5 Job Categories That Grew 3x Despite Automation — And Why the AI Era Will Repeat the Pattern
Nail salons, pet care, and tutoring each tripled in employment since 1990 despite automation fears. Here's why economists think AI will follow the same…
Five Job Categories That Grew 3x Despite Automation — And Why AI Will Repeat the Pattern
A Chicago Booth economist just made the most data-grounded case against the AI job apocalypse you’ll read this year. Alex Emas’s essay “What Will Be Scarce” landed in mainstream discourse when Ezra Klein cited it in an opinion piece arguing the AI job apocalypse probably won’t happen — not in the way the most fearful predictions suggest. The argument isn’t optimism for its own sake. It’s grounded in a specific historical claim: when automation displaces labor in one sector, surplus flows somewhere else. It doesn’t evaporate.
The anchor data point is this: nail salons, pet care, exam prep and tutoring, and athletic coaches and umpires each had fewer than 100,000 workers in 1990. Today each category employs between 150,000 and 350,000 people. These aren’t sectors that were protected from automation. They grew alongside it.
That pattern — labor surplus from one sector flowing into new categories rather than disappearing — is the core of the historical case. And it has implications for how you should think about what AI actually does to employment over the next decade.
US Agriculture Went from 70% of Employment to Under 5% — and Unemployment Didn’t Spike
Start with the most dramatic example in the data. In 1850, roughly 70% of US employment was in agriculture. Today it’s under 5%. That is a near-total displacement of the dominant employment category in the country over 170 years.
Coding agents automate the 5%. Remy runs the 95%.
The bottleneck was never typing the code. It was knowing what to build.
The standard fear response would predict: mass unemployment, social collapse, a permanent underclass of displaced farm workers. None of that happened. What happened instead was a diversification of the labor market into sectors that barely existed in 1850 — leisure and hospitality, private education and health, professional and business services.
The A16Z piece by David that circulated this week shows this as a chart of US employment by sector since 1850. If you haven’t seen it, it’s worth finding. The visual story isn’t a cliff where farm jobs fell and nothing replaced them. It’s a fan — multiple new sectors growing as agriculture shrank.
The mechanism matters here. More productive farming didn’t just make food cheaper. It enabled a population boom, because the world could support more people. More people meant more demand for everything else. The surplus didn’t disappear; it compounded. This same compounding logic is why economists studying AI diffusion tend to focus on the AutoResearch loop pattern as a model for augmentation rather than replacement — AI runs the repetitive cycles, humans direct the strategy, and total output expands rather than headcount contracting.
Spreadsheets Killed Bookkeeping Jobs — and Created More Finance Jobs Than They Destroyed
The agriculture example spans 170 years, which makes it easy to dismiss as too slow to be relevant. The spreadsheet example is tighter.
When spreadsheets arrived, bookkeepers and accounting clerks saw a steady decline that lasted roughly 30 years. If you froze the frame at the moment of disruption, it looked like a straightforward substitution story: software replaces humans, humans lose jobs.
But the full picture is different. Financial analysts grew. Accountants and auditors grew. The spreadsheet didn’t eliminate financial work — it made financial analysis cheap enough that more businesses could afford it, which expanded the total market for financial expertise.
This is what economists call the Jevons paradox applied to labor: a productivity improvement in a sector often increases total demand for that sector’s output, because the lower cost opens up new markets. The tool that automates the routine work tends to expand the scope of the non-routine work that surrounds it.
The same dynamic is visible in software engineering right now. Despite years of predictions that AI coding tools would reduce demand for engineers, the data on public market earnings calls shows companies mentioning AI “augmenting” workers over “substituting” them at an 8:1 ratio. That’s not a rounding error. That’s a signal about how companies are actually deploying these tools internally — as force multipliers, not replacements. It’s also worth noting how model capability has expanded the surface area of what’s automatable: the comparison between GPT-5.4 and Claude Opus 4.6 illustrates how different frontier models are being optimized for different workflow roles, which suggests specialization rather than wholesale job elimination.
