The seven percent

The Federal Reserve publishes a survey almost nobody in my line of work has read. It is called the Small Business Credit Survey. The 2025 round was fielded from September 3 to November 14 of last year, it collected 6,525 responses from employer firms with 1 to 499 people in all fifty states, and the report came out on March 3, 2026.

Two sentences in it explain my entire business better than I can.

Forty-six percent of firms reported that the business or its employees currently use AI. Of those that use it, just seven percent had fully integrated it into the business.

Read them together. Almost half of small employers are already using this. Among the ones using it, thirteen out of fourteen have not finished. Another 44 percent are partially integrated and about half are still calling it experimenting, which in most of the shops I walk into means one browser tab, on one person's laptop.

What "not finished" looks like from the inside

It looks like this. The owner has an account. The office manager has heard about the account. The estimate still gets typed once into the quoting tool and once into the invoice, by hand, on a Thursday, because the two things were never connected and nobody had a spare afternoon to connect them.

Nothing there is broken. The software works. The subscription is current. The task it was bought to remove is still being done by a person, which means the money went out and the hour never came back.

That is the seven percent problem, and it is not a technology problem. Everything needed to fix that shop existed two years ago and cost about forty dollars a month.

The category is selling the wrong half

Look at how this gets sold. The demo. The capability. The list of what the model can do, run on somebody else's data, in a controlled twenty minutes.

The demo has not been the hard part for a long time. What the demo skips is every part that comes after: the shape of the actual data, the person who has to open the thing on a Monday, the step where it does something wrong and somebody has to know whether that is normal.

The Fed asked about that directly, and the answers are worth writing down. Among firms already using it, the top two challenges were accuracy at 46 percent and adapting the tools to meet business needs at 43 percent. Among firms planning to start within twelve months, the top two were finding tools that meet business needs at 54 percent and the time required to implement or train employees at 37 percent.

None of those is a limitation of the technology. Three of the four are implementation, and the fourth is knowing when to trust the output.

A separate San Francisco Fed brief published July 15, 2026 asked small business owners the same question in their own words and got the same answer back: barriers included time and capacity constraints related to staff training and system upgrades, and knowledge gaps regarding implementation strategies.

Not cost. Not capability. Training and implementation, which is a polite way of saying that somebody has to stay in the building after the demo ends.

What the seven percent probably did differently

The survey does not say, so what follows is my read from the shops I have been inside, and I would rather label it than dress it up as a finding.

They picked one job, not a category. Not "use AI in the business." One named thing that happens every week and that somebody hates. Crew scheduling. Intake. The follow-up nobody makes.

It went into one named person's hands. Not the company's. A person, with a name, who is now the one who runs it and knows what it is for. A system that belongs to everybody belongs to nobody, and it dies the first week somebody is on vacation.

Somebody showed them where it breaks. This is the part almost nobody does, and it is the part I care most about. When I hand a system over, I break it in front of the person who owns it. I feed it the bad input. I show them the wrong answer it produces and what that wrong answer looks like, so the first time they meet it is standing next to me and not alone on a Friday afternoon with a customer waiting.

It kept working when the builder left. That is the actual finish line. Not the day it works. The day it works without the person who made it.

The part that is bad for vendors

If the job is finished when somebody else can run it without you, then the job ends. That is inconvenient for anyone whose business model is a permanent seat at the table, and it is the reason the handover keeps getting skipped by people who are otherwise perfectly competent.

I would rather be hired again than kept. It is a worse recurring revenue story and a much better business, and it is the only version of this I know how to defend to somebody's face.

How to tell which one you are

One question, and it takes about four seconds. Name the person at your company who would notice within a day if the thing stopped working.

If a name comes out, you are probably in the seven percent. If the answer is a department, a vendor, or a pause, you are in the ninety-three, and the fix is not a better tool.

Sources

Federal Reserve Banks, 2026 Report on Employer Firms: Findings from the 2025 Small Business Credit Survey, published March 3, 2026. Survey fielded September 3 to November 14, 2025; 6,525 responses from firms with 1 to 499 employees. Read directly at fedsmallbusiness.org, verified August 27, 2026.

Federal Reserve Bank of San Francisco, Rocio Sanchez-Moyano, Natalie Holmes and Sarah Simms, AI Adoption Among Small Businesses: Qualitative Insights from the Small Business Credit Survey, published July 15, 2026, drawing on the 2024 Small Business Credit Survey. Read directly at frbsf.org, verified August 27, 2026.