The Robots Can Already Do Your Job. You're Just Still Cheaper.

They don't need a better robot to replace you. They just need a cheaper one, and they get to vote on the price.

Click to enlarge

The Same Tractor, Both Ends of the Squeeze

If you have read this blog for a while, you have met John Deere before.

Back in Part 2 of the “It Was Never a Dream” series, Deere was my example of Leveragism in a barn: the company that turned a wrench into a service call, so a farmer who paid for the tractor still could not fix it without the dealer's blessing.

This week Deere showed up again, in a place I did not expect. In the robot study I am about to walk you through, the job of agricultural equipment operator scores 2.1 out of 3 on robot exposure, among the highest of any occupation in the country. The example robot cited for that job is a John Deere autonomous tractor.

Sit with that for a second. Same company. Same machine category. Both ends of the squeeze.

• On one end, you cannot fix the machine. The repair is locked behind the dealer.
• On the other end, the machine does not need you to drive it. The operator is the next line item.

That is not a coincidence I am going to pretend is a conspiracy. It is just what it looks like when the people who build the machine also own every point where a human used to be necessary.

What the Study Actually Measured

The paper is called “What work can robots do?” by Russell Legate-Yang and Maxim Massenkoff at Anthropic, published September 30, 2026. