Small Business AI Agents: What's Real

Short answer: the AI agents worth having in a small business do a narrow, boring job, and a person approves the output before it goes anywhere. The ones that make the news are the opposite: broad, unsupervised, and pointed at customers.

I build these for a living, so I have a stake. I’ll flag where the evidence is thin, and it’s thinner than the vendor blogs make it sound.

Who’s actually using this

Fewer people than you’d guess. The Census Bureau’s business survey put AI use at 19.8% of US firms as of May 3, 2026, and it hovered between 17% and 20% from December 2025 through May 2026. Firms with at least 250 employees were at 37%. For firms with four or fewer employees, less than 20% reported using AI. Retail sat around 14% (Census Bureau, checked October 29, 2026).

The Federal Reserve’s own note says adoption rises with firm size, but that adoption among the smallest firms is stronger than their size alone would predict. On how much people use it, a different survey in the same note, of business executives, found a plurality of AI users (35%) spending up to an hour a week (Federal Reserve FEDS Note, April 2026, checked October 29, 2026).

One caution on those numbers. The question asks whether a firm used AI in the past two weeks. Census originally framed it around AI used “in producing goods or services” and changed the wording in November 2025 to “in any business function,” so you can’t line it up against older figures, and lighter uses can now count.

What I take from it: small firms aren’t behind on some race. Most haven’t found a job worth handing over yet. That’s information, and it’s a reasonable place to be.

What holds up, and why

Here’s the part I can’t dress up. I looked for neutral evidence, meaning not from a company selling the thing, that the usual small-business jobs work: drafting review replies, entering invoices, sorting email, chasing quote follow-ups, pulling facts out of incoming paperwork. I didn’t find any. What exists is vendor case studies, and I wouldn’t lean on those.

The closest solid evidence isn’t from small firms. A study of 5,179 customer-support agents found that a generative-AI conversational assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled ones (NBER, Brynjolfsson et al., checked October 29, 2026). That’s a large support operation, nothing like a five-person shop, and the abstract doesn’t say who approved what the AI wrote. MIT’s widely quoted “95% of pilots fail” report, as Fortune reports it, found the biggest return in back-office automation, and that tools bought from specialized vendors and partnerships succeed about 67% of the time while internal builds succeed one-third as often. The report is about enterprise pilots, so it may not carry over to a small shop (Fortune, August 2025, checked October 29, 2026). Note the in-house finding cuts against people like me who build custom.

So I won’t tell you the boring jobs are proven. I’ll tell you why I think they’re a safe place to start. Three properties:

A person approves the output. The AI drafts, you or your staff click yes.
Mistakes are cheap and visible. A bad draft reply gets caught and fixed before anyone sees it. A mis-read invoice gets flagged next to the PDF it came from.
The job is narrow. One task, with a clear right answer you can check.

Take invoice intake. Vendor invoices arrive as PDFs and emails. The agent reads each one, pulls the vendor, date, amount and line items, and lines them up for entry in your accounting software, flagging anything that doesn’t add up, like a total that doesn’t match its lines or a vendor it hasn’t seen. Your bookkeeper reviews the flagged ones and spot-checks the rest. Nothing posts to the books without a person. If it misreads a number, the failure shows up on a review screen instead of in your tax filing.

That’s the whole design principle. I’m not claiming it makes you faster by some percentage; I haven’t got a neutral number to give you. I’m saying the downside is bounded, and you can measure the upside yourself in a few weeks.

What goes wrong

The failures that make headlines share a shape: the AI spoke directly to customers or the public, with no person in between.

Cursor’s AI support bot told users that being logged out across devices was “expected behavior” under a new policy. No such policy existed, there was no human behind that email, and users reported cancelling. The company’s cofounder acknowledged an “incorrect response from a front-line AI support bot” (Fortune, checked October 29, 2026). New York City’s MyCity chatbot told business owners they could take a cut of workers’ tips, refuse cash, and that landlords could discriminate against Section 8 voucher holders. In January 2026 the new mayor put its cost at around half a million dollars and singled it out as a cost to cut, and by February 4 the city had taken the bot down (The Markup, January 30, 2026, updated February 4, checked October 29, 2026). And a company is responsible for what its AI tells customers; I wrote about the Air Canada case in what an AI agent costs.

The second kind of failure is security, and it’s subtler. Two entries on OWASP’s list of top risks for AI applications matter for a small business:

Prompt injection. Text the AI reads, like an email, a web page or a PDF, can carry instructions that the AI follows as if you’d given them (OWASP LLM01, checked October 29, 2026). Their example is a support chatbot told to ignore its rules, grab private data and send emails.
Excessive agency. OWASP’s term for damaging actions taken in response to unexpected, ambiguous or manipulated output, usually because the agent was given too many tools, too many permissions or too much freedom (OWASP LLM06, checked October 29, 2026).

In plain words: an agent that reads outside email and can also send email or touch your books can be talked into things. The fix is boring. Give it the smallest set of permissions that does the job, and keep a person between its output and anything that can’t be undone.

Should you? Five questions

Can you describe the job in two sentences? If not, it’s not ready.
Does someone do this by hand today, repeatedly? Then you have something to compare against. If nobody does it, an agent won’t invent the need.
What does a wrong answer cost, and who sees it first? If the answer is “a customer,” add an approval step or leave it alone.
What can it touch? List the accounts and actions. Cut the list.
How will you know it’s working? Pick the number before you start: errors caught, minutes per invoice, replies approved without edits.

If you get stuck on question 3 or 4, that’s usually the answer.

Where I come in

The first step here is a $950 AI Opportunity Assessment: I look at two or three of your workflows, rank them, and hand you a working prototype of one on sample data plus a written plan. If something’s worth running, a $4,500 pilot puts it beside the person who does the job today for two to four weeks, and you get a plain go or no-go report. Some assessments end in “don’t build this,” and I’d rather you hear that from me. The details are on the AI automation page, and the cost side is in what an AI agent really costs.

If you’re also wondering about the site those customers land on, that’s a different conversation, and the Columbus web developer page covers it.

Email [email protected] with the job you’d like off your plate.