Pax8 released its Q3 SMB AI Pulse on October 12, based on 1,150 small business leaders across the United States, United Kingdom, Germany and Sweden. Two-thirds of AI-engaged small businesses have built solutions that are now in regular use. And 35 percent have discovered unapproved workflows, tools or agents running in their operations. Not employees pasting things into a chat window, which is the version of this problem everyone already knows about. Things that were built, connected to business systems, and are running.
What they are building
The construction numbers are the striking part. 52 percent are developing automated workflows that connect AI to other business systems. 46 percent are building or testing AI agents that complete multi-step tasks on their own. 39 percent have built custom integrations against an AI provider's API. These are not pilots in a sandbox. They are small businesses wiring a model into the systems that run the company.
Oversight is not keeping up. Formal, documented review of internally built AI tops out at 45 percent across the markets surveyed. And more than half of decision makers report spending hours per week checking AI's work, which is the quiet cost nobody puts in the business case.
One note on reading the figures: the UK subset differs from the global headline, with 54 percent having developed solutions in regular use and 24 percent finding unapproved AI in the past six months. The scope differs, so the two sets should not be mixed.
How the risk has changed
Shadow AI used to mean an employee with a personal account and a bad habit. The exposure was data leaving: someone pasting client information into a free tool. Serious, but bounded, and addressable with a policy and a conversation.
What this survey describes is different in kind. An unapproved workflow or agent is a thing that acts. It has credentials. It reads from and writes to real systems. It runs on a schedule or on a trigger, often after the person who built it has moved on to something else. The failure modes change accordingly: a workflow that keeps emailing customers after a campaign should have ended, an agent with standing access to a mailbox because that was the quickest way to make it work, an integration whose API key sits in a spreadsheet. Nobody is being reckless. Somebody solved a problem on a Thursday and never told anyone.
This is the same gap we wrote about in the services gap a week ago, seen from the other side. There, the issue was licences bought and never adopted. Here it is the opposite: enthusiastic building with no review behind it. Both come from the same absence, which is that nobody owns the decision.
What to actually do
- Go find them. Ask directly, with no blame attached: has anyone built an automation or connected an AI tool to our systems? The answer is usually yes, and usually more than one.
- Inventory by credential, not by tool. For each automation, write down what account it runs as and what that account can reach. That single column tells you nearly everything about your actual exposure.
- Give every automation a named owner. Not a department, a person, and specifically someone who still works here and knows it exists.
- Review before it touches anything live. A two-question gate is enough for most small businesses: what can this reach, and what happens when it gets something wrong. Documented, once, in a shared file.
- Scope the credentials down. An agent should have its own account with the narrowest permissions that make it useful, never an administrator login and never a departed employee's.
- Put a human in front of money and access. Payments, bank detail changes and permission grants are the two or three steps that should never run unattended, whatever the tool promises.
- Count the checking time. If someone spends five hours a week verifying output, that belongs in the ROI calculation. It may still be worth it. It should not be invisible.
The takeaway
The honest reading is that small businesses are doing something genuinely impressive here, building real automation without the teams that used to be required, and the governance simply has not caught up. The fix is not to slow the building down. It is to know what exists, what each thing can reach, and who owns it, which is roughly an afternoon of work and the difference between useful automation and a system acting on your behalf that nobody is watching.