TL;DR: Selling AI automations to small businesses is being pitched everywhere as the obvious solo business of 2026, usually with no numbers attached. Here are the numbers. The US Census Bureau’s Business Trends and Outlook Survey put national business AI use at 19.8% as of 3 May 2026, having hovered between 17% and 20% since December 2025, with adoption at 37% among firms with 250 or more employees but under 20% among firms with fewer than 20. Federal Reserve analysis published in April 2026 puts work-related generative AI use at 41% of the workforce over the same period. Firms are at roughly 19%; workers are at roughly 41%. That gap is not a measurement error, and it is not a reason for despair about the market. It is the market: individual employees are already using these tools informally while the firm itself has nothing integrated. Applying the Census non-adoption rate to the 6,274,916 US employer firms counted by the Small Business Administration gives roughly 5.0 million firms with no AI in their business functions. Below: the original addressable-firm arithmetic, a transparent three-scenario unit-economics model showing why the delivery mix matters more than the rate, and a durability test for the one risk that actually kills this business.
Every few years a business model arrives with enough surface plausibility that thousands of people start it in the same quarter. The one-person automation agency, meaning a solo operator who builds AI and workflow automations for small and mid-sized businesses, is that model right now. The pitch is clean: the tools got good, the small businesses have not adopted them, you sit in the middle.
The pitch is also, in outline, correct. What is missing from nearly every version of it is the arithmetic that determines whether the middle is a good place to sit. So let us build it.
What the official data actually says about the market
Two US government sources measure this, and they disagree in a way that turns out to be the most useful fact available.
The Census Bureau’s Business Trends and Outlook Survey, a firm-level survey, reported national AI use of 19.8% as of 3 May 2026, having ranged between 17% and 20% since mid-December 2025, with 20% to 23% of businesses expecting to use it within six months.
Table 1. Verified business AI adoption by firm size and sector (US Census Bureau, Business Trends and Outlook Survey, 14 December 2025 to 3 May 2026)
| Segment | Current AI use | Expected use in next 6 months |
|---|---|---|
| National, all firms | 19.8% | 20% to 23% |
| Firms with 250 or more employees | 37% | not separately stated |
| Firms with 100 to 249 employees | 32% | not separately stated |
| Firms with fewer than 20 employees | under 20%, no significant change over the period | modest, under 5 points |
| Information sector | 39.7% | 42% |
| Finance and insurance | 33.9% | 39% |
| Retail trade | about 14% | about 17% |
Now the second source. A Federal Reserve FEDS Note published on 3 April 2026 compared three separate measurement approaches and found they produce very different answers.
Table 2. Verified AI adoption by measurement approach (Federal Reserve Board, FEDS Notes, April 2026)
| Survey | What it weights | Adoption reported | Reference period |
|---|---|---|---|
| Business Trends and Outlook Survey | Firms | 18% of firms | December 2025 |
| Real-Time Population Survey | Workers | 41% work-related generative AI use | November 2025 |
| Survey of Business Uncertainty | Employment | 78% of the labour force at adopting firms; 54% for large language models specifically | November 2025 |
The Federal Reserve’s own explanation for the divergence matters. The firm-weighted survey mirrors the actual population of businesses, which is overwhelmingly small. The employment-weighted survey oversamples large employers. Question framing differs, and so does who answers: executives in one case, general respondents in the other.
Read the two tables together and the commercial insight falls out. Roughly 41% of workers are using generative AI for work while roughly 19% of firms report AI in their business functions. The ratio is about 2.07 to 1. Millions of people are using these tools privately, at their own initiative, inside firms that have integrated nothing.
That is the product. You are not selling AI to businesses that have never heard of it. You are selling the conversion of unmanaged individual usage into an owned, documented, repeatable business process. Those are very different sales, and the second one is far easier.
The addressable-firm arithmetic
The Small Business Administration’s Office of Advocacy, drawing on Census SUSB and Nonemployer Statistics, counts 34,752,434 small businesses in the US: 28,477,518 nonemployer firms and 6,274,916 employer firms, alongside 19,688 large businesses. Small businesses account for 99.9% of all firms and 45.9% of private sector employees, which is 59.0 million workers.
Employer firms are the realistic buyer for an automation retainer, because a nonemployer firm is one person and rarely has a process worth automating for someone else’s fee. So apply the Census adoption band to that base.
