Someone in your business has pasted a customer list into a free AI tool. You do not know who, which tool, or where that list is now. That is not a staff problem; it is a counting problem, and it is the first one to fix.
Your employees adopted AI before you had a chance to decide about it
The adoption numbers come from two directions, the employees and the businesses that employ them, and the gap between them is the problem. Asking employees, Gallup found in June 2025 that the share of U.S. workers using AI at work at least a few times a year had risen from 21% in 2023 to 40%. Frequent use, a few times a week or more, had gone from 11% to 19%, and daily use had doubled in the previous twelve months, from 4% to 8%. Asking those same employees about their organizations, Gallup found that 44% said their organization had begun integrating AI, but only 30% reported general guidelines or formal policies for its use and only 22% said a clear plan had been communicated. The fourteen-point distance between integration and standards is Gallup’s own measure; the distance between what workers do and what anyone has written down is larger.
Asking the businesses themselves produces a smaller number still. The Census Bureau’s Business Trends and Outlook Survey, summarized in a December 2024 research note, tracked the share of firms reporting AI use in their operations. Firms with one to four employees ran second only to the largest firms, rising from 4.6% to 5.8% across the year to August 2024 on the Bureau’s smoothed trend line rather than any single two-week reading, while firms with five to ninety-nine employees sat around 3% to 4%. The surveys measure different things: the owner reports what the business uses, the employee reports what they use. A business whose owner would answer “no” to the Census and whose staff would answer “weekly” to Gallup has shadow AI, and that is the position most small firms we assess are in. The tools arrived through a browser tab and a personal card, not a purchase order, so nobody was ever asked.
Uncounted tools are where the expensive incidents start
The cost of not knowing is now measured. IBM’s 2025 Cost of a Data Breach Report, based on breaches at 600 organizations between March 2024 and February 2025, found that “one in five organizations reported a breach due to shadow AI, and only 37% have policies to manage AI or detect shadow AI.” Organizations with high levels of shadow AI “observed an average of $670,000 in higher breach costs” than those with little or none. The data that leaked through unsanctioned tools skewed toward what a small business can least afford to lose: shadow AI incidents compromised personally identifiable information in 65% of cases, against 53% across all breaches, with intellectual property next at 40%. Among breached organizations, 63% either had no AI governance policy or were still writing one.
Those figures describe larger organizations than yours, and the $4.44 million global average breach cost in the same report is not your number. The pattern is. A breach through an unsanctioned tool is a breach through a system nobody was watching, holding data nobody classified, with a vendor contract nobody read. Every one of those failures traces back to the tool never having been on a list. The enterprise version of this argument is that the inventory is the control every other control depends on. The small-business version is simpler: you have no other controls yet, and this is the one you can afford.
A policy is a rule about things you know exist. The inventory is how you find out what there is to rule on.
A policy is a rule about things you know exist. The inventory is how you find out what there is to rule on.
The template already exists, and it was written for businesses without a security team
You do not need to design the inventory. In February 2024 the National Institute of Standards and Technology (NIST) published Special Publication 1300, the Cybersecurity Framework 2.0 Small Business Quick-Start Guide, written for firms with modest or no cybersecurity plans. Under its Identify function, the guide calls for “creating and maintaining an inventory of hardware, software, systems, and services,” and for “inventorying and classifying your business data.” The guide supplies a starter table with six columns: the software, hardware, system or service; its official use; its administrator or owner; any sensitive data it can access; whether multi-factor authentication is required; and the business risk if access to it is lost. It also asks the owner directly whether the technologies staff use are secure and approved. An AI tool is a service that can access data, and it belongs in that table like anything else.
Our assessment is that four columns need adding for AI tools, because they decide whether a tool can stay. Whether the vendor trains its models on what you type, under the plan you are on. How long the vendor keeps your inputs and whether you can delete them. Who in the business approved the tool, and when. And what categories of data are permitted in it, written as a short list rather than a judgment call. The first two come from the vendor’s terms and privacy policy; a free consumer plan and a paid business plan from the same vendor routinely answer them differently, which is why the plan matters as much as the product.
A two-week count the owner can run without hiring anyone
The inventory is a project with an owner, a deadline and three inputs, and the owner is you or the person who runs operations. Set a two-week window and announce an amnesty: nobody is in trouble for what they have been using, and the point is to make the useful tools official. Then pull from three sources. First, the card and expense statements for the last twelve months, which surface every paid subscription, including the ones individuals expensed. Second, the Google Workspace admin console or Microsoft 365 admin center, under connected or third-party apps, which shows what staff have linked to their work accounts, and which is where the unpaid tools appear. Third, a ten-minute conversation with every employee, asking what they use, for what, and what they paste into it. Most of the list comes from the third source, and the first two exist to catch what people forget.
Every row then gets one of three decisions, recorded in the table. Approved, with the data rules written beside it. Approved with conditions, usually an upgrade to a plan that does not train on inputs, or a restriction on which data may go in. Or stopped, with a named alternative that does the same job, because a ban without a substitute returns as shadow AI within the month. Review the table quarterly, and add a line to onboarding so that new tools enter through the list rather than around it. NIST’s Govern function describes the goal as policies you “communicate, enforce, and maintain,” and all three verbs require that the policy refer to something real.
The owners who get ahead of this will not be the ones who wrote the strictest AI policy. They will be the ones who can say, in a single page, which tools their business uses, what goes into each, and who said so. Everything else a small business will eventually need, from a customer’s security questionnaire to a cyber-insurance application to the first incident, starts by asking for that page. Build it now, while it is short.