Why Small Businesses Get Stuck Experimenting With AI

Something quietly vanished from small business AI research this quarter.

Pax8 surveys US small business leaders every three months about their technology decisions. Between the first and second quarter of 2026, the share who said they were "interested but haven't started" with AI dropped from 9 percent to 1.5 percent.

That's an 83 percent collapse in three months. The fence has emptied.

The more useful finding is what happened to those people. They didn't become AI adopters. They became AI experimenters. Nearly one in three small businesses using AI right now are stuck in testing, unable to move anything into the actual business.

Pilot purgatory. You tried the tool. It worked. Nothing changed.

If that describes your company, the survey has something worth knowing about why.

When Pax8 asked stuck experimenters what was blocking them, the top three answers were lack of internal expertise at 28 percent, cost or unclear return at 24 percent, and security concerns at 21 percent. Reasonable answers. Probably not the real one.

Look instead at what separates the businesses that moved forward. Among small businesses actively using AI, 91 percent say leadership is aligned on what AI is for. Among experimenters, that drops to 68 percent. Among non-users, 32 percent.

Then the governance number. Only 11 percent of experimenters have a written AI policy. Across all small businesses surveyed, it's 23 percent.

Those two numbers explain more than the barrier list does. Not budget. Not expertise.

Here's where I want to slow down, because "write an AI policy" is the most boring advice available, and most people hear it as a compliance chore.

It isn't. A policy is a decision. The reason so many pilots stall is that nobody ever made one.

Think about what testing actually is. Testing is permission to not decide. You can run AI on your invoices for six months and never answer whether it's part of how you do invoices now. Nobody owns it. Nobody has to defend it. And nobody on your team knows whether to build around it or wait for it to disappear.

That ambiguity costs you in ways that never show up on a line item. Your bookkeeper keeps doing the manual version in parallel, just in case. Your operations lead won't rewrite a process around a tool that might be gone by fall. You're paying for the software and the old workflow at the same time.

So the fix isn't a bigger budget or an outside expert. It's one page.

Write down three things. What AI is allowed to touch in your business. What it is not allowed to touch, specifically, including whatever your version of customer financials or employee records happens to be. And who gets to change those two lists.

Then pick exactly one workflow you've been testing, put a person's name next to it, and set a date when the old way stops. Not a pilot extension. A retirement date for the manual version.

That's a Thursday afternoon of work. It's also the thing the data keeps pointing at.

One caveat, because I don't want to oversell someone else's research. Pax8 sells through managed service providers, and its conclusion is that stuck businesses should hire a technology partner. That's a convenient finding for a company whose customers are technology partners. The barrier data holds up on its own. Read the recommendation with the source in mind. Plenty of small businesses get unstuck without hiring anybody.

There's one more number in the report I keep returning to. Among businesses that slowed or cut technology spending this year because of the economy, 49 percent still say they won't be competitive within three years without AI.

Half the people pulling back believe the thing they're pulling back from is existential.

That isn't stupidity. That's what it looks like when conviction outruns capacity. You believe it matters, you don't have room to act on it, so you keep testing. Testing is cheap. Testing feels like progress. Testing asks nothing of anyone.

The businesses that got past it didn't find more room. They stopped treating the decision as something they could postpone until the tools settled down.

The tools are not going to settle down.