Nearly half of small businesses use AI. Very few have changed how the work gets done.

We read what the Federal Reserve, the Census Bureau and a dozen other studies say about AI in small companies. The gap they describe is not about the technology.

A white scale model of a small business district. Eight buildings each have one small orange light. One building near the front is lit in every room, with lines of orange light joining its rooms.

We build AI agents for small and mid-size companies, so we have a habit of reading every survey on the subject. This autumn we went back through them properly: the government numbers, the bank numbers, the vendor numbers. We opened every link ourselves, dropped the claims we couldn’t trace, and marked the ones we could only read second hand.

One pattern runs through all of them. Plenty of small businesses use AI. Almost none of them have let it change how a job actually gets done.

In short

  • In the Federal Reserve’s 2026 report, 46% of small employer firms use AI, and 83% of those use it for writing or marketing.
  • Only 7% of the firms using it say it’s fully integrated into how they work.
  • The studies keep naming the same three obstacles: the wrong problem, no connection to the real work, and no trust. None of them is about the model.

How many small businesses use AI? It depends who asks

You’ll see wildly different numbers quoted, and they’re all real. They just come from different questions put to different people.

Who asked Who answered Share using AI
US Census Bureau, May 2026 All US businesses 19.8%
Federal Reserve Banks, 2026 report 6,525 firms with 1 to 499 employees 46%
US Chamber of Commerce, October 2026 Small businesses 66%
Goldman Sachs 10,000 Small Businesses, March 2026 1,256 owners in its program 76%

The Census Bureau asks whether a firm used AI in a business function, and it counts every business in the country, including the very smallest. In its May 2026 release it also noted that use among firms with fewer than 20 employees hadn’t changed significantly since December. The Chamber is a business advocacy group, and we couldn’t open its full method. Goldman surveys owners who signed up for a growth program, so they lean ambitious.

We use the Federal Reserve’s figure as the honest middle. It’s a large sample, it’s neutral, and it only counts firms with employees. Nearly half.

What they use it for

Here the surveys agree with each other. In the Federal Reserve’s 2026 report, 83% of the firms using AI use it for writing or marketing. A Census Bureau working paper from April 2026 found that 57% of AI-using firms use it in three or fewer business functions, and that the most common one, at 52%, is sales and marketing.

So picture the typical case. Somebody in the office has a chat window open. They use it to draft an email, tidy a job ad, write a product description. That’s useful. It also touches exactly one desk, and it leaves the rest of the company running the way it ran in 2019.

Very few have built it into the work

This is the number that matters. Three different surveys, asked of three different groups, landed in the same place.

  • 7%. In the Federal Reserve’s 2026 report, that’s the share of AI-using small firms that say it’s fully integrated. About half say they’re still experimenting.
  • 12%. In ServiceTitan’s AI in the Trades report from December 2025, that’s the share of more than 1,000 contractors who have embedded AI in their processes. ServiceTitan sells software to contractors, so keep that in mind.
  • 14%. In Goldman Sachs’ March 2026 survey, as reported by CPA Practice Advisor, that’s the share of owners who say AI is fully embedded in their core operations.

Seven, twelve, fourteen. Whatever the exact figure, it isn’t many. Most of the companies counted as “using AI” are using it the way they’d use a better spellchecker.

Why does it stall?

The same reasons come up again and again, and we find them reassuring. Why? Because not one of them is “the AI wasn’t smart enough.”

The wrong problem. In a 2024 report on why AI projects fail, RAND drew on interviews with 65 practitioners and academics. The first root cause they named was that people misunderstand, or fail to communicate, what problem actually needs solving. Those interviews predate today’s agents, but nothing we’ve seen since contradicts it. “We should be doing something with AI” isn’t a problem. “Vendor invoices don’t get checked against the order until month end” is.

No connection to the real work. In that ServiceTitan survey, the two top barriers were lack of training and integration complexity, at 44% each. A chat window can’t see your orders, your price list or your inbox. Until something connects the AI to where the work lives, a person has to copy things in and out by hand, and the time saved goes straight back out.

No trust. In an April 2026 survey of 942 owners by Bluevine and Centiment, 78% said they don’t fully trust AI to handle even low-level tasks without oversight. Bluevine is a financial technology company, and this was a three day online poll. Still, we think those owners are right. The answer isn’t to talk them out of it. It’s to build the oversight in.

A white scale model of a desk. A small orange light rests on a stack of papers. Beside it stands one dark graphite stamp.

An agent drafts. A person approves. In every picture on this site, orange is the agent and graphite is the person.

What about the numbers from big companies?

You’ve probably seen the headline that 95% of AI pilots fail. We’d be careful with it. It comes from a 2025 report by MIT’s NANDA project about generative AI pilots at large enterprises, and what it describes is pilots with no measurable effect on profit, not failures. Its method has been disputed. We don’t lean on it.

Two others are worth knowing, with their limits stated. S&P Global Market Intelligence found that 42% of companies had abandoned most of their AI initiatives, up from 17% a year before, as reported by CIO Dive in March 2025. That’s an enterprise sample and the data is older. And in June 2025 Gartner forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. That one is a prediction, not a measurement.

