Most Small Businesses Tried AI. Few Built It In.

If you own a Las Vegas law firm, a dealership, or a restaurant group, someone has probably told you that every competitor already runs on AI. Whether that is true depends on which number you read. Five measurements published in 2026 put small business AI use anywhere from 17.7% to 51.8%, and none of them is wrong. They count different things.

This post puts those five side by side, explains what each one actually measures, and separates the businesses that tried AI from the ones that built it into daily work. Every figure below was read from the primary source on September 17, 2026. The short version: a lot of businesses have tried it, and very few have wired it in.

The adoption ladder: five 2026 measurements side by side

The table orders the measurements from the loosest definition of using AI at the top to the strictest at the bottom. Read it as a ladder. Each rung down asks more of a business before it counts.

Each source is linked at the page holding its number: the Census BTOS data tables, the Fed's 2026 Report on Employer Firms, the JPMorgan Chase Institute study of AI use by small businesses, the Nevada State Bank 2026 small business survey report (PDF), and Stanford AI Index 2026, chapter 4, section 4.3 on page 197 (PDF).

The rung 5 figure is ours: 46% of firms use AI and 7% of those users are fully integrated, so 0.46 times 0.07 is about 3.2% of all employer firms. That is the number the headlines skip. From the top rung to that one, the share falls by a factor of about sixteen.

Why the numbers run from 17.7% to 51.8%

Four differences explain almost all of the spread: who gets asked, what question they answer, what time window counts, and whether money has to change hands.

The Nevada State Bank figure sits highest partly because of who answered. It is a bank sponsored survey that businesses chose to fill out, not a probability sample, and owners interested in a topic are more likely to answer questions about it. It also asks whether a business has implemented AI, with no time window attached. That is a useful read on sentiment among engaged Nevada owners. It is not an estimate for the whole state.

The Fed's 46% is also a question with no time window, put to employer firms. A business that used a chat tool to draft one job post counts the same as a business that runs its scheduling on AI. The Census question is narrower: did the business use AI in a business function in the prior two weeks. A business that experimented in the spring and stopped would say yes to the Fed and no to Census.

Census also has the strongest method. It is a probability sample with a standard error of 0.28 points on the national figure, and it records uncertainty honestly: 67.2% said no and 9.6% said they did not know.

JPMorgan Chase Institute measures something else entirely: payments. If a business paid an AI service, it counts. Free tools are invisible to it, which is why its all firm figure, about 17.7% at the end of 2025, is the lowest on the ladder. Split out, employer firms read 26.1% and nonemployer firms 15.3%. That employer figure sitting above the Census 23.2% is not a contradiction. Paying for a subscription and using it in a business function during a given two week window are different events.

The direction every measure agrees on is up. Census cycles in late 2025 sat between 17.2% and 17.8%. The 2026 cycles read 17.7, 17.5, 19.8, 19.5, 20.6, 21.7, 21.8, 22.4 and 23.2, a rise of 5.5 points since the first cycle of the year (our subtraction), and 27.3% of businesses expect to use AI within six months. JPMorgan found that the 2025 cohort of businesses reached a 10% adoption rate in six months, where the 2019 cohort took over six years.

Size cuts in an unexpected way. In the September Census release, businesses with 1 to 4 employees read 23.6%, above the 5 to 9 band at 21.0% and the 10 to 19 band at 20.8%. Larger firms lead, at 34.5% for 100 to 249 employees and 38.4% for 250 or more. The very smallest shops are not the laggards anymore. The five to twenty person business in the middle is.

Trying is not integrating

The Fed survey is the only one of the five that asks how deep the use goes, and the answer is sobering. Of the firms using AI, about half said they were experimenting, 44% had partially integrated it into business processes, and just 7% had fully integrated it.

Applied to all employer firms (our arithmetic): 46% times 44% is about 20.2% of firms partially integrated, and 46% times 7% is about 3.2% fully integrated. The remaining users, roughly 49% of them, are experimenting, which is about 22.5% of all firms. So out of every hundred employer firms, about 54 do not use AI, about 23 are trying it, about 20 have partly built it in, and about 3 run on it.

What firms use it for explains the pattern. The most common tasks were writing or marketing at 83% of users, individual productivity at 61%, and planning or analysis at 51%. In each of those, a person asks, reads the answer, edits it and moves on. The tool makes one person faster; the business process around that person does not change. That is what experimenting looks like from the inside. It is not a failure, just a different thing from integration.

The top challenge users reported was accuracy, at 46%. That fits. When AI output goes straight to a customer or into a record with nobody checking, a wrong answer costs something. Owners who have seen one confident wrong answer tend to keep a person reviewing every output, and a person reviewing every output by hand is exactly what keeps a tool on the experimenting rung.

The Stanford AI Index points the same way at the far end of the ladder. On AI agents, software that takes actions across several steps rather than answering a single prompt, it reports that across most business functions a majority of respondents had no agent use at all, and scaled use was in the single digits for nearly all functions. Those respondents are larger organizations, not small businesses. If scaled agent use is rare there, a small business without agents is not behind.