Nail Salons: A Category That Barely Existed in 1990 Now Employs 150,000+ People
Here’s where the argument gets specific in a way that’s hard to dismiss.
Nail salons as a distinct employment category had fewer than 100,000 workers in 1990. The industry grew substantially through the 1990s and 2000s — not because of some unique economic protection, but because rising productivity elsewhere put more disposable income in people’s hands, and people spent some of that income on personal services that had previously been unaffordable luxuries.
This is the mechanism Emas focuses on in “What Will Be Scarce”: when one sector gets disrupted, the surplus flows somewhere. The question is where. His answer is the “relational sector” — goods and services where the value comes not just from the output but from who provides it and how. A manicure from a skilled technician you trust is not the same product as a manicure from a stranger, even if the physical result is identical.
That relational dimension is, by definition, resistant to AI substitution in a way that pure information work is not. You can automate the production of a financial report. You cannot automate the relationship between a client and the advisor they’ve worked with for fifteen years.
Pet Care: From Niche to 200,000+ Workers as Productivity Gains Created New Spending Categories
Pet care follows the same trajectory. Under 100,000 workers in 1990, now well over 150,000 — with some estimates putting the figure closer to 200,000 when you include adjacent roles.
The interesting thing about pet care as an example is that it’s not a sector that benefited from any particular technological protection. Robots can’t walk your dog, but that’s not why the sector grew. It grew because rising incomes — partly driven by productivity gains elsewhere in the economy — meant people had money to spend on services for their pets that they previously would have done themselves or foregone entirely.
This is the “new categories” argument in its clearest form. The jobs that AI creates won’t all be “AI jobs” in the sense of requiring AI expertise. Many of them will be in categories that don’t exist yet, serving demand that doesn’t exist yet, enabled by the productivity surplus that AI generates.
When you’re building AI-powered workflows — say, using MindStudio to chain models across 200+ integrations and automate research or customer service tasks — the goal is usually to free up human time for higher-value work, not to eliminate the humans doing that work. The 8:1 augmentation-to-substitution ratio on earnings calls suggests that’s how most companies are actually thinking about deployment.
Exam Prep and Tutoring: A Sector That Tripled as Education Became More Accessible
Exam prep and tutoring had fewer than 100,000 workers in 1990. Today the category employs somewhere between 150,000 and 350,000 people, depending on how you count adjacent roles.
The growth here is partly a story about rising educational stakes — more competitive college admissions, more professional certifications, more standardized testing. But it’s also a story about accessibility. As other productivity gains made tutoring services cheaper and more scalable, demand expanded beyond the wealthy families who could always afford private tutors.
This is the pattern that should make AI pessimists pause. The fear is that AI tutoring tools will eliminate human tutors. The historical pattern suggests the more likely outcome is that AI makes tutoring cheap enough that demand expands dramatically, and the total number of people employed in education-adjacent roles grows — even if the nature of those roles shifts.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
The relational dimension matters here too. Parents who can afford a human tutor for their child often prefer one, even when AI alternatives are available and cheaper. The human relationship is part of the product. The same logic applies to browser-based automation tools: using Claude Code with Playwright for web scraping and form filling automates the mechanical layer of research work, but the judgment about what to research and what to do with the results still requires a human in the loop.
Athletic Coaches, Umpires, and Officials: 350,000 Jobs in a Category That Barely Registered in 1990
Athletic coaches, umpires, and related roles represent perhaps the clearest example of a category that grew almost entirely because of rising disposable income and changing preferences — not because of any protection from automation.
Under 100,000 workers in 1990. Now up to 350,000, making it the largest growth story in this set of examples. The growth tracks the expansion of youth sports, adult recreational leagues, fitness culture, and the broader “wellness economy.”
None of this was predictable from the vantage point of 1990. If you’d asked an economist in 1990 to forecast where labor surplus from manufacturing automation would flow, “youth soccer coaches” would not have been a top answer. But that’s exactly the point. The new categories that absorb displaced labor are often categories that don’t exist yet in any meaningful form at the time of the disruption.
