Christopher Wheat, Chi Mac, and Andrea Passalacqua at JPMorgan Chase Institute pulled de-identified bank records from Chase Business Banking accounts. They did not send a survey. They counted what small business owners actually spend on AI, and the answer was 17.7%. Goldman Sachs asked owners the same kind of question around the same time and got 76%.
Fifty-eight points separate those two numbers. That is not a rounding error. It is a measurement failure, and the research tells us exactly what causes it.
The Platitude
"AI is transforming small business." You have heard it at conferences. You have read it in every second headline. You have watched peers nod along as if they all run some new system you missed.
The pressure is familiar. A new tool arrives, the press calls it a revolution, and every owner who has not adopted it feels like the last one standing.
It sounds right. It is not.
What the Data Show
The research on this is worth a careful look.
Goldman Sachs polled small business owners in early 2026 and got 76%. Thryv ran a similar survey in 2025 and got 55%. The U.S. Census Bureau's Business Trends and Outlook Survey, tracked by the Federal Reserve, put the real figure between 17% and 20%. JPMorgan's bank records landed at 17.7%.
Two kinds of measurement. Two very different answers.
There is a name for this in the literature. Researchers Ling and Imas, in a study published through CHI 2026, measured the gap head-on. When people face social pressure to report using AI, the self-reported number inflates by up to 40 percentage points. The force behind it is called social desirability bias: the pull to answer a survey the way you think you should, not the way things are.
The Federal Reserve flagged the same pattern in its April 2026 FEDS Note. Senior leaders, the Fed found, face added pressure to report AI use as a mark of good management. The note named social desirability bias as a likely cause of the inflated numbers.
The headline number does not fail people. It fails the system that those people try to measure against.
The Structural Flaw
Surveys record speech. Transactions record spending. The flaw is not in the person who fills out the form. The flaw is that one tool captures what an owner says, and the other captures what an owner funds. Those are not the same signal. The gap between them is where the panic lives.
What Real Adoption Looks Like
Start with what is true. Then build from there.
JPMorgan's data show that 72.5% of small businesses that pay for AI use one tool. Not a suite. Not a stack. One subscription. Only 9.4% use three or more.
The median entry cost for businesses that started paying in 2024 was $20 per month. By 2025, that had risen to $29. Compare that to the 2019 group, which entered at $50 per month.
Even Goldman Sachs's own survey found that only 14% of respondents had fully put AI to work in core operations.
The word "transformation" does not describe a single $20 subscription.
Once that is clear, three moves follow from it.
Move 1: Pick One Workflow
Look at the past month of your work. Find the task that eats the most hours and earns the least per hour. That is your test case. Not the most visible task. The highest-volume, lowest-return task.
This requires sitting with your own calendar and being honest about where time goes. That is harder than picking a tool.
Move 2: Run One Paid Tool
Choose one tool. Pay for it. Run the task through it for 30 days. Do not add a second tool. Do not test three platforms at once. One workflow, one tool, one month. The goal is not to find the best AI product on the market. The goal is to find out if any AI product earns back its cost on that specific task.
Move 3: Measure the Return
At the end of 30 days, count two things: hours saved and dollars kept. If the tool saved five hours at your effective rate, subtract the cost of the tool. If the net number is positive, keep it running. If not, cancel and test a different task.
No survey will tell you whether the tool works for your operation. Your bank statement will.
What the System Shows
Running this for 30 days does something the adoption headlines never did:
You see your actual hourly cost on the task you tested. You see whether the tool handles your real work or just a clean demo version of it. You see the gap between what the vendor promises and what the output still needs you to fix by hand. And you see, in dollars, whether that subscription earns its spot or sits unused.
The Feedback Loop
At the end of that first month, ask three things.
→ What moved the number? Which part of the workflow got faster or cheaper in a way you can measure?
→ What looked like progress but left no trace? Did the tool create output that still needed your full review, costing time instead of saving it?
→ What friction showed up more than once? Was it the tool, the way you fed it input, or the fact that the task was wrong for the test?
That is the difference between advice that sounds right and a system that proves itself.
The best candidate for your next role might not live in the same country. Oyster helps you hire globally in 180+ countries. Payroll, compliance, and benefits included.
Where You Stand
Wheat's team did not ask owners what they use. They looked at what showed up in the bank record. That method is open to you right now.
The adoption number is not 76%. It is not 55%. When you count what people spend instead of what people say, it is 17.7%.
Your bank statement has always been a better scorecard than a survey headline.
