Shadow AI in Your Small Business: The AI You Didn't Know You Were Running

Someone on your team built something with AI. Maybe it answers customer emails, pulls numbers out of your accounting system, or fills in a proposal while they grab a coffee. It works, so nobody talks about it, and you may not even know it exists. That is shadow AI in your small business, and right now it is probably running somewhere you have not looked.

Shadow AI is not a new problem with a new name. It is an old operations problem that AI has made faster, cheaper, and easier to create. This article covers what the newest data shows, why I believe every AI agent should be managed like an employee, what I have watched happen when one person's creation becomes a whole department's lifeline, and four steps to find what is running before it costs you.

What Is Shadow AI?

Shadow AI is any AI software, workflow, or agent running in your business without anyone in charge approving it. The term grew out of "shadow IT," the older label for systems employees set up on their own without IT knowing.

New research shows how common it has become. In its Q3 2026 SMB AI Pulse Report, Pax8 surveyed 1,150 business leaders at companies with five to 499 employees across the United States, the United Kingdom, Germany, and Sweden. Among businesses already engaged with AI, 67% had built at least one AI system that is now in regular use. In the past six months alone, 35% had discovered AI running that nobody approved.

The numbers for American businesses are sharper. According to the full Pax8 report, 81% of AI-engaged US businesses have built something that is now in use, and 42% found shadow AI in the last six months. Only 36% have a formal, written review of what they built. Another 41% check things consistently but keep no record of it. Of US businesses paying for AI, 57% received a bill higher than they expected, and 68% of US decision makers spend three or more hours a week checking AI's work.

Read those numbers with the source in mind. Pax8 sells AI and cloud products through IT service providers, and the report closes by telling those providers that an AI inventory is an easy first job to sell. The companies surveyed also run larger than many small businesses. The pattern still matches what I have seen for years.

One line in that report deserves more attention than it gets. It points out that a policy covering how employees use AI does not cover the workflows and agents the business has built itself. I have written before about why a one page AI policy matters for what people paste into a chat window. That policy still matters. It just stops at the chat window, and the agents your people are building live well past it.

Why Is AI Creating So Much Shadow AI?

AI has made building things easy. Someone with no programming background can describe what they want and get a working agent, a connected workflow, or a spreadsheet that reaches into other systems. AI assistants like Copilot can write Excel macros that go out and grab data, and the person using them does not have to understand a single line of the code being put in.

That ease is the problem. The people building these things are usually not trained to write programs. They do not know to put guardrails in place, and they often do not know what they are giving the AI access to. With no rules to guide them, an agent takes on the habits of whoever built it. If the builder is careful, the agent is careful. If the builder is in a hurry, the agent is in a hurry, with access to your customer files.

I see companies make one of two mistakes. Some hand everyone Copilot, tell them to go use AI, and put nothing in place to manage what it can reach. Others lock it down so tight that all anyone can do is paste text into a chat window. The second choice is safer, and it leaves most of the value on the table. The first choice is how shadow AI starts.

It looks like the Wild West. You know the scene from the old Western movies. Everyone in the saloon pulls out a gun and starts firing into the ceiling, and nobody thinks about the roof. In your business, every one of those holes is a leak, and what leaks out is your data.

Is Shadow AI Really a New Problem?

No. AI just makes it happen faster and in more places at once. I have watched the same pattern play out with spreadsheets for decades.

At a government contractor I worked for, which did work for both the Department of Defense and the Department of Energy, the business ran on Deltek Costpoint, an ERP system the government side favors. I never cared for it. One person built an Excel sheet with macros that reached into Costpoint and pulled out all kinds of data. It was useful, so it spread. The company had somewhere between 100 and 200 people on the floor, and everyone in the office knew about that sheet and wanted access to it.

IT did not know it existed. By the time IT got involved and approved it, the whole company was already using it. The man who built it maintained it himself. It was still in full swing when I left, and I do not know what happened to it after that. If its builder ever moved on, I would not bet that IT took it over. IT manages systems. I have never heard an IT group offer to maintain somebody's spreadsheet.

At another company I have worked for, one spreadsheet became the template for every proposal. It even carried the name of the person who built it. I will call it the Dooley template. Everyone used it, and proposals went out for signature on it. Then its builder retired, and someone else had to go back through the macros to work out what the sheet was pulling and where it was sending it.

Excel macros are some of the hardest work to pick up from another person. I have used Excel my entire career, and macros still blow my mind. Unless you are an Excel expert, untangling someone else's is slow, careful work.

