Quick answer
Public data cannot yet say. 44 of the 45 months since December 2022 are Census-projected, not counts of businesses. The one counted result: of every 100 tax-ID applications, fewer became employer businesses in 2021-22 than in 2015-2020, about 15% lower after adjusting for a trend, and much larger without that adjustment. It says nothing about AI.
Key takeaways
- What we know: of every 100 applications for a tax ID, fewer became businesses with employees in 2021-22 than in 2015-2020. It was about 15% lower after adjusting for a trend, and much larger without that adjustment.
- What we do not know: whether that changed after 2022. 44 of the 45 months from December 2022 to August 2026 are Census-projected, not counts.
- What this does not tell you: anything about AI. The data count tax-ID applications and first payrolls, nothing else.
- The fall in the Census-projected ratio after 2022 goes with a change in the kind of applications (the high-propensity share fell from 40% to 32% in the US), not with evidence of fewer companies forming.
- California is not distinguishable from other states at this precision, and California tech cannot be measured.
Are more real companies starting, or just more business applications?
Public data cannot yet say whether employer businesses kept pace with business applications after late 2022.
- What we know: of every 100 applications for a tax ID, fewer became businesses with employees in 2021-22 than in 2015-2020.
- What we do not know: whether that changed after 2022, because 44 of 45 months since then are Census estimates, not counts.
- What this does not tell you: anything about AI. No public dataset measures AI agents or products built with AI in this series.
We measured tax-ID (EIN) applications and the employer businesses the Census links to them, in its Business Formation Statistics. Window: application months January 2015 to August 2026, file retrieved 6 October 2026.
What is a business application, and what is an employer business formation?
A business application is a request for a tax ID. A formation is a new business that then pays wages. The Census puts it this way: "A business application is an application for an EIN made predominantly for business purposes." An EIN is the nine-digit Employer Identification Number the IRS issues.
The Census's second definition: "A business formation is the birth of a new wage-paying business that originates from a business application." The series here counts those that show up within four quarters of the application month.
"BFS only consider entirely new employer business formations originating from business applications; new non-employer businesses or new employer businesses formed by existing firms are not included in the business formation measures." A solo founder who never hires is, at most, an application.
Why can't the data tell what happened after 2022?
Counted formations stop long before applications do. The formation series ends at "the most recent quarter for which the administrative data identifies employer business startup activity based on first payroll observation." For later months the Census estimates, using "an econometric model that generates estimates of the likelihood that a business application turns into an employer business." In the file we used, counted formations end with applications filed in December 2022: 1 of 45 months since then is counted, and the rest are Census-projected.
What did counted data show for 2021-22?
In counted data, of every 100 applications, fewer became employer businesses in 2021-22 than in 2015-2020: about 15% lower after adjusting for a straight-line trend. It is the one counted finding; without the adjustment the gap is much larger.
Three cautions:
- The size depends on whether you adjust for the downward trend that was already under way. Without the adjustment the gap is much larger, and the counted ratio fell every year from 2015 to 2019.
- 2021 and 2022 are the pandemic surge in applications, which pulls the ratio down on its own.
- It predates late 2022 and is not about AI.
What happened after late 2022?
On the Census-projected measure, fewer applications are projected to become employer businesses after late 2022, but that tracks a change in the kind of applications. "High-propensity" applications are the ones the Census flags as more likely to hire.
| Group | Of 100 applications, how many became employer businesses: 2015-2020, counted | Application months December 2022 to August 2026, Census-projected | Share of applications flagged as more likely to hire, before and after |
|---|---|---|---|
| US, all sectors | 9.1% | 5.7% | 40% to 32% |
| California | 13% | 8.2% | 49% to 39% |
| Information | 7.1% | 4.5% | 29% to 19% |
| Professional, scientific and technical services | 9.7% | 6.8% | 36% to 26% |
California note: where both exist (2015-01 to 2020-02), projections ran below counted figures for California by 12.1%, so part of its drop is that mismatch.
What to take from it: the drop goes with the shrinking share of applications flagged as more likely to hire, and is not evidence of fewer companies forming. Also, where both exist (2015-01 to 2020-02), projections ran below counted figures by 15.6% for the US. Part of any gap is that mismatch.
Does the answer depend on your assumptions?
Yes. For the US, the answer goes from small to large depending on one assumption. US total, 2015-2020 counted versus after 2022 Census-projected:
| If you assume… | The Census-projected ratio after 2022 is… |
|---|---|
| the earlier downward trend would have continued | a Census-projected ratio about 9% lower, but the range includes no change |
| the trend would not have continued | a much larger fall in the Census-projected ratio |
| the trend would have continued, and Census projections are used in both periods | exploratory, defined after seeing the results: a Census-projected ratio slightly lower, not distinguishable from no change |
| the trend would not have continued, and Census projections are used in both periods | exploratory, defined after seeing the results: a much larger fall in the Census-projected ratio |
The first row is the main estimate for the US (counted before the cut, Census-projected after); from here on, "the main estimate" means that. The size of the fall depends on that assumption, and the data cannot say which assumption is right. The fall goes with the changed mix of applications, with projections running below counted figures, and with a ratio that was already falling every year from 2015.
