The Census Bureau's newest number: 24.2% of US employer businesses said they used AI in the two weeks from Oct. 5 to Oct. 18. That comes from cycle 202622 of the Business Trends and Outlook Survey, published Nov. 5 (Census BTOS, National.xlsx). Call it about 1 in 4.
If you run a six-person shop and wonder whether you're behind, this is the right survey to check. It's also rougher than the headline sounds.
One housekeeping note first. Search census.gov for "AI" and you'll come up short. After Executive Order 14434, signed Sept. 29 and published Oct. 2, told agencies to use "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI," the BTOS site relabeled its charts "SI (formerly AI)." Census says the instrument wording changed "beginning with the October 19, 2026 collection" (Census BTOS About). Census kept it in the same time series. Search "SI."
The question, word for word
Here's the current-use question as Census prints it in the data file (Census BTOS, National.xlsx):
In the last two weeks, did this business use Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI) in any of its business functions? (Examples of SI use include: SI-powered data analysis; generative SI for marketing, product design, or coding; virtual SI agents for customer service; SI-powered forecasting or pricing; SI-driven computer vision for quality control, etc.)
The answers are Yes, No and Do not know. It's a yes or no on use, not on how much, how well, or whether it paid off. One AI scheduling tool and a fully automated back office both count as a single yes.
Who answers? Census contacts sampled businesses by letter or email, and someone at the business logs in and reports online. It's voluntary and Census estimates about nine minutes (Census BTOS respondent page). Census doesn't require the owner to answer. Responses are weighted to represent US employer businesses outside farming, with nonresponse adjustments by state, sector and size (Census BTOS methodology). Census also files BTOS under its experimental data products (Census experimental data products).
Why "1 in 4" needs an asterisk
The wording changed a year ago. Until fall 2025 the question asked about AI "in producing goods or services." From Nov. 17, 2025 it says "any of its business functions," and Census started a new time series because of "a level shift observed in conjunction with the new question wording" (Census wording update). The old question's last reading was 10.0% for Sept. 8 to Sept. 21, 2025 (Census, AI Core Questions.xlsx). The new one opened at 17.3% for Nov. 3 to Nov. 16, 2025 (Census BTOS, National.xlsx). Some of that jump is adoption and some is wording, and the survey can't split the two.
The latest cycle is mostly noise. The national figure slipped from 24.6% to 24.2%, with a published standard error of 0.61 points (Census BTOS, National.xlsx). That's a wash. The bigger move: "Do not know" jumped from 9.6% to 13.9% in the first cycle with the SI wording. Maybe "Super Intelligence" confused some respondents. It's one cycle, and we can't prove a cause.
Fewer businesses answered. The 1.2 million figure is the sample, not the responses. Census splits it into six panels of about 200,000, and each panel reports once every 12 weeks (Census, Nov. 5 release). Cycle 202622 got about 17,500 responses, down from 24,000 in 202621. The unit response rate fell from 12% to 9% (Census, URR.xlsx).
Find your row: size
Share answering yes, by employee count (Census, Employment Size Class.xlsx):
| Employees | Oct. 5 to 18 (202622) | Sept. 21 to Oct. 4 (202621) |
|---|---|---|
| 1 to 4 | 22.8% | 24.3% |
| 5 to 9 | 23.4% | 22.6% |
| 10 to 19 | 25.3% | 24.6% |
| 20 to 49 | 29.2% | 26.0% |
| 50 to 99 | 33.2% | 29.0% |
| 100 to 249 | 40.0% | 35.3% |
| 250 or more | 44.1% | 43.6% |
| All businesses | 24.2% | 24.6% |
The smallest firms are barely behind the average. Businesses with 1 to 4 employees sit 1.4 points under the national number (24.2 minus 22.8 = 1.4). The separation builds from about 20 employees up. From the smallest row to the largest is 21.3 points (44.1 minus 22.8 = 21.3).
Don't read the rows under 20 employees as a staircase. In the prior cycle, 5 to 9 came in below 1 to 4. This cycle it's a hair above. Those rows sit within a few points of each other and swap order between cycles. Illustrative math, using the standard error Census publishes for the 1 to 4 row (0.86 points) and a 90% interval: 1.645 × 0.86 = 1.41, so roughly 21.4% to 24.2%.
Bigger firms using AI more doesn't mean AI made them bigger. A 300-person company has more functions where a tool could show up, so more chances to say yes.
Find your row: sector
Sector moves the number far more than headcount does. Latest cycle, 202622 (Census, Sector.xlsx), with response counts (Census, URR.xlsx):
| Sector | Using AI | Standard error | Responses |
|---|---|---|---|
| Information | 45.0% | 3.04 | ~350 |
| Professional, scientific and technical | 41.6% | 1.34 | ~2,700 |
| Finance and insurance | 39.2% | 2.51 | ~700 |
| Real estate | 28.2% | 1.71 | ~800 |
| Health care and social assistance | 24.6% | 1.27 | ~1,900 |
| Administrative and support services | 22.2% | 2.33 | ~850 |
| Retail trade | 19.1% | 0.99 | ~1,900 |
| Construction | 17.1% | 1.19 | ~1,400 |
| Other services (repair, personal care) | 15.7% | 1.13 | ~1,000 |
| Accommodation and food services | 12.1% | 0.96 | ~1,500 |
A restaurant owner comparing herself to the national 24.2% is grading against a curve built partly by software firms and consultants. Her peers sit at 12.1%. Information has the widest error bar on the list, on about 350 responses.
Walkthrough: find your exact cell
Hypothetical: a seven-person HVAC contractor. HVAC falls under construction in NAICS.
- Go to the BTOS data downloads page and grab Sector by Employment Size Class.xlsx.
- Open it in Excel or Google Sheets, on the Response Estimates tab. Turn on filters.
- Filter Sector to 23 (construction), Empsize to B (5 to 9 employees, per the file's data dictionary), Question ID to 7, Answer to Yes.
- Read the column headed 202622. Check the Response Standard Errors tab for the same row.
Our contractor's cell: 19.2%, standard error 1.17 (Census, Sector by Employment Size Class.xlsx). So roughly 1 in 5 comparable firms said yes. Cells marked S aren't published, and small cells swing hard between cycles, so read three cycles before deciding anything.
Question 8, new this cycle, asks whether employees used generative SI, naming ChatGPT, Gemini, Copilot, Claude and others. Nationally 26.8% said yes (Census BTOS, National.xlsx). If your crew drafts quotes in ChatGPT, that's the closer comparison.
What the gap actually means
Say your row reads 19% and you're a no. About four in five of your peers didn't say yes either. A benchmark tells you what's normal. It doesn't tell you what's worth doing.
The owner polls we've covered put AI use far higher than Census does. Those polls ask different questions of different people, often an organization's own network. Census is the better peer benchmark because it draws a probability sample and publishes its error bars. The polls are better at telling you what owners worry about, and accuracy tops that list (our accuracy post).
So skip "am I behind" and ask the narrower question: is there a task, like dispatch, quoting or follow-up emails, where your time costs more than a tool would? If there is, a starter setup doesn't have to be expensive (our under $150 a month post). If there isn't, being in the 77% of 1 to 4 employee firms that didn't say yes is a fine place to be. (Illustrative: 100 minus 22.8 = 77.2% answered no or do not know.)
Want help finding your row and picking the one task worth automating? Write to us: team@lirium.ai.