Reading the entrepreneurship boom like a scientist

Payment processor Stripe reports a striking pattern: businesses that signed up after 2023 are reaching a million dollars in cumulative revenue roughly three times faster than the 2019 cohort, solo founders made up 63% of companies formed through its Atlas service in the most recent quarter reported, an all-time high, and the number of solopreneurs earning seven figures reportedly more than doubled between 2023 and 2025. Economist Liya Palagashvili, of the Mercatus Center at George Mason University, adds labour-market evidence pointing the same way: solo business applications in high-AI-adoption sectors rose nearly 27% since early 2024, while applications in sectors AI has barely touched stayed roughly flat. Read together, the story is coherent and genuinely interesting: AI is lowering the number of people it takes to build a real business.

It is also a story built on business-registration data, and that specific category of evidence has misled before, in the same direction, for reasons that had nothing to do with any underlying technology.

The baseline has moved before

Stripe's own reporting flags two earlier episodes. Increased IRS registration pressure on gig workers produced a steady rise in non-employer business registrations that had nothing to do with a change in how many people were actually starting businesses; it reflected more of the existing activity being formally registered. Separately, eligibility rules attached to 2020 Paycheck Protection Program loans drove a registration spike large enough to roughly double the baseline, and that elevated baseline never fully unwound. Anyone reading a registration count today is reading a series with at least two known step-changes in it that were caused by policy, not by entrepreneurial appetite.

None of this means the current data is wrong. It means a registration count answers a narrower question than the headline built from it: it tells you how many entities were formally registered, not directly how many businesses meaningfully started, scaled, or were run by one motivated person rather than a team that simply had not hired yet. Stripe itself is explicit about the second gap, noting that registration data cannot cleanly separate solopreneurs from firms that intend to hire as soon as revenue allows.

Grading the claims, not dismissing them

This is the same discipline worth applying to any claim about AI’s effect on an organisation, and it does not require rejecting the entrepreneurship data. It requires sorting it. A registration count, taken alone, is a weaker form of evidence than a registration count corroborated by an independent series measuring a different thing the same story would predict. Here, several independent measures do move together: Palagashvili's federal self-employment data (roughly 20% growth from 2022 to 2025 in AI-exposed occupations, flat elsewhere), Stripe's revenue-cohort data (which is not a registration count and is not exposed to the same policy artefacts), and international registration growth across several jurisdictions with different tax and loan regimes, which makes a single national policy explanation less likely to be doing all the work. Convergence across independently-sourced measures, each with different failure modes, is a stronger form of evidence than any one of them alone, precisely because a policy artefact in one series is unlikely to also produce the same shift in an unrelated series measuring something else.

What should stay explicitly uncertain is the magnitude and the mechanism: how much of the acceleration is AI lowering the cost of building something real, versus AI-adjacent hype pulling registration forward, versus a labour market shifting people toward self-employment for reasons unrelated to AI capability itself. The directional claim, that something changed around 2023 to 2024 in AI-heavy sectors and not in others, is reasonably well supported by convergent evidence. The size of the effect, and how much of it persists once the current AI enthusiasm ordinary settles into routine use, is not yet established by data of this kind, and treating a three-year trend as a permanent structural shift is exactly the overreach that a registration-count history should make anyone cautious about.

Why this matters past one statistic

The entrepreneurship numbers are a clean example because the failure mode is documented and admitted by the source itself, which is rare. Most operational and adoption claims an organisation encounters, about a competitor's AI programme, a vendor's case study, an analyst's productivity estimate, arrive without that admission attached. The question worth asking of any of them is the one this data forces into the open: what specifically was measured, what could move that measurement for reasons unrelated to the claim being made, and does an independent source, measuring something different, move the same way. A number that survives that question is worth acting on. A number that has not been asked it yet is a headline, not evidence.