The Self-Employment Boom the Employment Report Cannot See

In July 2026, Americans filed 578,926 applications for a new Employer Identification Number. That is the highest monthly total the Census Bureau has recorded since the Business Formation Statistics series began in 2004 — higher than any month of the 2020 surge, and 8.1 percent above June.

In the same month, the Bureau of Labor Statistics counted roughly 16.49 million Americans whose main job is self-employment. A year earlier the figure was 16.48 million.

Business creation set an all-time record. The population of business owners did not move.

Both numbers are correct. Neither is a revision of the other. They describe different objects, and the distance between them has been widening for six years. What that distance measures is the arrival of an economic arrangement that no federal statistical program was built to recognize.

America is not experiencing a conventional self-employment boom. It is experiencing a business-creation-without-occupational-transition boom.

The Current Population Survey asks about a person’s main job and classifies the person accordingly. It produces two self-employment series. Unincorporated self-employment covers sole proprietors and partners working for themselves. Incorporated self-employment covers people who own a corporation and draw a paycheck from it — the CPS files them under wage and salary workers, because on paper that is what they are.

Add the two together and you get the working measure of Americans whose principal occupation is running a business. In 2024 that averaged about 16.74 million: 9.92 million unincorporated, 6.82 million incorporated. In 2025 it averaged about 16.77 million: 9.84 million unincorporated, 6.94 million incorporated. Unincorporated self-employment fell. Incorporated self-employment rose by roughly the same amount. The total moved by less than a tenth of a percent.

The last eighteen months contain real movement inside that flat total. In April 2025 the combined figure stood at about 16.90 million. In April 2026 it stood at about 16.27 million — a drop of roughly 622,000, or 3.7 percent. By July 2026 it had recovered to about 16.49 million, split 9.84 million unincorporated and 6.65 million incorporated.

Three cautions belong with those figures, and they matter more than the decimal places.

The first is sampling error. These are household-survey estimates drawn from a sample of about 60,000 households, and month-to-month swings of a few hundred thousand in a subcategory this size sit comfortably inside the noise. Response rates have made this worse: the November 2025 CPS response rate was 64.0 percent, a series low against a prior twelve-month average of 68.4 percent. The year-over-year comparisons are the ones worth trusting. On that basis, the seven months through July 2026 averaged about 16.37 million against about 16.64 million for the same months of 2025 — down roughly 269,000, or 1.6 percent.

The second is a hole in the record. Household-survey data for October 2025 were never collected, because of the lapse in appropriations, and BLS has stated they were not collected retroactively. October 2025 has no self-employment estimate, no unemployment rate, and no year-ago base for October 2026. The Bureau built 2025 annual averages from the eleven available months and says plainly that they “are not strictly comparable with annual averages for other years.” That caveat applies to the 2025 figures above.

The third is a level shift. Updated population controls, normally introduced with January data, were delayed and applied with the February 2026 estimates. Population control revisions move CPS levels without any underlying change in behavior, and they are not carried back through earlier months. A comparison that crosses that boundary — April 2025 against April 2026, for instance — carries a discontinuity that has nothing to do with how many Americans are running businesses.

With those caveats applied, the finding holds: primary self-employment is flat to modestly lower, and has been for two years.

The Census Bureau counts something else entirely. Business Formation Statistics are built from IRS Form SS-4 filings — applications for an Employer Identification Number — and then from the first appearance of payroll tax liability against that EIN. They are administrative records, not a survey, so they carry no sampling error.

Those records are moving fast. Seasonally adjusted applications totaled about 5.25 million in 2024 and about 5.64 million in 2025. Through the first seven months of 2026 they ran about 3.67 million, against about 3.12 million in the same period of 2025 — an increase of roughly 18 percent. July 2026 alone was about 23 percent above July 2025.

The formation side of the same release is where the picture separates. Census projects that about 29,959 of the businesses arising from the July 2026 application cohort will have payroll employees within four quarters. That is a real increase — roughly 18 percent above the July 2025 projection — but it is a projection built on top of an application base that grew faster still.

Stretch the comparison to a period unaffected by the pandemic and the divergence becomes unmistakable. Measured January through July, against the same months of 2019:

The conversion rate tells the same story more plainly. In the first seven months of 2019, about 7.1 percent of business applications were projected to become employer businesses within four quarters. In the first seven months of 2026, about 5.6 percent were. In July 2026 specifically, 578,926 applications yielded a projection of 29,959 employers — about one in nineteen.

One methodological change deserves explicit mention, because it cuts against the tidy version of this argument and is more interesting than the tidy version. In its February 2026 annual update, Census removed corporate applications carrying an internet-sales industry code from both the high-propensity and corporate application series. The Bureau noticed a spike in those filings, checked whether they were producing businesses with employees, and found they historically were not. That reclassification lowers the recent high-propensity counts and therefore widens the measured gap between total applications and high-propensity ones.

It is also, read correctly, direct evidence for the thesis. The federal agency that maintains the business register examined a fast-growing class of registrations that look like corporations on the form, confirmed that they seldom acquire payroll, and moved them out of the category reserved for businesses expected to hire. Something is being created at scale that is a real legal entity and is not on a path to becoming an employer.

Every apparent contradiction in this data dissolves once the units are named precisely. Six things are routinely treated as one thing, and they are six things.