The most useful finding from big companies is about who succeeds. In McKinsey’s 2026 State of AI survey, as reported by Fortune, only 37% of organizations said AI was having a meaningful effect on earnings. Among the small group getting the most from it, nearly three quarters had fundamentally redesigned how the work flows. That’s a correlation in large firms, not proof. But it matches everything above: the gain doesn’t come from adding a tool to the old job. It comes from changing the job.

Where the manual work still sits

Two numbers show how much room there is.

Ardent Partners, in its accounts payable benchmarks published in January 2026, puts the average cost of processing one invoice at $9.84 and the average time at 8.2 days, with 18.4% of invoices needing a person to sort out a problem. That covers finance departments generally, not only small firms, and Ardent’s research is sponsored by companies that sell payables software.

Invoca, which tracks phone calls for businesses, reported in July 2026 that only 52% of callers to home services companies reach a person. That’s from Invoca’s own customers, so it isn’t the whole industry. But if you run a plumbing or HVAC company, you already know roughly what an unanswered call is worth to you.

Neither of these is glamorous. Both are daily, measurable and expensive, which is exactly what makes them good places to start.

What we take from all this

We sell this work, so read our conclusions with the same care we asked you to bring to the vendor surveys. Here’s what the evidence says to us.

  1. Name one job, precisely. Not “use AI”. One desk, one task, one result you can count.
  2. Connect it to the real work. The agent has to read the actual order, invoice or call log. If a person is copying text into a chat box, you haven’t started.
  3. Redesign the job around it. Decide what the agent drafts, what it checks, and where a person steps in.
  4. Keep a person approving. Anything that spends money, sets a price or reaches a customer gets a human yes. This is how you earn the trust those 78% are withholding.
  5. Use plain rules where they’re enough. Plenty of office work needs no AI at all, only software that does the same correct thing every morning.
  6. Measure it. Days per invoice. Calls answered. Quotes sent the same day.

That list is, more or less, how we work. It’s also why we think this is within reach of a company with thirty people and no IT department: it’s a few months of engineering, not a program.

The surveys will keep reporting that more small businesses “use AI” every year. The number to watch is the small one.

Sources

Every link was opened and checked on November 15, 2026. Where we could only read a source through another publication, we say so.

  1. Federal Reserve Banks, Small Business Credit Survey: 2026 Report on Employer Firms, March 3, 2026. 6,525 firms with 1 to 499 employees, surveyed September to November 2025. https://www.fedsmallbusiness.org/2026-report-on-employer-firms
  2. US Census Bureau, Business Trends and Outlook Survey, story published May 26, 2026. The question wording changed in November 2025, so earlier rates aren’t comparable. https://census.gov/library/stories/2026/05/ai-use-businesses.html
  3. US Census Bureau, working paper CES-WP-26-25, April 2026. A research paper covering firms of all sizes, not an official statistic. https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html
  4. US Chamber of Commerce, 2026 Empowering Small Business report, October 2, 2026. https://www.uschamber.com/technology/2026-empowering-small-business-report
  5. Goldman Sachs 10,000 Small Businesses Voices survey, March 2026, 1,256 program participants, as reported by CPA Practice Advisor on March 17, 2026. https://www.cpapracticeadvisor.com/2026/03/17/small-businesses-embrace-ai-but-need-training-and-support-to-fully-harness-it/179886/
  6. ServiceTitan, AI in the Trades report, December 16, 2025. More than 1,000 contractors, surveyed by Thrive Analytics. Paid for by a software vendor. https://www.servicetitan.com/blog/2026-ai-in-the-trades-report-takeaways
  7. Bluevine and Centiment survey of 942 US owners of firms with 2 to 249 employees, April 7 to 9, 2026, published by Stacker on July 15, 2026. https://stacker.com/stories/small-business/34-small-businesses-dont-trust-ai-basic-tasks
  8. RAND, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (RRA2680-1), August 13, 2024. https://rand.org/pubs/research_reports/RRA2680-1.html
  9. MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Fortune on August 18, 2025. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
  10. S&P Global Market Intelligence survey of more than 1,000 companies, as reported by CIO Dive on March 14, 2025. https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/
  11. Gartner forecast of June 25, 2025, as reported by SD Times on June 30, 2025. https://sdtimes.com/agentic-ai/gartner-more-than-40-of-agentic-ai-projects-will-be-canceled-in-next-few-years/
  12. McKinsey, The State of AI in 2026, 1,719 respondents worldwide, as reported by Fortune on September 2, 2026. https://fortune.com/2026/09/02/companies-getting-the-most-from-ai-rethinking-how-work-gets-done-cfo/
  13. Ardent Partners, State of ePayables 2025, part nine, January 22, 2026. https://payablesplace.ardentpartners.com/2026/01/state-of-epayables-part-nine-ap-benchmarks-and-best-in-class-performance/
  14. Invoca, Home Services Lead Conversion Benchmarks Report 2026, July 2026. https://invoca.com/reports/the-invoca-home-services-lead-conversion-benchmarks-report-2026