Why AI adoption statistics disagree, and which one applies to you.

Three numbers, all true

A business owner reading these on the same morning could conclude they're either far behind or comfortably normal. The Federal Reserve published a note in April 2026 explaining why the numbers disagree, and the reasons are worth knowing.

Why the surveys disagree

• They count different things. The Census survey counts firms, and most US firms are small. The executive survey weights by employment, so each large company counts for many workers. Large firms adopt AI more often, so any measure weighted by workers comes out higher.
• They ask different questions. The Census survey originally asked about AI used in producing goods or services. Others ask about any business function, or whether anyone uses generative AI at all. A company where one person drafts emails with ChatGPT says yes to the second question and no to the first.
• Different people answer. An executive may know more about AI use across a company than whoever fills in a government form, and may also feel more pressure to report that the company is using it.

The high numbers mostly measure whether AI is being tried. The Census number is closer to whether it's running inside the work.

What the Census data shows by size and sector

The Census Bureau's May 2026 release covers December 2025 to May 2026. Overall use hovered between 17% and 20% of firms. By size:

• Firms with at least 250 employees: 37%.
• Firms with 100 to 249 employees: 32%.
• Firms with 20 to 99 employees: use increased over the period.
• Firms with fewer than 20 employees: no significant change, and fewer than 20% of firms with four or fewer employees use AI.

By sector, as of early May 2026: information 39.7%, finance and insurance 33.9%, and retail trade around 14%. Looking six months ahead, 20% to 23% of businesses expected to be using AI, with finance and insurance near 39%.

A number to be careful with

A figure of 89% is widely quoted online as the share of small businesses using AI, often attributed to the US Chamber of Commerce. The Chamber's own page for its 2025 report gives 58%. If you're putting an adoption number in a board paper or a funding application, cite the primary source and say what it measured.

What this means for a company with 20 to 500 people

• You're probably not behind on trying AI. The Federal Reserve's note puts the share of workers using generative AI at work at about 41%, so some of your staff are likely using it already, with or without a policy.
• You may well be behind on running it. A minority of firms use AI inside the work that produces revenue, and in the Census data, firms with 20 to 99 employees are where that share grew over the last six months.
• Compare yourself to your sector, not to headlines. A retailer and an insurer are starting from different places.
• Measure use in operations, not tool sign-ups. The useful question is which workflow AI now handles, how often, and who checks it.

Moving from trying to running usually starts with one workflow that is repeated often, costs real hours and has a clear right answer: invoices matched to orders, calls answered and booked, documents read and filed. Our AI automation for small and mid-sized businesses page covers how we pick it, and the AI agent cost guide covers what running it costs.

How to measure your own adoption

A survey answer tells you little about your own company. Four questions tell you more, and each can be answered in an afternoon.

1. Which workflows does AI handle today, without someone re-typing its output? Name them. "People use ChatGPT" doesn't count as one.
2. How often does each one run, and how many hours a month did it take before?
3. Who checks the output, and how often do they correct it? A correction rate you can measure is the difference between a pilot and a system.
4. What data does it touch, and is that written down anywhere, including which vendors process it?

If the answer to the first question is "none", you are with most US firms in the Census data, and you have a clean starting point. If it's two or three workflows with measured hours and a named checker, you are ahead of most companies your size, whatever the headlines say.

What running it looks like in practice

The difference between trying AI and running it shows up in small, specific places:

• An invoice arrives by email. It's read, matched to a purchase order and posted to the accounting system, and only the ones that don't match reach a person.
• A patient calls after hours. The call is answered, the appointment is booked in the practice's own schedule, and the front desk sees it in the morning.
• A support ticket comes in. A draft answer, built from the company's own documentation, waits for an agent to approve it, and the approval rate is tracked each week.

Each of these has a system of record it writes to, a person who checks it, and a number that says whether it's working. That is the pattern the Census question is closest to measuring.

Sources

• US Census Bureau: Large firms with at least 20 employees biggest AI users (May 26, 2026)
• Federal Reserve: Monitoring AI adoption in the U.S. economy, FEDS Notes (April 3, 2026)
• US Chamber of Commerce: Empowering Small Business report, fourth edition (August 18, 2025)