At that same company, another spreadsheet ties into SAP and tracks every part shortage, customer request, and deviation that needs renewing. It even has forecasting built in. One person built it, and a very tech savvy colleague took it over and kept adding to it. Now an entire department gets all of its information from that one Excel file. Some people came to depend on it completely. The people who built and expanded it have left, the sheet is slowly breaking, and those jobs still run 100% on its data.

Swap the spreadsheet for an AI agent and every one of these stories gets worse. An agent does not just pull data. It takes action, and someone who could never have written those macros can build one in an afternoon. AI doesn't fix a bad process, it just makes the mess move even faster.

Should an AI Agent Be Managed Like an Employee?

Yes. An AI agent is an employee. It is an asset doing work in your business and making decisions based on logic someone else created or trained. You would never let a new hire wander through every file in the building, change records, and delete documents with no manager and no rules. An agent should not get that freedom either.

The owner has to know what is happening in the business. It is your business, and nobody else is going to worry about it. The same goes for every manager. Nobody else is going to keep track of what the people in your department are doing, and that now includes what they have built.

Employees get rules on their first day. Agents need rules too, and the rules need to exist before anyone builds one. Every agent should point to the same baseline from the start: these are the only folders it can read, these are the only folders it can write into, it never writes over a document, it never deletes anything, it archives instead, and here is exactly what can and cannot go out to the internet.

I run my own business this way. Every agent in my content system points to one core rules document before it does anything. No agent deletes a file. An agent writes only its own work and never changes another agent's output. If an instruction is unclear, the agent stops and flags it instead of guessing. A single run cannot create, move, or rename more than 20 files without my approval.

I keep those rules in a Word document, not in a format built only for AI to read. AI can read either one. People need to read the rules too, because people have to change them. Things change, and sometimes you discover a back door you left wide open and need to close it today. When the rules are written for a person, you can find the problem and fix it.

Who Takes Over When the Builder Leaves?

The Pax8 report names a risk it calls technical debt: systems that become hard to maintain once the person who built them leaves the company. That is the Dooley template. In operations language, it is owner dependency, and it is the reason every agent needs a succession plan.

Every agent someone builds has to be able to hand off to somebody else. That is not only a question for the person who owns the business. Every person in a company should be thinking about who comes after them, how they will hand off the job, and how they will train the next person. Most people do not. In most of my jobs, I have had a succession plan, or at least everything I knew kept in one place and ready to hand off.

I learned this at the very start of my career as a CNC programmer. We programmed five axis machines that cut castings the size of a car. A single program for one operation ran 50 to 100 pages, and one part could take up to seven programs: rough one side, rough the other, finish each side, then machine the two together. Every programmer had a personal method. When I was handed someone else's program, I had no idea what it was doing and had to go through it line by line.

So I started leaving notes. Before each section, I wrote what that section did: the rough milling that cleaned up the casting, the probing checks that made sure there was enough material, the drill paths. Where someone else's macro did not work right and I had to put in a manual change, I wrote a note saying exactly what it was doing and why it had to be there. The next person could open the program and find their place.

I still build that way. When I build something, for my own business or for a client, I build it so I can hand it off. The same thinking applies to the business as a whole. When an owner sells, retires, or steps back, every agent and workflow nobody can explain goes along with the business.

How Do You Find the Shadow AI in Your Small Business?

Finding shadow AI in your small business does not take new software. It takes four steps, and they only work in this order.

1. Ask what has been built. Hold a meeting with your people. Ask what AI they use, what they have built, and what they are testing. Make it safe to answer honestly, because the goal is a complete picture, not a reprimand.

2. Make the list and put a name on everything. For every agent, workflow, and AI-built spreadsheet, write down who built it, who owns it now, what it touches, and what it costs. To find the cost, look at how often it runs, since usage charges grow with every run. Start with anything that touches money, customers, or your core business system. If nobody is willing to own something, turn it off.

3. Set the baseline rules. Write down where agents can and cannot go: which folders they can read, which they can write into, no writing over documents, no deleting, archive instead, and what can go out to the internet. Make every new agent point to those rules from day one, and write them so a person can read and change them.

4. Document it and plan the handoff. Once something is running, the person who built it documents how it works, including a simple flow chart that shows every decision point. People do not draw flow charts much anymore, and they are still the fastest way for someone else to understand what a program does. Then put a date on the calendar to repeat this review, because your people will keep building.

You cannot manage AI that nobody wrote down, and somebody in your business is building more of it this week.