Industrial Revolution 2.0: The One-Person Factory
Factories made workers more productive by combining machines with divided labor, and corporations coordinated mills across cities. AI may put a whole factory under one person's direction. Here is what the evidence shows, what it doesn't, and who gets the gains.
At 5:25 a.m. Chicago time on November 13, two issues opened in this site’s repository. Issue #204 carried my direction for a post called Own Your Memory. Issue #203, filed four seconds earlier, asked for its cover.
Five minutes later a Claude Code writer opened pull request #206 with a draft and a research report. At 5:39 a Codex reviewer, a separate session that had not written a word of it, sent the draft back. One of its findings was a date. The draft said Anthropic cut off OpenAI’s access to Claude in August 2025. The reviewer pointed out that Wired’s original report ran on August 1 but dated the cutoff to the Tuesday before, July 29. The writer couldn’t fetch Wired to confirm it, so the post now says what can be checked: the date Wired reported it. The research report records why. A second commit added the primary statements behind the post’s other examples.
Meanwhile a Codex image worker made the cover pair, and its pull request, #210, merged at 5:55. The post merged at 6:08. From filing to merge took 43 minutes.
I chose the subject. Agents did the research, the writing, the cover and the review.
That is a small factory. I think it is the beginning of a much larger idea: the one-person factory.
What the first factories combined
It’s worth being precise about what the first Industrial Revolution actually combined, because “factory” names three different things.
Machinery. In cotton spinning, the labor to spin 100 pounds of cotton fell from more than 50,000 hours in the early 1700s to 2,000 with Crompton’s mule in 1780, 1,000 with a 100-spindle mule around 1790 and 300 with power-assisted mules around 1795, per figures Daron Acemoglu and Simon Johnson cite from Stanley Chapman (Acemoglu and Johnson, 2024, p. 601).
Factory organization. Machines don’t explain everything. The Boston Manufacturing Company, founded at Waltham, Massachusetts, in 1814, put every step of making cloth “under a single roof” (National Park Service). Mills in this system were split into departments: picking, carding, spinning, weaving and cloth finishing, each room under an overseer who supervised 40 to 60 workers, most of them young women (UMass Lowell Center for Lowell History). A handloom weaver had also made and starched the warp. Starting with the first power looms at Waltham, separate workers did those jobs, often on new specialized machines (Bessen, 2008, p. 12).
Corporate coordination. Above the mill sat the corporation. Its directors were, in the Park Service’s words, “the true masters of the mill.” A treasurer, usually in Boston, bought the cotton and helped set the price of cloth. A salaried agent ran the mill on site and could hire and fire (National Park Service). The investors behind Waltham went on to build mills in Lowell, Chicopee, Manchester and Lawrence, and controlled a fifth of America’s cotton production by 1850 (National Park Service).
One weave room, one century
Follow the weaver. James Bessen reconstructed output at benchmark mills from census and payroll records (Bessen, 2008, Table A1):
| Year | Mill | Yards per loom-hour | Looms per weaver |
|---|---|---|---|
| 1810 | A handloom (estimate) | 1.5 | 1 |
| 1819 | Boston Manufacturing Co., Waltham | 2.1 | 1.84 |
| 1835 | Lawrence Co. Mill No. 2, Lowell | 3.9 | 1.84 |
| 1855 | Lawrence Co. Mill No. 2, Lowell | 3.8 | 3.43 |
| 1901 | Three coarse-cloth mills | 5.1 | 7.3 |
By 1902, a weaver produced “over 25 times as many yards of cloth in an hour” as a weaver a century earlier, running seven or eight power looms instead of one handloom. Output per loom rose almost four times. Looms per weaver rose seven times (pp. 2, 13).
The striking row is Lowell from 1835 to 1855. The looms didn’t get faster: 3.9 yards an hour, then 3.8. Each weaver went from 1.8 looms to 3.4. Bessen estimates that over the whole century, one third to one half of the technical change came not from inventions but from what workers learned on the job (pp. 19, 28).
That is the pattern I care about. The gain came from one person directing more machines at once, and learning how.
Industrial Revolution 2.0
Here is my argument, and I want to be clear that it is an argument, not a finding.
The factory raised a worker’s output by surrounding the worker with machines and dividing the work around them. The corporation coordinated many factories, departments and cities. Both needed many people, because only people could do the specialized work and only people could coordinate it.
AI agents can now do much of the specialized work, and, as in the example above, a good deal of the coordination. Research and writing, illustration and review used to be separate departments. In my repository they are separate agent roles, and one person directs all of them. I think AI will give one person, or a small team, more productive reach than a weaver with eight looms, and more than many corporations had.