The jobs that AI creates over the next decade will likely follow the same pattern — categories that are hard to name today because the demand that creates them doesn’t exist yet. The productivity surplus from AI will put money in people’s pockets and time in their schedules, and they will spend both on things that are currently unaffordable or unavailable.
The Layoff Headlines Are Misleading — and Markets Are Starting to Notice
The week that surfaced all this data also produced layoff headlines from Coinbase and Cloudflare, both of which prominently cited AI as a factor. Most outlets ran the AI angle without much scrutiny.
But the specifics don’t support the narrative. Cloudflare laid off 1,100 people — after hiring 2,000 new people just a few months earlier. That looks like an overhiring correction, not AI displacement. Coinbase’s layoffs came just before an earnings report showing transaction revenue fell 40% year-over-year. That’s a crypto market problem, not an AI problem.
The point isn’t that AI has zero role in any layoffs. Companies are recalibrating, and on average they will probably be smaller in five years than they are today even while producing more. But the reflexive “AI did it” framing on every layoff announcement is increasingly getting pushback from people who look at the company-specific numbers before accepting the headline.
This matters for how you build. If you’re designing AI workflows or agents, the question isn’t “how do I automate this job?” — it’s “how do I free up this person to do the work that actually requires them?” That distinction shapes what you build and how you measure success. Tools like Remy take a similar approach to the abstraction problem in software: you write a spec — annotated markdown — and the full-stack TypeScript application gets compiled from it, including backend, database, auth, and deployment, so the developer’s attention goes to intent and precision rather than boilerplate. The source of truth shifts up the stack; the work doesn’t disappear.
The Timeline Is the Variable That Changes Everything
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
The strongest version of the job displacement argument isn’t “AI will eliminate jobs” — it’s “AI will eliminate jobs faster than new categories can absorb the displaced workers.” That’s a timeline argument, not a structural one.
And here’s where the current evidence is genuinely reassuring. The biggest AI labs are distracted by painful enterprise deployment problems — Anthropic and OpenAI both launched massive joint ventures this week specifically to solve the gap between model capability and actual enterprise adoption. That gap is real and large. The capability overhang is enormous, but closing it requires solving authentication, compliance, workflow integration, and organizational change management. None of that happens in a year or two.
If the timeline for AI to actually diffuse through the enterprise is a decade rather than two years, the historical pattern becomes much more applicable. Reskilling is possible over a decade. New job categories can emerge over a decade. The adjustment is painful but manageable.
Larry Fink at BlackRock put it plainly this week: “Not only is there not an AI bubble, but there is the opposite. We have supply shortages. Demand is growing much faster than anyone has anticipated.” Jamie Dimon at JPMorgan said he believes “the trillion dollar investment in data centers will make sense.” These aren’t AI enthusiasts making the case — these are capital allocators whose job is to be right about where the economy is going.
The historical pattern from agriculture to spreadsheets to nail salons is consistent: automation displaces specific tasks, not work itself. The surplus flows somewhere. The question is always where — and the answer is usually categories that didn’t exist yet when the question was being asked.
That’s not a guarantee. But it’s a much stronger prior than the apocalypse framing suggests.
If you want to go deeper on the economic argument, Alex Emas’s essay “What Will Be Scarce” is the primary source worth reading. Ezra Klein’s piece in the Times is a good entry point if you want the mainstream framing first. And if you’re thinking about what AI-augmented workflows actually look like in practice — as opposed to the substitution story — the AutoResearch loop pattern that Karpathy described is one concrete example of augmentation rather than replacement: AI runs experiments overnight, humans evaluate and direct in the morning.
The five categories above — agriculture, bookkeeping, nail salons, pet care, tutoring, athletic coaching — aren’t cherry-picked outliers. They’re the pattern. And the pattern has held across 170 years of automation waves more disruptive than anything AI has produced so far.