Table 3. Addressable non-adopting employer firms, United States (CEOtudent editorial calculation from SBA Office of Advocacy firm counts and US Census BTOS adoption rates)
| Basis | Adoption rate applied | Non-adopting employer firms |
|---|---|---|
| Upper end of the BTOS band | 20% | 5,019,933 |
| BTOS reading at 3 May 2026 | 19.8% | 5,032,483 |
| Lower end of the BTOS band | 17% | 5,208,180 |
Roughly five million US employer firms have no AI in their business functions. The number is large enough that market size is not your constraint, which is the important conclusion. If you cannot make this business work, it will not be because there were not enough prospects. It will be because of the delivery economics, which is what almost nobody models.
Two refinements worth noting. First, sector concentration runs opposite to intuition: the Information sector is at 39.7% adoption and retail trade at about 14%, so the least-served buyers are the least technical ones, which raises both the education cost per sale and the value of the work. Second, the Census found adoption rose among firms with at least 20 employees during the period but did not change significantly among firms below that, and smaller firms expect gains of under five points. The bottom of the market is not adopting on its own. That is durable demand and slow demand at the same time.
The unit economics almost nobody models
Here is where most accounts of this business stop and where the actual answer lives.
A solo operator has one hard constraint that no amount of tooling removes: hours. The question is not what you charge. It is how many separate sales you must make per year to hit your number, because selling is the activity that does not compound and cannot be automated away by the tools you are selling.
The model below is illustrative, not survey data. We are not aware of any authoritative source for typical AI-automation project fees, and we are not going to invent one. Treat the fee and hour figures as placeholders and substitute your own; what matters is the structure the arithmetic reveals, which holds at any rate level.
Table 4. Three delivery mixes at the same revenue target (CEOtudent editorial model; illustrative assumptions, not market data)
| Assumption | Value |
|---|---|
| Annual revenue target | 120,000 currency units |
| Project fee | 6,000 per project |
| Retainer fee | 2,000 per month |
| Delivery hours per project | 40 |
| Delivery hours per retainer client per month | 8 |
| Scenario | Retainer clients | Projects per year | Delivery hours per year | Effective revenue per delivery hour | New sales required per year |
|---|---|---|---|---|---|
| A. Project only | 0 | 20 | 800 | 150 | 20 |
| B. Retainer heavy | 5 | 0 | 480 | 250 | 5 |
| C. Mixed | 3 | 8 | 608 | 197 | 11 |
Same revenue in all three rows. The retainer-heavy mix delivers it in 40% of the hours of the project-only mix and requires one quarter of the annual sales activity. Nothing about the hourly rate changed; only the shape of the engagement did.
This is the single most consequential finding in the model, and it inverts how most people start. New operators sell projects because projects are easier to close: a defined deliverable, a fixed price, no ongoing commitment from the buyer. Then they discover that a project-only business has to re-fill its entire pipeline every year and that the 800 delivery hours plus the sales effort for 20 closes does not leave room for anything else.
The correction is not to charge more. It is to make the retainer the default product and the project the on-ramp to it.
There is a second-order point hiding in the hours column. In scenario B you have 480 delivery hours against a working year that has far more available. The surplus is not idle time, it is the capacity that lets you build reusable automation components, which is the only mechanism by which a solo operator’s margins improve over time. A project-only business consumes the surplus and stays flat forever.
The durability problem, and a test for it
Now the risk that actually kills this business, and the one most enthusiastic accounts ignore.
Every automation you build sits somewhere on a spectrum between “thin wrapper around a capability the model will have natively next year” and “durable integration of a specific firm’s specific mess.” The first kind gets absorbed. We looked at the general version of this problem in whether you can still build a profitable AI wrapper product in 2026, and the services version follows the same logic with one important difference: a services business can migrate when the ground shifts, because the client relationship, not the artifact, is the asset.
That difference only protects you if you chose durable work in the first place. Here is the test.
Table 5. Automation durability test (CEOtudent editorial framework)
| Question about the engagement | Absorbed by the next model release | Durable |
|---|---|---|
| What is the hard part? | Prompting, phrasing, output formatting | Connecting systems that were never designed to talk to each other |
| Where does the knowledge live? | In the tool you configured | In the client’s specific processes, exceptions and edge cases |
| Who could rebuild it? | Any competent user with the same tool, in an afternoon | Someone who spent weeks learning how this business actually runs |
| What happens when the model gets better? | Your automation becomes a checkbox in the product | Your automation gets cheaper to run and you keep the margin |
| What is the client paying for? | Access to a capability | Accountability for an outcome |
| How is success measured? | The automation runs | A business metric moved and someone owns it |
Read the right column carefully. Every durable answer has the same property: the value is in knowledge of the specific client, not in knowledge of the tool. Tool knowledge depreciates with every model release. Client knowledge appreciates with every month of the relationship.
