The Cambrian Explosion of Tiny Companies

Future Shock keeps ending up in the parts of the economy nobody photographs: the billing queues, the approval chains, the structured paperwork we followed in "The Boring Transformation." This piece starts in another of those rooms, with a single payment that refused to match.

Before Caitlin Leksana started building software for accounts receivable, she watched finance teams lose days to a single payment. A wire would land without a usable reference number, and someone had to work out which customer it came from, which invoices it covered, why it was a few hundred dollars short, and then record all of that across systems that disagreed with one another. It is the kind of work that has always required people, because it is mostly judgment about small exceptions, repeated a few thousand times a month.

Leksana and her co-founder Timmy Galvin, who hit the same wall at their previous company, started Fazeshift to hand that work to AI agents that chase the payments, match the cash, draft the messages and flag the disputes. Harvard Business School, which studied the company as one example in a June working paper on AI-native firms, describes it serving enterprise finance departments with about ten employees.

Accounts receivable was never a glamorous department, but it was a real one, and it grew with the business. More customers produced more invoices and disputes, which meant hiring more people to work them. Service companies traditionally scaled by adding the humans who delivered the service. Fazeshift breaks that relationship between customer growth and headcount.

Across the venture-backed companies the HBS researchers examined, AI-native service firms ran roughly 70 percent smaller than comparable peers, with hierarchies nearly a full level flatter. The mean AI startup in their Y Combinator sample had about thirteen employees; its non-AI counterpart had forty-two. The numbers suggest that these firms are being built around a different division of labor.

Fazeshift's headcount captures only one part of the operation. The routine work now lives in software, customer finance teams approve what the agents propose, and unusual exceptions still reach a specialist. The company became smaller because the capability it sells no longer had to fit inside its payroll. That makes headcount a much less reliable guide to where the work happens and who remains responsible for it.

Why companies owned their capabilities

A company has always been a bundle of jobs held together because coordinating them any other way was too slow or too expensive. If a firm needed design, bookkeeping, legal review, customer support and software, it hired people who could do those things and kept them on staff, because assembling that combination fresh for every project would have been untenable. The economist Ronald Coase pointed this out ninety years ago noting that firms exist because going to the open market for every task carries its own costs, and past a certain point it is cheaper to bring the work inside and manage it.

The expensive aspect of bringing more functions in-house was permanence. A business needs some set of capabilities over and over again, so it hires to keep them on hand. Then when the team grows to meet demand, the company hires managers to coordinate the people it hired.

What changes with general-purpose AI is that a founder can attempt work from a neighboring occupation without first hiring someone, contracting someone, or buying expensive enterprise software. They can draft the contract, rough out the design, write the first version of the code, and reserve a human specialist for the parts that carry real risk. Owning a capability and having access to one is beginning to blur for a lot of businesses in 2026.

People are already reaching across those lines. OpenAI, studying how its own product gets used at work, found that 43.5 percent of occupation-specific work messages involved tasks historically tied to a different occupation, and that the crossover ran higher inside the smallest workplaces, where there are fewer specialists to route things to.

That figure measures attempts, not competence. Asking a model a legal question crosses a task boundary; it does not confer a lawyer's judgment about whether the answer is right.

The minimum viable organization is shrinking

Startups show the effect first because they can build around AI from the beginning. Their org charts emerge without an inherited organization that needs to be re-arranged.

The clearest evidence comes from a 2026 working paper by finance researchers at UNC Kenan-Flagler and Penn, who studied nearly 95,000 U.S. startups and used ChatGPT's release as a natural experiment. Companies whose work was more exposed to generative AI cut employment by about 8 percent within two quarters, and the reduction held. Those same firms became more productive and raised money more readily. The number of startups in exposed industries rose about 7 percent. More companies, each carrying fewer people, and no net loss of jobs in the aggregate because the new firms absorbed what the shrinking ones shed.

For a century, "this company is doing well" and "this company is hiring" were close to the same sentence. Investors read headcount as momentum. That link is starting to unfurl. A startup can now show product output, revenue and fundraising traction while its payroll stays flat (though watch out for that AI bill to take up a significant portion of their budget).

The shape of the firm changes along with its size. The HBS study found AI-native startups about 25 percent smaller than their peers, with hierarchies half a level flatter, roughly 15 percent fewer managers, and a heavier concentration of engineers, all at comparable valuations. There are fewer of the coordinating roles whose job is to move information between other people, because there are fewer other people to coordinate.

One obvious objection: maybe this is just productivity. Give everyone a chatbot and the same company does more with fewer hands. But that is not what the HBS researchers found. Simply handing employees ChatGPT did not predict a smaller firm; building the capability into the product did. Gamma, the presentation startup in their study, reached tens of millions in revenue with a few dozen people because slide design, layout and copywriting live inside the software, performed when a customer asks, rather than in a department that does the work by hand. When the product does the job, the capability moves out of the org chart and into the thing being sold.

The floor did not fall evenly

If cheap capability lands in everyone's browser, the businesses with the least internal capacity should gain the most. The solo operator, the one who never had a design department or a legal team, has the most to gain from tools that behave like one.

The JPMorganChase Institute found the reverse. It tracked payments to AI services across millions of small-business accounts, using actual transactions rather than survey answers about whether someone had "tried AI." Employer firms adopted at nearly twice the rate of nonemployer firms, 26.1 percent against 15.3 percent. The gap held within the same revenue bands, so it is not simply that employer firms are bigger.

