A Business Application Is a Signal, Not an Employer
This series studies a specific conversion: what happens between a person's business application and the appearance of a business that pays at least one employee. The opening question is why the application—often treated as a startup in headlines—cannot answer that question by itself. Keeping those events separate gives the reader a defensible way to read early-stage data, judge what an application signal is worth, and see where claims about jobs and survival begin.
The distinction sounds administrative until the numbers are placed side by side. A person can decide to try self-employment, file for an Employer Identification Number, predict that they will hire, and still never become an employer. Another applicant can report no expected employees and later hire. A business that crosses the employer threshold can then remain tiny, grow, or disappear. These are not interchangeable outcomes. They are stages in a path from intention to paid employment, and each stage answers a different question.
The event people casually call “starting a business”
There are at least three relevant events. The first is entrepreneurial activity in a broad population measure: someone who did not own a business in the first month of a survey and begins one in the next month while working at least 15 hours per week. The Kauffman Foundation’s rate of new entrepreneurs includes incorporated and unincorporated businesses, employers and nonemployers, and businesses of every size. It is useful for asking how often adults move into business activity. It is not a count of new employers.
The second event is an application. In the Census study, the relevant applications are filings that state an intent to start a new business. The form records such details as the expected maximum number of employees in the next 12 months. That response is an early statement of belief: the applicant’s view of demand, profitability, ability, competition, and uncertainty. It is not a promise to hire, and it is not a record that hiring has already occurred. The filing may be needed for banking, licensing, payroll, or other administrative reasons, but even when it clearly expresses a new-business plan, the plan remains a plan.
The third event is employer entry: a linked business record shows positive paid employment. The Census researchers connect roughly 17 million new-business applications filed between April 2017 and December 2021 with later observations in the Longitudinal Business Database. For their employment-generation question, an application becomes an entry when the linked business records positive employment within a matched one-year window. An application that becomes a nonemployer business is grouped with no employment generation in that outcome, even though a nonemployer business may provide meaningful income to its owner.
That last qualification matters. “Not yet an employer” does not mean “not real,” “not useful,” or “not economically meaningful.” It means that the particular downstream outcome being measured—paid employment—has not been observed. The distinction is analytical, not a judgment about the worth of self-employment.
The same word hides multiple denominators
The Kauffman report makes the broader measurement problem visible by separating four indicators. The rate of new entrepreneurs measures movement into business ownership. The opportunity share asks whether new entrepreneurs came from a labor-market position other than unemployment and active job search; it distinguishes opportunity from necessity, but does not measure business quality. Startup early job creation measures jobs created by startup businesses in their first year per 1,000 people. Startup early survival measures the percentage of new employer establishments still active one year later.
Those indicators are complementary, not successive names for the same event. A population may have a high rate of new entrepreneurs because many people begin small or nonemployer activities, while startup job creation remains weak. A cohort may create jobs but have lower one-year survival. A higher opportunity share may describe motivation without telling us whether a business will hire or persist. The Kauffman report therefore warns against treating any one indicator as a complete picture.
The 2025 figures show why that warning is practical. The rate of new entrepreneurs was 0.36 percent—about 360 out of every 100,000 adults in a given month—and remained above its 2019 pre-pandemic level of 0.31 percent. Yet the opportunity share was 83.3 percent, below the 86.9 percent recorded in 2019, after falling to 69.8 percent in 2020. The broad movement into business activity was elevated, but its composition had not fully returned to its pre-pandemic pattern. Nor did the downstream measures tell a uniformly upbeat story: startup early job creation was 5.3 jobs per 1,000 people, while one-year survival was 77.9 percent, below its 2019 level of 79.4 percent.
The numbers should not be combined into a homemade conversion rate. The activity and opportunity measures come from monthly Current Population Survey transitions; job creation and survival use startup and establishment data with different horizons and denominators. The Census application study uses a narrower population and links expectations to employer outcomes. Their value lies in the contrast between lenses. When a headline says that entrepreneurship increased, the next question should be: increased at which stage?
The conversion has two tests
Employer entry itself contains a distinction that is easy to miss. The first test is whether the business becomes an employer at all. This is the extensive margin: the boundary between zero and positive paid employment. The second test is how many employees the business reaches once it crosses that boundary. This is the intensive margin: the scale of employment conditional on entry.
An application can therefore be wrong in two different ways. It can predict employer entry that never occurs. Or it can correctly anticipate entry while misjudging the number of employees. The reverse errors also matter: a business can hire despite reporting zero expected employment, or an entrant can grow beyond its initial plan. A single “startup success” label conceals these paths.
The distinction also explains why more activity can coexist with mixed employer outcomes. During the first year of the pandemic, the Kauffman rate of new entrepreneurs rose from 0.31 percent in 2019 to 0.38 percent in 2020, while the opportunity share plunged. People were moving into business activity, but the economic circumstances behind that movement had changed. In the Census data, applications whose twelve-month realization windows overlapped the pandemic showed a widening gap between expected and realized employment. A rise in intentions or transitions could not by itself reveal how many businesses would hire, at what scale, or for how long.
The practical implication is a sequence of questions rather than a single verdict:
- Did someone express an intention to start a business or move into business activity?
- Did the application become an employer within the relevant horizon?
- If it did, how did realized employment compare with the expected scale?
- How many jobs did the startup cohort create, and how many establishments remained active after a year?
The sequence is not a claim that every business must grow or hire to be successful. It is a way to avoid claiming paid-employment effects from evidence that only establishes intention or self-employment.
A forecast is not a verdict
The hiring number on an application is most usefully understood as a forecast generated before the business has faced all of the conditions that will shape its outcome. It can contain information. Someone who expects to hire more may know something about a customer base, a contract, a skill, or a local market that an outside observer cannot see. At the same time, the expectation can be systematically too high, especially when the applicant’s belief is formed under uncertainty or when only people with sufficiently positive expectations choose to apply.
That creates an important middle position between dismissal and credulity. An application forecast is not meaningless paperwork simply because some applicants fail to enter. Nor is it a reliable headcount simply because expectations are correlated with later employment. A signal can help rank likely outcomes while remaining poorly calibrated in absolute terms. It can also be associated with later performance without proving that optimism caused that performance.
This is why the language of the evidence matters. A linked administrative record can show whether employer entry occurred within a defined window. A regression can show that expectations predict realized employment. A model can explain how private information, noisy beliefs, and selection might produce both patterns. None of those steps alone establishes a universal conversion rule or identifies the precise psychological motive behind every forecast.
The central mental model is therefore a pipeline with filters, not a straight line from paperwork to jobs:
The labels in this map are deliberately precise. The application and expected-entry stages describe beliefs or intentions. Realized employer entry requires observed paid employment. Job creation asks about the employment produced by a startup cohort, while survival asks whether new employer establishments remain active after one year. Even the broad activity measure is observed through a different survey transition. The arrows show a useful order for reasoning, not a guarantee that every case follows the same path.
The first discipline, then, is simple: never let the earliest observable signal stand in for the later outcome it is meant to predict. A business application tells us that someone has crossed an administrative and cognitive threshold—enough intention to file and enough expectation to report a plan. It does not yet tell us that an employer exists. To learn that, we must follow the plan to its first conversion: the point at which paid employment appears, or fails to appear, in the linked record.