The Lowell weaver is a better analogy than the inventor. Output per weaver grew because each weaver ran more looms, and because the weavers and the mill learned how. The one-person factory is the same move: more agents per person, and a person who gets better at directing them.
What the evidence shows
Some evidence points this way. It is narrower than the argument.
Individuals with AI matched teams without it. Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani and colleagues ran a field experiment at Procter & Gamble in 2024 (Organization Science, 2026). In the working-paper version, 776 professionals each spent a one-day virtual workshop developing a solution to a real business need for their business unit (the journal version reports 791). They were randomized to work alone or in a pair of one R&D and one commercial specialist, with or without GPT-4 (NBER Working Paper 33641, pp. 4, 9-10). Blind experts scored quality. Teams without AI scored 0.24 standard deviations above individuals without AI. Individuals with AI scored 0.37 above, which the authors describe as comparable to teams without AI, and spent 16.4% less time. Teams with AI were 9.2 percentage points more likely to produce a top-10% solution. With AI, the gap between R&D and commercial ideas disappeared (pp. 14-17).
What it doesn’t show: this was one day, one company and pairs of people. The authors note that one-day virtual collaborations did not capture the extended coordination and iteration of real teams (p. 20). A person with AI matched a pair on a brainstorm. That is not the same as one person running a company.
Cheap work expands output. Mollick has argued that companies which treat AI only as a way to cut costs will miss its value, the way a steam-era firm that fired workers would have missed “building world-spanning companies by expanding their outputs” (One Useful Thing, June 2024). He also argues that the people doing the work are the ones who will discover what it can do. That fits Lowell. It fits my repository too, where the process gets rewritten by the people, and agents, running it.
The gains need organization first. Erik Brynjolfsson, Daniel Rock and Chad Syverson argue that general-purpose technologies like AI require large, mostly intangible complementary investments in new processes, business models and skills, which national accounts measure poorly. Productivity looks low while those investments are made and high once they pay off. They call it the Productivity J-curve (American Economic Journal: Macroeconomics, 2021). A model alone does not make a one-person factory. The rules, roles, review and tooling around it are the investment, and they take time to pay back.
Where people are still essential
In the 43-minute example, I chose the subject. Everything else was agents. But look at what the agents depended on.
- Judgment about what to make. No agent decided that a post about memory sovereignty should exist.
- Primary sources. The reviewer’s correction rested on a reporter who wrote the original story. Agents checked facts. People produced them.
- The rules. Independent review, the order a cover and post merge in, who may deploy: people wrote those rules, and they change only through reviewed pull requests.
- Responsibility. When the post is wrong, I am wrong.
The P&G study found the same split. In the journal version’s summary, AI mainly improved the quality of ideas, while “human judgment retains value in evaluative selection” (Organization Science, 2026).
Who captures the gains
This is the hard part, and the first Industrial Revolution is a warning.
Output soared. Cotton went from about 1% of British GDP in the early 1780s to 7-8% by 1811-13. Meanwhile real wages for handloom weavers “more than halved between 1806 and 1820,” and their numbers fell from a peak of 240,000 to 3,000 by 1862. Economy-wide real wages were largely flat into the 1830s (Acemoglu and Johnson, 2024, pp. 598, 603, 609, 611). David Ricardo, who had told Parliament in 1819 that machinery did not reduce the demand for labor, added a chapter “On Machinery” in 1821 saying he had changed his mind (p. 598).
Acemoglu and Johnson’s conclusion is that automation helps workers only when it creates new tasks or workers have the power to claim a share. Their summary: “If AI is used to create new tasks and increase human capabilities, the benefits would be more likely shared” (p. 617).
So who captures the gains of a one-person factory? My bet is the person who owns it. That’s the hopeful version of this story. In 1814 you needed Boston capital to build a mill. In 2026 the factory is a repository, a few subscriptions and the rules you write, so the person who directs the work can also own it. That is why I argued you should own your factory, not rent it.
The other version is also possible. If the factory belongs to a platform and the person only rents a seat, the gains go where they went in 1820. Acemoglu and Johnson are right that this is a choice, not a law.
The weaver with eight looms
The weaver of 1902 did not invent the power loom. The weaver ran seven or eight of them and made 25 times as much cloth an hour as a weaver a century earlier, and much of that gain came from learning how.
That is the one-person factory. You don’t need to invent the agents. You need to learn to run many of them, organize the work around them, and own what they make.