This is also why the retainer is not merely better economics. It is the structure that lets client knowledge accumulate at all. A project ends before you have learned enough about the business to be hard to replace.
What this means practically
Five things follow from the numbers above.
Sell the conversion, not the technology. The 41% worker adoption against 19% firm adoption means your prospect’s staff are already using these tools. Your opening is not “you should try AI.” It is “your team is already using this without oversight, without documentation, and without anyone owning the result.” That is a problem the owner recognises immediately.
Target below 20 employees and outside the technical sectors. Census data shows firms under 20 employees have not moved and expect gains under five points, while Information sits at 39.7% adoption. The underserved buyers are in retail, trades, professional services and similar. They will need more education per sale and they will also have less competition for their attention.
Make the retainer the product from day one. Scenario B needs five clients where scenario A needs twenty closes. If you must open with a project, price and scope it as a diagnostic that concludes with a retainer proposal.
Choose work that fails the absorption test. Before quoting, run Table 5. If most answers land in the left column, either reprice it as a cheap on-ramp or decline it. Building a business on work that is one release from becoming a product feature is the specific failure mode here.
Build components with the surplus hours. The gap between 480 delivery hours and a full working year is the only leverage a solo operator has. Spend it on reusable pieces, not on more sales calls.
For the groundwork underneath all of this: the 80/20 of knowledge work automation covers how to identify which tasks are worth automating in the first place, which is the diagnostic skill this business is built on. Agent literacy covers the concepts you need before delegating real work to agents, and what agentic browsers and computer-use AI actually do is an honest assessment of the tools most of these automations are built on. On the wider question of what happens to solo revenue models as AI reshapes distribution, affiliate income after AI search is a useful comparison case.
FAQ
Is the market really five million businesses?
Roughly five million US employer firms currently report no AI in their business functions, calculated from the SBA Office of Advocacy count of 6,274,916 employer firms and the Census BTOS adoption band of 17% to 20%. That is the count of firms without AI, not the count of firms that will buy an automation retainer. Treat it as evidence that prospect volume is not your binding constraint, not as a revenue projection.
Why do the adoption numbers vary so much between sources?
Because they measure different things, and the Federal Reserve’s April 2026 note is explicit about why. Firm-weighted surveys mirror the real business population, which is mostly very small firms. Employment-weighted surveys oversample large employers, which is how you get 78% of the labour force at adopting firms alongside 18% of firms. Question framing and who answers also differ. None of the numbers is wrong; they answer different questions.
Should I charge by the hour?
The model in Table 4 suggests the question is misframed. Scenario B produces an effective 250 per delivery hour against scenario A’s 150 without any change in what you charge, purely from the engagement structure. Hourly pricing also caps you at exactly the constraint you are trying to escape, since your hours do not scale and your reusable components do.
What if the client’s staff can just do this themselves?
Often they can, and if the work fails the Table 5 durability test they eventually will. That is the correct filter rather than an objection to overcome. The work worth selling is the work that requires knowing how a specific business runs, which the staff know and cannot systematise, and which no tool ships with.
Is it too late to start in 2026?
The Census data argues against that reading. Firms under 20 employees showed no significant change in adoption between December 2025 and May 2026 and expect gains under five points ahead. The bottom of the market is not adopting on its own timetable, which means the demand is durable rather than urgent. Slow-moving demand is harder to sell into and much harder for competitors to exhaust.
How many clients can one person actually serve?
The constraint in the model is delivery hours plus sales activity, not client count in isolation. Scenario B’s five retainer clients consume 480 delivery hours; scenario A’s twenty projects consume 800 plus four times the sales effort. Whatever your real fees, run your own version of that table before deciding how many clients you can carry, because the answer changes completely with the mix.
Sources
US Census Bureau, Business Trends and Outlook Survey, business AI use data, May 2026.
Board of Governors of the Federal Reserve System, Monitoring AI Adoption in the U.S. Economy, FEDS Notes, April 2026.
US Small Business Administration Office of Advocacy, Frequently Asked Questions About Small Business, 2024, drawing on Census Statistics of U.S. Businesses and Nonemployer Statistics.
US Census Bureau, Statistics of U.S. Businesses, annual firm and establishment data by enterprise size.
World Economic Forum, Future of Jobs Report 2025, January 2025.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
