Kief Studio

Shadow AI: 82% of Surveyed Organizations Found AI Agents Nobody Remembered Turning On

Automate

  • Published
  • Updated
  • 10 min read
Soft violet-magenta and gold curtains of light ripple across the upper part of a dark frame above an empty black horizon.

TL;DR: Shadow AI is AI that runs in your business without anyone deciding to run it. In a 2026 survey of 418 IT and security professionals, 82% had found AI agents they did not know about in the past year. A small business fixes shadow AI with a list, an owner for every agent, and a monthly check, not a ban.

Updated October 2026: the 82% figure comes from a vendor-commissioned survey of IT and security staff, not from all companies, and claims we could not source were removed.

Here is a question almost no small business owner can answer: how many AI agents are running inside your tools today, and who decided to turn each one on? That gap is what people call shadow AI, and it is not a big-company problem.

Not "do you use AI." You probably do. I mean the actual agents. The helpers inside your invoicing app, your email, your CRM and your project board that draft replies, move data, tag records and start tasks while you are doing something else. Who scoped those? Who is on the hook when one of them does the wrong thing at 2am on a Saturday?

For most small teams, the honest answer is nobody. That is not a personal failing. It is where the software went.

What did the 82% shadow AI survey actually find?

In April 2026 the Cloud Security Alliance published a report called "Autonomous but Not Controlled." Its headline finding: "While 68% report high confidence in their visibility, 82% have discovered shadow AI agents in the past year." (Cloud Security Alliance (opens in a new tab))

Who was asked, and who paid for it

Read the fine print before you repeat the number. The report is "based on 418 responses from IT and security professionals," and it was "Commissioned by Token Security," a company that sells agent security. For 41% of respondents, finding an unknown agent "happened more than once." (IT Brief (opens in a new tab)) CSA's own page confirms the sponsor: "This survey report, commissioned by Token Security, explores how organizations are currently managing AI agent governance." (Cloud Security Alliance (opens in a new tab))

So treat it as directional. A vendor that sells the fix paid for the survey, and the people answering were IT and security professionals, not owners of five-person shops. The original version of this note said "82% of companies." That overstated it. It is 82% of the organizations those respondents work at.

Where the unknown agents turned up

The hidden agents were "most commonly found in internal automation or scripting environments, cited by 51% of respondents," and "LLM platforms, including custom tools, assistants and plugins, followed at 47%." The same study found "65% of respondents had experienced at least one AI agent-related incident in the past year." (IT Brief (opens in a new tab))

Put two numbers side by side: 68% confident they could see their agents, 82% surprised by one anyway. The risky state was never "we have no AI." It is "I'm pretty sure I know what my tools are doing."

Why is shadow AI a bigger problem for a small business?

The organizations in that survey have people whose job is knowing what runs on their network. They still got surprised four times out of five. A small business usually has nobody with that job. Shadow AI does not need a big network to cause trouble; it only needs one tool with an agent switched on and nobody watching it.

How shadow AI gets in without a decision

Over the last two years, most business software grew an AI feature. Some arrive as an opt-in button, some as a new default after a plan change, some as a browser extension a staff member installed to save ten minutes. Nobody sat down and chose to run a fleet of agents. You renewed some subscriptions and said yes to some prompts.

AI use is spreading fast across American firms, which means more of these switches are getting flipped. The Census Bureau's 2026 AI supplement found that "18% of firms used AI in a business function, rising to 32% on an employment-weighted basis," and "adoption is expected to reach 22% within six months." (U.S. Census Bureau (opens in a new tab)) Each of those firms has at least one tool where an AI feature is on. Some of them know it.

Agents act, they don't just store

An old shared spreadsheet only sits there. An agent does things. It can pull records from your customer list, send them to an outside service, write a summary and email it, with no person looking at any step. That is useful when you meant it to happen. It is a problem when you did not.

