Picture the Monday after a proposal deadline at a 40-person engineering firm. An estimator uploads the client’s 120-page RFP to a free AI tool on his personal account, asks for a summary of the requirements, and has the outline he needs in ten minutes. He saved a day. Nobody asked whether he could. The client’s project details now sit on a server the firm never contracted with, under terms nobody read, in an account the firm can’t close when he leaves. That is shadow AI, and every company I talk to has some.
A definition a CEO can use
Shadow AI is any artificial intelligence tool, feature or agent that touches company or client data without the company having decided that it may. The decision is the dividing line, not the technology. The same chatbot is sanctioned when it runs under a business account with retention settings chosen on purpose and a signed agreement about whether the vendor may train on your data. It is shadow when it runs on an employee’s personal login.
It arrives four ways. Personal accounts on public chatbots and writing tools. Browser extensions and desktop apps that read whatever is on the screen. AI features switched on inside software the company already pays for, usually with a routine update. And vendors or contractors who run their own AI on your files. The fourth is the one companies miss most often, because nothing shows up in your own logs.
The scale is measured now
Microsoft’s 2024 Work Trend Index, a survey of 31,000 knowledge workers in 31 countries, found 75 percent already using generative AI at work and 78 percent of those users bringing their own tools. Verizon’s 2026 Data Breach Investigations Report says the share of employees using AI tools frequently rose from 15 percent to 45 percent in a single year. IBM’s 2025 Cost of a Data Breach report found one in five breached organizations reporting a breach that involved shadow AI, at an average of $670,000 in added cost, and its 2026 report puts the global average breach at $4.99 million, with only 40 percent of organizations using access controls on their AI models and data.
When I led a shadow AI assessment at a Fortune 500 company, the inventory came back with 346 unsanctioned AI tools across 130 departments. The number surprised the executives. The pattern didn’t surprise me. Mid-market firms show the same shape at a smaller scale.
What the inventory is for
The goal is not zero shadow AI. Your people are using these tools because they work, and a ban pushes the use onto personal phones where you’ll never see it. The goal is zero unknown AI: a current list of what is in use, what data goes into it and who owns each decision. That list is what a client questionnaire, an insurer, a board member or an assessor will eventually ask for, and it’s the first step of the SAFE framework I use with leadership teams. You can build the first version in a week with tools you already own. Once it exists, it becomes the first block of the one-page AI brief a board should see every quarter.
The first inventory: five days, no new software
Day one, identity. Start where sign-ins live. In Microsoft 365, the Entra admin center lists every enterprise application and every third-party app your users have consented to, along with the permissions they granted. Google Workspace shows the same under API controls and the OAuth token audit log. Sort by AI vendors, then by anything with access to mail, files or calendars. This step alone usually finds the meeting recorders, because they ask for calendar access.
Day two, network and devices. Your DNS filter or secure web gateway already categorizes traffic, and most now carry a generative AI category, so a 30-day report of AI destinations by user takes minutes. Then pull the browser extension report from your device management tool. Extensions that summarize pages or draft replies read everything on the screen, including the client portal and the bank.
Day three, the money trail. Search corporate card statements and expense reports for AI vendors and for any subscription under $30 a month that nobody recognizes. Personal reimbursements are where the personal accounts hide.
Day four, people. Ask every department head one question: what are your people using to get work done faster? Then ask the staff directly, in a five-question anonymous survey with amnesty stated up front. This is how you learn about the tools the logs can’t see: the personal phone, the home laptop, the vendor’s assistant.
Day five, the register. Put everything in one table: the tool, who uses it, what for, what data goes in (public, internal, client-confidential, regulated), whether the account is personal or business, whether the vendor may train on the input, and who owns it. Then sort every row into one of three piles: approve as is, approve once fenced, or retire and replace.
If you hold controlled data, the register is also a scoping document
For defense suppliers the inventory does double duty. DFARS 252.204-7012 applies NIST SP 800-171 to any contractor system that processes, stores or transmits covered defense information, and it requires an external cloud service that holds that information to meet security requirements equivalent to the FedRAMP Moderate baseline. A public chatbot on a personal account satisfies neither, and the self-assessment you sign doesn’t carve it out. The register tells you which rows have to be fenced first.
Frequently asked questions
What is shadow AI, in one sentence?
Shadow AI is any AI tool, feature or agent that processes company or client data without a company decision behind it. The missing decision, not the technology, is what makes it shadow.
How is shadow AI different from shadow IT?
Shadow IT is unapproved software or hardware. Shadow AI is a subset with a twist: the data leaves in the prompt, so download and file-transfer controls never see it, and the vendor may keep or train on what was pasted in.
How do you find shadow AI without buying new software?
Use what you already have: identity consent logs, DNS or web-filter reports, browser extension inventories, card statements and expense reports, and a direct question to every department head. A first pass takes about a week. Repeat it quarterly, because vendors switch on new AI features with updates.
My team at Stealth Technology Group runs this inventory for clients as the first step of our AI strategy and governance work, and I walk leadership teams through the results in plain language. If you want to talk through your own first inventory, get in touch.