidents involving unapproved AI tools more than doubled to 43% of organizations, and those breaches cost an average of $5.39 million, against a $4.99 million average for all breaches.
Verizon sees the same thing from the other side. Shadow AI is now the third most common non-malicious insider action in data from its data loss prevention tools (software that flags sensitive files leaving a company), a fourfold increase in a year. Source code was the most common thing submitted to outside AI tools.
Your business may never show up in IBM's averages, but the leak path is the same: a customer list sitting in a personal account you cannot audit or delete.
Where Shadow AI Hides in a Small Business
| Where it hides | How to find it |
|---|---|
| Personal chatbot accounts ChatGPT, Claude or Gemini on a personal login |
Ask directly. It is the one check that needs no software. |
| AI browser extensions Add-ons that read the page you are on |
Open the extensions list in each work browser. Verizon found the average company had more than 15% of users with unauthorized AI extensions. |
| Meeting note-takers Bots that join calls and record them |
Check recent calendar invites and meeting attendee lists for bot names. |
| Apps signed in with Google Tools employees connected to work email or Drive |
In the Google Admin console, go to Security, then Access and data control, then API controls, and review "Accessed apps," with user counts per app. |
| AI switched on inside software you already pay for | Review the AI and data settings in your five most-used tools. |
On Google Workspace, the Google sign-in row is the fastest check. Microsoft 365 shops can ask their IT provider for the equivalent list.
Extensions deserve extra attention. Verizon notes that one of the main functions of these plugins is often to collect and keep information about what the user is browsing, so a staff member browsing your internal pages may be feeding them to a third party.
The last row catches people out. Atlassian, the company behind Jira and Confluence, started using customer data to improve its apps and AI on August 17, 2026. On Free and Standard plans, contributing "in-app data" (page text, Jira titles and comments) is on by default; an admin can switch it off under Atlassian Administration, then Security, then Data contribution. Metadata, meaning patterns and measurements about your content, stays on for every plan below Enterprise, apart from some compliance exceptions. Nobody did anything wrong; the setting was just never checked.
How to Find It Without a Crackdown
Banning AI does not remove the reason people reached for it, and it gives them a reason to stop telling you. Run an amnesty instead:
- Ask, with no penalty. A short form: which AI tools, for what, on which account, with what data.
- Run the two technical checks. The Google Admin app list and each work browser's extensions.
- Sort, do not punish. Approve useful tools on business accounts, and replace or remove only what you cannot make safe.
- Check the defaults. Review the AI and data settings of your main software, starting with anything holding client files.
- Write it down. One page on which tools are approved and what never goes into them. Our small business AI policy template takes an afternoon.
For what happens to pasted text, see where your data goes in ChatGPT.
The Bottom Line
Shadow AI is a sign your team wants better tools. Find it with questions, not punishment, then give people an approved version that is just as easy. The risk is not that your staff use AI. It is not knowing where your data went.
Sources
- Verizon, "2026 Data Breach Investigations Report" (May 19, 2026): AI use, personal accounts, data loss findings, browser extensions. verizon.com
- Verizon, "Vulnerability exploitation top breach entry point, 2026 industry-wide DBIR finds" (May 19, 2026): definition of shadow AI. verizon.com
- IBM, "IBM Report: 13% Of Organizations Reported Breaches Of AI Models Or Applications, 97% Of Which Reported Lacking Proper AI Access Controls" (July 30, 2025): 2025 shadow AI findings. newsroom.ibm.com
- IBM, "Manage AI data risk with IBM Guardium Exposure Manager" (July 30, 2026): 2026 unapproved-AI findings. ibm.com
- IBM, "IBM Study: One in Four Malicious Breaches are AI-Enabled, Costing Companies $6 Million on Average" (July 29, 2026): 2026 average breach cost. newsroom.ibm.com
- Google Workspace Admin Help, "Control which apps access Google Workspace data" (archived February 22, 2026): the Accessed apps list. web.archive.org
- Atlassian, "Data Contributions Frequently Asked Questions" (archived October 27, 2026): start date, defaults and opt-outs by plan. web.archive.org
- The Register, "Atlassian's new data collection policy protects rich customers while AI eats the rest" (April 18, 2026): metadata, in-app data and opt-out limits. theregister.com
Not sure what AI is already running in your business?
Our AI readiness assessment starts with a no-blame inventory of the tools your team uses, then turns the useful ones into approved, safe defaults.