OWASP, the open security project, has a name for the failure. Its list of top risks for AI applications includes "Excessive Agency," and it traces the root cause to "excessive functionality," "excessive permissions," and "excessive autonomy." (OWASP GenAI Security Project (opens in a new tab)) Shadow AI is usually all three at once: an agent with more reach than its job needs, permissions nobody reviewed, and no human checkpoint. Our agentic security article covers the technical side.

Is shadow AI just drift with a new name?

This is the reframe I want you to take away. The agents are not the danger. Drift is.

Real automation is something you decided to run, scoped to a job, with a name next to it. Shadow AI is the opposite: software doing work nobody catalogued, scoped or signed off on.

AI multiplies a messy process

AI does not fix a messy process. It speeds it up. If your intake or your follow-up only ever lived in one person's head, and an agent takes it over, now nobody can explain how it works. Not even that person.

RAND looked at why AI projects go wrong by interviewing 65 experienced data scientists and engineers. The report opens with a hedge worth keeping: "By some estimates, more than 80 percent of AI projects fail," which it describes as twice the failure rate of IT projects without AI. (RAND (opens in a new tab)) That 80% is a borrowed estimate, not RAND's own measurement, so I would not quote it as a fact about your business. The pattern RAND describes is the useful part: projects fail on unclear purpose and on the organization around the technology far more often than on the technology itself.

We wrote about the same pattern in half the automation in your business exists only in one person's browser tabs. Shadow AI is that problem with a model attached.

What is Microsoft doing about shadow AI?

If you want a sense of how normal this has become, look at what the biggest office software vendor shipped. The Microsoft 365 admin center now has a Shadow AI page. Microsoft's documentation says it "helps IT administrators discover, monitor, and govern unmanaged AI agents used within their organization," and defines shadow AI as "consumer-facing AI applications and standalone agents deployed across your organization without IT visibility or approval." (Microsoft Learn (opens in a new tab))

What the shadow AI page can and can't see

It is useful, and it is narrow. As of Microsoft's October 2026 documentation, the page is a public preview inside the Frontier program. It needs Microsoft Defender for Endpoint on your devices, a "Microsoft 365 E5 license to view Shadow AI Agents," and Intune enrollment for managed Windows devices. It detects a short list of desktop agents and can block only one of them, and "Shadow AI blocking currently apply only to managed Windows devices enrolled with Microsoft Intune." (Microsoft Learn (opens in a new tab))

So the tool at the top of the market finds some shadow AI, on some machines, at the highest license tier. It does not see the AI feature someone switched on inside a web app, or the Mac in the front office. Most small businesses do not run E5 and Intune at all. "I'll just keep track of it myself" was never a real plan, and a dashboard alone is not one either. We looked at the cost side of Microsoft's approach in Microsoft just started charging per employee to govern AI agents.

How do you find the shadow AI in your own tools?

You do not need enterprise software to get most of the way there. You need an afternoon and a spreadsheet. Here is the order I would do it in.

A one-afternoon shadow AI inventory

  1. List every paid tool. Pull the last three months of card and bank statements. Every recurring software charge goes on the list. Most owners find two or three they forgot.
  2. Open the admin settings of each one. Look for anything named AI, assistant, agent, copilot, automation or "smart." Write down what is on and what it can touch: email, files, customer records, payments.
  3. Check connected apps. In your email and file storage admin pages, look at third-party apps with access. Each one is a door. Remove the ones nobody can explain.
  4. Ask your team one question. "What AI tools or extensions do you use for work, including free ones?" Make it safe to answer. You want the list, not a confession.
  5. Check browser extensions on the shared machines. An extension that reads every page can read your web apps too.

Questions for every agent on the list

  • What job is it doing, in one sentence?
  • What data can it read, and what can it change or send?
  • Who turned it on, and who owns it now?
  • Where does its output go before a customer sees it?
  • If it stopped tomorrow, would anyone notice?

If nobody can answer the third question, that agent is shadow AI. If the answer to the last question is "no," turn it off.

Why ban shadow AI when you can own it?