One likely explanation is that adoption still requires time to evaluate tools, money to absorb failed experiments, and enough organizational slack for someone to figure out where the technology fits. The smallest firms have less of all three. The OECD found the same divide across countries: 40 percent of firms with at least 250 employees used AI in 2024, compared with 12 percent of small firms, whose use remained concentrated in peripheral tasks.

So the technical floor fell faster than the others. The cost of generating a competent-looking design, contract or analysis collapsed. The cost of capital, of distribution, of customer trust, of insurance, of regulatory standing and of the attention needed to integrate any of it did not. The leverage is landing first on firms that already had organizational capacity, which is why the AI-native startups in the HBS data skew toward elite credentials, technical founders and Silicon Valley rather than toward everyone with an idea.

This may also be several trends wearing one big overcoat. America's surge in business formation began in 2020, more than two years before ChatGPT, amid the pandemic, remote work, stimulus and a wave of people leaving jobs. Team sizes started compressing as cheap venture money dried up, which would push founders toward smaller headcounts with or without AI. The available data cannot cleanly separate those forces from the AI effect.

Even so, one thing the confounds cannot explain away sits inside the startup data itself. Compare firms in AI-exposed industries against those that are not, and the exposed ones are more numerous and run leaner, with the flatter, manager-light shape the HBS data already described. That is a narrower claim than "AI caused the startup boom," and it is the one the evidence actually supports.

What the Org Chart Leaves Out

The founder alone at a laptop is a poor model for these companies. Payroll tells us who the firm employs, but not which capabilities live in rented software, which work customers now perform, when specialists enter for difficult cases, or how much shared infrastructure supports the operation.

Contractor payments make one part of that hidden structure visible. Gusto, analyzing payment and banking records across its base, estimated that America's roughly 30 million one-person businesses paid $72 billion to contractors in 2025, while more than four in five U.S. small businesses have no employees. The one-person company often means one permanent employee surrounded by a changing group of contractors.

The emerging firm has a small permanent team coordinating work that sits in several places. Repeatable tasks are embedded in the product, customers handle part of the workflow, platforms supply models, compute, payments and distribution, and human specialists enter for exceptions. Solo founders usually do not remain alone; they remain thin, delaying permanent hires until the business has proved it needs them.

MEDVi shows how difficult those boundaries can become to read. The telehealth company reportedly booked $401 million in sales in its first full year with two employees, according to company financials reviewed by The New York Times. Doctors, pharmacy services, shipping and compliance all ran through outside partners, making the operation vastly larger than its payroll suggested.

In February 2026, the FDA sent MEDVi a warning letter finding that its website displayed compounded semaglutide and tirzepatide with MEDVi on the label, implying the company was the compounder when it was not, and made claims suggesting an FDA evaluation that had never occurred. The warning exposed a mismatch between the company customers could see and the operation behind it. When drugs, medical services and compliance sit across several organizations, an org chart no longer tells a regulator where responsibility resides.

The missing rungs

Employment losses in the startup study concentrated in junior and execution roles. The AI-native firms in the HBS sample likewise carried roughly 15 percent fewer entry-level workers and a heavier weighting toward senior staff. These firms are not shrinking evenly. They are keeping the people who can set direction and judge the work while removing many of the roles in which people learn to do either.

Junior work was never merely cheap production. Matching a stray wire transfer against the right invoice, the task Fazeshift now hands to an agent, taught someone what a real payment dispute looked like before they were responsible for resolving the difficult ones. Drafting a dull contract taught an associate which clauses were boilerplate, which created risk, and when the document needed a more experienced set of eyes. The output mattered, but so did the repeated exposure to ordinary cases that made unusual ones recognizable.

A small firm can now buy senior judgment without maintaining much of the ladder that produced it. That works for the individual company because the wider labor market already contains experienced accountants, lawyers, engineers and operators. It becomes harder to sustain when every company wants senior people and fewer are willing to train junior ones. The firm consumes expertise accumulated elsewhere while contributing less to the next generation of it.

The same tension follows a founder who uses a model to cross occupational boundaries. Generating a contract, financial analysis or piece of code is becoming cheap. Knowing whether it is correct still depends on judgment, and that judgment was often built through the supervised work now being automated. For displaced workers, the cost is already visible. In the startup study, about half remained unemployed for up to six months, while many who found work moved into lower-paid roles less exposed to AI. New firms may eventually create other paths, but the evidence so far does not show that they are replacing the ones being removed.

Headcount still tells us how many people a company employs. It tells us much less about the operation those people can assemble. A ten-person firm can serve enterprise customers by drawing on contractors, model APIs, cloud systems, payment rails and work performed by the customer. Its org chart gets smaller because the operation has been spread across other companies and other people.

That arrangement makes the tiny-company boom less decentralized than it first appears. More firms can compete at the customer-facing layer because none of them has to build its own models, computing infrastructure, payment network, marketplace or distribution system. The same services that lower the cost of starting a company also give a small group of providers influence over which companies can operate and on what terms. A Cambrian explosion of firms can rest on an infrastructure that remains narrow underneath.

The unmatched wire from the opening shows how little permanent organization a company may now need. A founder can sell accounts-receivable capability without first building the department that once delivered it. As the same shift reaches design, bookkeeping, legal work, support and software, more specialized ideas can cross the line into companies, many short-lived and some durable. The number of companies may explode even as the amount of company inside each one collapses.