Where Nevada sits, and the number that does not exist

The September Census release puts Nevada at 30.3%, about seven points above the national 23.2%. Before that goes in a pitch deck, look at the standard error: 2.19 points. A rough 95% range (our arithmetic, 1.96 standard errors either way) runs from about 26.0% to 34.6%.

The prior three Nevada cycles read 21.3%, 28.3% and 15.7%. Consecutive readings that swing by 12 to 15 points are telling you about sample size as much as about businesses. Averaging the last four cycles (our arithmetic, not a Census figure) gives 23.9%, within a point of the national rate. The fair reading is that Nevada businesses use AI at about the national rate, perhaps a little above, and a single cycle cannot say more.

There is no Las Vegas number. Census publishes estimates for 25 metro areas, and Las Vegas is not one of them. Anyone quoting a Census AI adoption rate for Las Vegas is quoting something Census did not publish. The closest local read is the Nevada State Bank survey's Southern Nevada figure of 51.6%, which carries the opt in caveat above. It tells you engaged Southern Nevada respondents answered about the same as the rest of the state's respondents, not that half of Las Vegas businesses run on AI.

A three question test for your rung

The Fed's three buckets are self reported, so two owners doing the same thing can answer differently. These three questions pin it down using what actually happens in your business.

1. If the AI tool vanished tomorrow, would a customer or a record notice within a week? If only an employee would notice, because their drafts got slower, the tool is a personal productivity aid.
2. Does at least one step now run without a person carrying text in and out of a chat window? A call summary that lands in your case management system on its own counts. Copy and paste does not.
3. Does that step have a named owner and a regular accuracy check? With accuracy the top concern among AI users, a process nobody reviews is not integrated. It is unmanaged.

A no to the first question puts you on the experimenting rung. Yes to the first two is partial integration. Yes to all three, for work that touches revenue, is the fully integrated rung that about 3 in 100 employer firms report.

Two worked examples

Take a six person personal injury firm. A paralegal uses a free chat tool most days to draft first versions of demand letters and client update emails. Census would count this firm if the use fell in its two week window. JPMorgan would not, because nobody paid. The Fed would record it as experimenting. On the test, question one is a no: if the tool disappeared, clients would get the same letters a little later.

Now put a number on the next rung, using our own assumptions. Assume the firm takes 25 new inquiries a week and someone spends 10 minutes after each call typing notes into the case system. That is 250 minutes, a little over four hours a week, of retyping. Automating that one step, with each summary landing in the record and a paralegal reviewing it, answers yes to all three questions for one process. Moving up a rung means one process, not a firm wide AI strategy. We describe what that kind of build involves on our business AI solutions page.

Second example: a 120 employee dealership. It sits in the 100 to 249 employee band, where Census reads 34.5%, so its size peers are ahead of the small shops. Assume it pays for a subscription that writes vehicle descriptions and marketing emails, and the BDC uses the same tool to draft follow up messages. JPMorgan would count it and Census would count it; in the Fed's terms it would probably call itself partially integrated. On the test, it stalls at question two: every draft still passes through someone's copy and paste.

The dealership's next rung is a step that runs on its own, such as after hours service requests landing in the scheduler, with a service advisor confirming them the next morning. Assume 15 such requests a week currently wait for a callback. The value sits in the ones that would otherwise book elsewhere, and the dealership can count those from its own records before building anything. Examples of this kind of process work for local operators are on our Las Vegas business automation page.

The comparison is the point. Most measures count both the small firm and the large dealership as AI users, and neither has integrated a single customer facing process. Size changed the tool budget, not the rung.

Where this analysis does not hold

The Census figures cover employer businesses only. A solo operator with no payroll is outside that sample. JPMorgan's nonemployer figure of 15.3% is the closest read for them, and it only sees paid tools. If you work alone on free tools, none of these five numbers really describes you.

The Stanford agent findings come from larger organizations, so treat them as a ceiling check, not a small business benchmark. The Fed survey was published March 3, 2026, and the Census series has risen since then, so the split between experimenting and integrated users may have shifted. None of these sources offers a newer depth measure to replace it.

Sector matters. JPMorgan found paid adoption of 39.3% in information businesses, 30.3% in professional services and 29.5% in educational services. A law firm's peers sit in that professional services band, well above the all firm average. These sources give no comparable figure for restaurants or auto dealers, so do not assume one.

Finally, integration is not the right goal for every task. If a task happens twice a month, experimenting is the correct rung, and building a pipeline for it would cost more than it saves. The test is meant for work that repeats every day.

What to do this week

1. List every AI tool in use, free or paid. Ask each person, not just the manager, and write down the task each tool does and how often. Most owners find at least one tool nobody mentioned.
2. Run the three question test on each task. Mark each one experimenting, partial or integrated. Expect most to land on experimenting, because that is where most businesses are.
3. Pick one daily task and time it for five days. Choose the one that repeats most and fails question two, write down the current steps, and log how long it takes each time. That log is the baseline any automation, ours or anyone's, has to beat.

If you want a second set of eyes on that list, our free 15 minute audit will show which rung you are on and what the next one would take.