Your first instinct might be to shut every AI feature off. I would resist that. Block the tools people find useful and they route around you into personal accounts and free extensions, which is shadow AI you can see even less of. My wife Amelia, Kief Studio's co-founder and CEO, makes the case in govern shadow AI by enabling it (opens in a new tab): give people a sanctioned way to do what they were already doing.

Give every agent a name

The NIST AI Risk Management Framework, which is free and written for organizations of any size, says the same thing in standards language. Its govern function calls for "Mechanisms are in place to inventory AI systems," and for "Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented" so they "are clear to individuals and teams throughout the organization." (NIST AI RMF Playbook (opens in a new tab))

For a small business that translates to two moves. Ownership: every agent doing work for you has a human name next to it. Consolidation: you run a set of tools you chose, not a pile of defaults you inherited. Our knowledge base article on AI agent governance for a small business walks through a simple version.

Put a person between the agent and the customer

For anything that leaves the building (an email, a quote, a post, a payment), keep a person in the loop until the agent has a track record. Then check that track record monthly. Agents that touch customer data deserve a written rule on what staff may paste into them; I covered that in can employees put customer data into AI tools (opens in a new tab).

Make it a monthly ten minutes

Shadow AI comes back with every software update. Put a recurring ten-minute check on the calendar: new tools on the card statement, new AI settings in the big three apps, new connected apps. Ten minutes a month beats a surprise from an agent nobody remembered turning on.

Frequently asked questions

What is shadow AI?

Shadow AI is any AI tool, feature or agent working in your business without anyone approving it or keeping track of it. It includes AI features switched on inside software you already pay for, not only apps staff install themselves.

Did 82% of companies really find unknown AI agents?

Not of all companies. In a Cloud Security Alliance survey of 418 IT and security professionals, commissioned by Token Security, 82% said their organization had discovered shadow AI agents in the past year. Treat it as directional.

Can Microsoft find shadow AI for me?

Partly. Microsoft's Shadow AI page is a preview that needs an E5 license, Defender for Endpoint and Intune, detects a short list of desktop agents and blocks only on managed Windows devices. It does not see AI features inside web apps.

Should a small business ban AI tools?

Usually not. Bans push people toward personal accounts and free extensions you can see even less. Pick the tools you sanction, give each agent an owner, and set a rule on what data may go into them.

How often should I check for shadow AI?

Do a full inventory once, then a ten-minute check every month: new charges on the card, new AI settings in your main apps, and new connected apps in email and file storage.

What is the first step with no IT department?

List every software charge from the last three months, then open each tool's admin settings and write down which AI features are on, what they can touch, and who owns them.

Where we fit

Kief Studio is in Shrewsbury, and our AI automation work starts where this note ends: pick the workflow, define what good output looks like, and measure how it runs today. We build automation with quality gates on every output, keep a human path for anything the model is unsure about, and run it for you on LTFI, reporting what it handled and what it sent to a person. When the job has several steps, our multi-agent workflows give each agent one task, only the tools it needs, and a log of every action, so nothing runs that nobody can name.

We do not build projects meant to replace staff with AI. We build systems that remove bottlenecks, and your people drive the work.

First conversation is free. No commitment.

Tell us which tools you are not sure about

Related articles

  1. Protect

    2026 Predictions for AI Agents and Web Safety: What Held Up

    We made 2026 predictions about AI agents and web safety. What the record shows in October 2026, what we got wrong, and how a small business should prepare.

    Amelia Gagne

  2. Automate

    AI Approval Workflow: What an Agent Asks Before It Acts

    An AI approval workflow for small business: what agents do alone, what waits for a yes, what only a person does, and how to stop approval fatigue.

    Brian Gagne

  3. Automate

    AI Agents for a Worcester Small Business: What They Do, What They Cost, Where They Fit

    What AI agents for small business do, what they cost, and where a person stays in the loop, for Worcester and Central Massachusetts owners. Cited and practical.

    Amelia Gagne