What the AI numbers actually say

Every week brings a new number about AI adoption. Small businesses are racing to adopt it; small businesses are ignoring it; most small businesses are stuck in between. The numbers are not lying. They are answering different questions and being repeated as if they answered the same one.

Three numbers, three definitions

The Federal Reserve's 2026 Small Business Credit Survey report, published in March, found that 46% of employer firms said their business or employees currently use AI, with another 15% planning to start. The same month, a Federal Reserve note drawing on Census Bureau business survey data put firm-level AI adoption at about 18% at the end of 2025. A US Bank survey of small-business owners reports 75% using generative AI.

Read the question, not the headline. The Fed survey asks about use by the business or its employees, which includes one person trying a chatbot for an hour. The Census series measures adoption in a way that changed methodology recently, which the Fed note itself flags. The bank survey samples its own customers, who are already inclined to answer a banking survey about technology. None of these is wrong; each is a different measurement.

Adoption is not integration

The more useful finding is usually in the second paragraph. A firm where one employee drafts social posts with a chatbot and a firm whose booking, invoicing and follow-up run through a tested agent both count as 'using AI' in most surveys. They are not the same business. If you are deciding what to build, the number you want is not adoption — it is how many firms have a workflow that runs without a person babysitting it, and that number is much smaller and much harder to find.

The capability numbers have the same problem

Benchmarks tell a similar story. A February 2026 study evaluated 14 models on two agent benchmarks and decomposed reliability into twelve dimensions. Accuracy rose faster than reliability across the models tested: the systems got better at the tasks, not proportionally better at being depended on. A demo that succeeds once and an agent that succeeds every Tuesday are different products, and only one of them shows up in a benchmark table.

Three questions before you quote a number

Who was asked? A survey of a bank's customers is not a survey of the economy. What counts as use? One prompt in a browser is not a workflow. And what is the denominator? '46% of firms' means something different when the sample is employer firms with employees than when it is anyone who registered a business. If a vendor cannot answer those three, the number is marketing.

We quote research on this site, so we have a rule: every figure appears with its source and its limits, and the full list lives on the sources page. When a study is narrow — one clinic, one country, one bank's customers — the card says so. It is slower than a statistic on a slide. It is also the only way the number stays useful after the slide is gone.