• Primary self-employmentA person whose main job is working for themselves. Counted by household survey. About 16.5 million.
• Incorporated versus unincorporatedA legal distinction, not an economic one. The incorporated self-employed appear inside the wage and salary count.
• An EIN applicationA filing. It is not a business. It is a request for a tax identifier, and many are never used commercially.
• A nonemployer businessA tax-filing entity with revenue and no payroll. About 30.4 million existed in 2023, with roughly $1.8 trillion in receipts.
• A high-propensity applicationA filing whose form characteristics historically predict payroll. A prediction about a filing, not a business.
• A projected employer formationA modeled estimate that a filing will produce first payroll within four quarters. About 30,000 a month.

The Census Bureau states the crucial point about nonemployers in its own documentation: most are self-employed individuals operating unincorporated businesses “which may or may not be the owner’s principal source of income.” That single clause is the seam. On one side of it sits an occupational statistic. On the other sits a business statistic. Nothing requires them to move together, and for six years they have not.

American employment statistics were designed for an economy in which a working person was principally one of two things. An employee. Or a proprietor. The survey instrument still assumes the choice is exclusive, because for most of the twentieth century it effectively was: operating a company required premises, staff, vendors and hours that could not coexist with a full-time job.

A third arrangement now exists at scale. A person holds a wage-and-salary position and operates a portfolio of businesses around it — a consulting practice, a digital product, an ecommerce store, a media property, a specialized service company. Revenue exists. Legal entities exist. Tax filings exist. Productive capacity exists.

The Current Population Survey classifies that person as an employee, because the wage job is the main job. That is not an error. It is the definition working exactly as written.

Multiple-jobholding statistics do not close this gap either. They require the respondent to report a second job, with hours, during a specific reference week. A business that generated revenue in April and required no hours in the survey week is invisible to that question as well.

So the transformation registers in Census administrative records, which see the entity, and disappears from the household survey, which sees the person. The employment report keeps counting the person as an employee, which he is, while the business register keeps counting the company, which exists.

It is tempting to attribute the divergence to AI. The evidence does not support that claim, and the honest version of the argument is stronger.

The gap between applications and employer formations opened in 2020 and 2021, before generative AI was commercially available. Applications surged in mid-2020; the projected conversion rate fell from about 7.1 percent in 2019 to about 5.9 percent in 2021 and has not recovered. Whatever started this, it was not a language model.

What AI changed is the ceiling. The reason a one-person business historically stayed small was not ambition. It was that certain functions could not be performed by one person at acceptable quality or speed: research, software development, financial analysis, bookkeeping, marketing production, content operations, customer support, administration. Each of those required an employee, an agency or a specialized vendor, and each of those required payroll, contracts and management attention.

Those functions are now available to a single operator at a cost and latency that did not previously exist. The minimum human organization required to run a real business has fallen. That is the defensible claim, and it is sufficient. It does not require AI to have caused the divergence. It requires only that AI make the far side of the divergence habitable — that a business created without payroll can now do work that previously demanded a staff.

Once that is true, business formation can separate cleanly from two things it was historically welded to: the founder’s occupational identity, and the creation of jobs.

Read that sequence and the shape is clear. Every measure of business creation rose. Every measure of business ownership as an occupation did not. And the statistical system itself spent the period both losing a month of household data and formally conceding that a large, growing class of new business entities is not on a path to employing anyone.

An incomplete category is not an academic problem. It propagates into every system that relies on the classification.

Employment statistics. The self-employment rate is treated as a proxy for entrepreneurial vitality. It is not measuring that anymore. It is measuring how many people have made business ownership their principal occupation, which is a different and increasingly narrow question.

Entrepreneurship policy. Programs designed around a founder who leaves a job to start a company are aimed at a shrinking share of the people actually starting companies. The application data says formation is booming; the occupational data says the transition is not happening. Policy built on either number alone will misfire.

Banking and credit. Underwriting still sorts applicants into wage earner or business owner and prices accordingly. A person with W-2 income and three Schedule C or K-1 businesses fits neither template well, and the mismatch shows up as declined credit for creditworthy borrowers.

Taxation. The tax system already handles this person — withholding on the wage job, estimated payments and self-employment tax on the rest. Tax data therefore sees the activity that the labor statistics do not. That asymmetry is worth exploiting: IRS filing counts are a better read on business activity than the self-employment rate.

Benefits. This is the quiet mechanism underneath everything else. Health insurance and retirement contributions attach to the wage job, which means the businesses do not have to carry them. The threshold at which starting a company makes economic sense drops sharply when the founder’s benefits are already covered. Portable benefits proposals are usually framed as protecting gig workers. They would also remove the last structural reason for many of these operators to keep the wage job at all.

Economic policy. A country can add substantial productive capacity without adding proportionate payroll. Measures keyed to job creation will register that capacity as weakness. Measures keyed to business registration will register vapor as strength. Neither reading is right, and no single existing series resolves it.

The employment report was designed to count jobs and to sort workers into classes. It performs both tasks accurately. It was never designed to recognize a person who remains an employee while assembling a portfolio of AI-enabled businesses around himself.

That person is not unemployed. He is not conventionally self-employed. He is becoming an economic organization, one whose scale is limited by attention and judgment rather than by headcount.

Once that distinction is visible, the contradiction disappears. More businesses can exist without more primary business owners. More economic capacity can exist without proportionate payroll growth. The 578,926 applications and the flat 16.5 million are both true, and they were always describing different things.

The employment report is not wrong. Its categories are becoming incomplete.