The U.S. Census Bureau said July 9 that it received 531,423 business applications in June on a seasonally adjusted basis, capping what was for the first half of the year the highest sustained rate of such applications in the 22-year history of the series. The surge is being interpreted as the first real evidence of an AI-driven entrepreneurial boom. Not necessarily. Artificial intelligence has made the mechanics of starting a business nearly free, and the mechanics were never the barrier that mattered.

When you apply for an employer identification number, the Internal Revenue Service asks when you expect to start paying wages. It is one line on a form, and you can leave it blank. That’s happening more often, and it matters more than the total number of forms submitted.

Through the 2010s the rate of new business formations was falling, triggering worries that the country’s dynamism was fading. Jorge Guzman of Columbia Business School and Scott Stern of MIT looked for a better measure. They took 28 million business registrations across 32 states covering 81% of U.S. gross domestic product and asked how many of those firms ever amounted to anything. Only 0.07% reached an initial public offering or a meaningful acquisition within six years. More than a third of those came from the top 1% of firms they rated highest quality based on their characteristics the day they registered.

Other economists have found the same pattern in different data. Vincent Sterk, Petr Sedláček and Benjamin Pugsley, using Census records covering nearly every private employer in the country, showed that most of the difference in how fast startups grow is set before they open, and a small subset of high-potential firms drives the gains in aggregate output. The count of new businesses doesn’t matter, but the count of high-quality new ones does. And in the 2010s that was going up.

So, what’s happening now? Even as total applications climb, those from people who told the IRS when they expected to start paying wages have been falling for almost two years and are down by double digits again.

Then why isn’t AI producing more high-quality startups? It’s certainly useful for entrepreneurs. I used to send speaking contracts to a lawyer. Now I just ask Claude. That scales. A working paper posted in December by Junhui Jeff Cai and five co-authors tracks 12.8 million Chinese firm registrations through the end of 2024 and finds new firm formation surging in the places where AI expertise was already concentrated. Small firms went from 33.8% of new registrations before ChatGPT to 51.4% after, and entry by large firms fell.

But look at what AI is helping the firms in that research do: draft legal documents, write marketing copy, basic coding, assemble the business plan. Those are all annoying, time-consuming, and potentially even costly. But they’re not hurdles that would have stopped many high-quality startups pre-AI.

The research on who AI helps points somewhere fascinating. Erik Brynjolfsson of Stanford University, with Danielle Li and Lindsey Raymond of MIT, studied 5,172 customer support agents as an AI assistant rolled out across them and found productivity up 15%, with the gains concentrated among the least experienced. The most experienced agents saw small gains in speed and small declines in quality. AI was acting as a leveler, compressing the differences between the most and least-skilled workers.

Customer support is a well-defined problem with clear right and wrong answers. That’s where AI shines. Entrepreneurship, as my entrepreneur wife often reminds me, is nothing like that. Berkeley’s Nicholas Otis and four co-authors gave 640 Kenyan entrepreneurs a GPT-4 business mentor over WhatsApp for five months and tracked their revenues and profits. On average, nothing happened. Underneath the average, the best entrepreneurs got better and the struggling ones got worse. Both groups asked the model similar questions and got similar advice. What separated them was which advice they acted on.

The model handed both groups the same answers. But a lack of answers wasn’t the problem. Knowing which answer to act on was. So cheaper execution is boosting the number of filings, not their quality. The mechanics were never what separated successful and unsuccessful startups. Guzman and Stern found quality is visible the day a firm registers, and Sterk and his coauthors found the same thing in the Census records.

AI can’t make more people with good judgment. So the number of high-quality startups doesn’t change. But the amount they make could, as AI tools cut their costs and magnify their value. This isn’t about who owns the models. It’s about what the models do. Even if these high-quality startups drive growth, it could all accrue to the people the economy already rewards. If the models get good enough to be trusted on questions with no checkable answer, that changes.

The applications filed in June do tell us something real. Starting a business has never been easier, which matters for the people starting them. What they can’t tell us is whether any of it will make the country richer, because that was never a question of mechanics. It comes down to one line on the form, the one asking when you expect to start paying wages. Fewer and fewer people are filling it in.

____

This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

Gautam Mukunda writes about corporate management and innovation. He teaches leadership at the Yale School of Management and is the author of "Indispensable: When Leaders Really Matter."

©2026 Bloomberg L.P. Visit bloomberg.com/opinion. Distributed by Tribune Content Agency, LLC.