Is California different?
California is not distinguishable from other states at this precision. Its main estimate is about 15% lower after adjusting for a trend. Read it with three checks. First, the Census projection ran 12.1% below counted figures before 2020. Second, the exploratory projected-in-both result (defined after seeing the results, outside the five pre-set tests) is a smaller fall. Third, with no trend adjustment the fall is much larger.
The direct test is cleaner: California against the other 50 states and DC shows no difference detected at this precision. Without a trend adjustment, 51 of 51 geographies have a lower average ratio after 2022, and California ranks 27 of 51 (1 is the largest fall). These are window means of the ratio, counted before the cut and Census-projected after, and the ratio was already falling every year from 2015. On the projected-in-both-windows basis California ranks 16 of 51. This is not a finding that formations fell everywhere.
Do tech sectors differ?
No difference detected at this precision, and California tech cannot be measured. The Census publishes sector detail only nationally. Information and Professional, scientific and technical services show no difference from the other 17 sectors at this precision. Two sectors count as tech here: Information, and Professional, scientific and technical services.
How we measured this, and where our starting belief fits
We started from an assumption, and the data could not settle it. We assumed AI makes products easier to build but not companies. We turned that into a question before looking at results: since late 2022, did tax-ID applications rise without a matching rise in employer businesses? We had already seen some of the 2026 figures before we fixed our method, so treat these as checks, not blind tests. The data cannot yet say whether employer businesses kept pace, so the assumption is neither supported nor contradicted.
We compared seasonally adjusted monthly ratios across windows. A "95% CI" is the range the true figure very likely falls in.
Technical detail (skippable): the US main estimate is Delta = -0.09 (95% CI -0.23 to 0.05, p = 0.559 after correcting for five tests, n = 107), not distinguishable from no change; with no trend adjustment Delta = -0.47; exploratory projected-in-both (defined after seeing the results) Delta = -0.04 (95% CI -0.12 to 0.03), and -0.30 with no trend. California's main estimate is Delta = -0.17 (95% CI -0.23 to -0.10, p = 0.002); exploratory projected-in-both Delta = -0.08 (95% CI -0.13 to -0.02), and -0.32 for that exploratory version with no trend; the main estimate with no trend is Delta = -0.45. The counted 2021-22 shift is -0.16 (95% CI -0.23 to -0.08, unadjusted p = 0.001, n = 86). California against other states is difference = -0.00 (95% CI -0.04 to 0.03); tech against other sectors is difference = -0.02 (95% CI -0.15 to 0.12). Months in the same year resemble each other, so the 107 months carry the information of effective n = 11; two standard checks flag the model: the trend and the before/after marker overlap heavily, and the straight-line fit is poor.
What is not covered, and what are the limits?
The data cannot show post-2022 formations as counted businesses, or why anything changed.
- The Census has not published error bands for its projections.
- Not in the sample: months before 2015, non-employer businesses, businesses formed without an EIN, Puerto Rico, and employer businesses started by existing firms.
- The months 2020-03 to 2022-11 are left out of the post-2022 comparison, so the trend is carried across a 33-month gap.
- No California-by-sector series exists, so no California tech result.
How can I check these figures?
Open the Census file and its definitions. The data are in the monthly file. The definitions, FAQs and about the data pages explain each series. Series BA_BA is applications, BF_BF4Q counted formations, BF_PBF4Q projected. Survey data: AI use by business size.
What can I do with this?
Use three rules. An application count is not a count of businesses that hire. When you see a rise in new businesses, ask whether it counts applications or first payrolls. Before you repeat any post-2022 formation figure, check the Census end-point: the last counted month here is applications filed in December 2022, and everything after is Census-projected. For what AI tools do for a small business today, see AI agents for small businesses.
Analysis by Deivy Hernandez, from public data files.
FAQ
Is business formation up since ChatGPT?
The data cannot say. The last month with counted employer businesses is applications filed in December 2022, so 1 of 45 months since then is counted and the rest are Census-projected. Applications for tax IDs were roughly 18% higher in December 2022 to August 2026 than in 2015-2020, after adjusting for a trend; whether employer businesses kept pace is not yet known.
Does a rise in business applications mean more businesses are opening?
Not by itself. An application is a request for a tax ID, filed for a side project, a one-person business or a business that never hires. The Census tracks separately how many become employer businesses.
When will counted data for 2023 to 2026 exist?
The Census page does not give a date. It says the end-point of the formation series is set by the most recent quarter for which tax records show first payroll. Check the Census data page for that end-point as it moves.
Is the Census projection a forecast?
No. It is a modelled estimate built from the details on each tax-ID application. When the mix of applications changes, the estimate changes with it. It is not a count.
Does this tell me whether AI helps people start companies?
No. The series count tax-ID applications and first-payroll businesses, not AI use or products built, so no result here can be tied to AI.