The story about the AI-powered solo operator promises lean, high-margin businesses run by one person. The Census Bureau publishes the numbers that let you check it, and the check is quick.
The Nonemployer Statistics count roughly 29.8 million one-person businesses in the United States, generating $1.7 trillion in receipts between them. Divide the second by the first and you get about $57,046 a year.
Three things about that number before it gets used for anything.
It is our division of two Census figures, not a statistic the Census Bureau publishes. It is a mean, and it is dragged upward by a small number of very large establishments: a few hundred, concentrated in finance and the arts, clear more than $5 million each. And the Census does not publish a median for this population, so the typical figure is not directly available. What can be said honestly is that the mean overstates the middle, and by an unknown amount.
Which leaves the forensic question. The tools of production have never been cheaper. Why has the average not moved?
Where the leverage actually went
The connection between hours worked and income earned has been weakening for a while, and the current wave severs it in a specific place: the marginal cost of producing a cognitive artifact. A page of copy, a logo, a block of code. That cost is falling toward zero.
When the cost of duplicating competent output collapses, the return on performing routine work collapses with it. What emerges is a split that borrows its shape from Peter Atwater’s 2020 K-shaped framing: one arm for people whose work is being commoditised by these tools, one for people whose existing position is being multiplied by them.
This split is an inference, not a measured finding. No dataset here separates the income of people using AI to consume systems from the income of people using it to build them. The argument is that the difference should show up, and the reasoning is laid out below so you can disagree with it.
The distinction it turns on is not capital against labour. It is consuming a system against owning one. Working harder, adding a credential, or raising an hourly rate are all responses that assume the constraint is your output. In an environment where automated workflows run continuously at near-zero marginal cost, output is not the constraint.
Ninety percent reliable is not reliable
The most common reason an AI strategy produces no money is a misreading of what a business is. Most people use these tools to make artifacts. A business is not a pile of artifacts. It is a set of loops that have to close.
Take a five-step loop: find the lead, qualify it, pitch, deliver, get paid. Suppose each step runs at 90% reliability. The loop does not close 90% of the time. It closes 0.9 to the fifth power, about 59%. (Illustrative arithmetic. The point is the exponent, not the specific rate.)
A loop that fails four times in ten is not a business. It is a research project with invoices attached.
METR’s measurements put a number on where the frontier sits. As of March 2025, frontier agents reached a task-completion time horizon of roughly one hour of expert human work at a 50% success rate, and the May 2026 update has that horizon continuing to double roughly every seven months. The doubling is real and fast. The 50% is the part that matters for anyone trying to build on it: a coin flip is not delegation.
METR also reports a large gap by domain. Reliability on software tasks runs far ahead of reliability on messy, cross-domain work that touches the physical world or several systems at once. Most real businesses live in the second category. Until the loop runs near reliably end to end, you are not operating a system. You are supervising an unreliable one.
Four tiers of leverage
There is a ladder here, and the rungs are further apart than they look.
- Manual prompting. Type, wait, copy, paste. Bounded by how fast you can type and read. Nothing compounds.
- Contextual copilots. Custom assistants and retrieval over your own documents. Your proprietary data is now in the loop, but you are still driving every step.
- Workflow automation. Triggers fire chains of actions through APIs. You move off the steps and onto the seams between them.
- Autonomous chains. Agents take a goal, decompose it, execute, check their own failures, and iterate.
Most people never leave the first rung, which means working as an unpaid assistant to their own computer. That is a job with extra steps, not an asset.
Why more output did not mean more income
There is an assumption buried in most AI advice: that producing more will earn more. Here is the counter-case, as a worked example. It is a constructed illustration rather than a real business.
A freelance writer produces four long-form articles a week at ₹6,000 each, so about ₹96,000 a month. She adopts AI tooling and gets to twenty articles a week. On paper that is a fivefold gain.
It does not arrive, because her clients and her competitors bought the same tools in the same quarter. The barrier to producing an article fell for everyone at once. Supply expanded, price per article fell, and she is now working considerably harder to hold the same ₹96,000.
The category she is competing in is the crowded one. Anthropic’s Economic Index, which measures what people actually use these tools for, shows content creation as one of the largest single categories of usage, while the categories that actually run a business, process automation, operations, sales and support, together account for a much smaller share. The crowd went where the tool was easiest to point.
Compare a solo compliance specialist in a niche manufacturing sector, holding a private corpus of documents and fifteen years of relationships. The same tools multiply that position instead of competing with it.
The general rule this points at: these tools multiply whatever advantage you already hold. Applied to a position that is already a commodity, they multiply a commodity.
What became scarce
When anyone can generate a competent artifact for nothing, the artifact stops being the scarce thing. The bottleneck moves, and it moves to three places:
- Demand. Finding people with a specific, acute, expensive problem.
- Distribution. Having permission to reach them at all.
- Trust. Being the person they will actually pay to fix it.
None of the three is produced by better tooling. The advantage goes to whoever sits at the seam: enough judgement to notice when the system has failed quietly, and enough standing to carry the relationship while it is fixed. Doubling the length of task an agent can complete does not manufacture a customer.
Artifacts or loops
The $57,000 average is not a verdict on the tools. The tools work. It is a statement about distribution and trust, and about the fact that universal availability is the same thing as no competitive advantage.
So the useful audit is not of your subscriptions. It is of your loops. Is the software on your screen helping you perform chores faster, at the same rate everyone else is now performing them? Or is it closing a loop that runs without you?
One of those is a faster way to hold a job. The other is an asset.
Stay & Analyze — Or Join Them.
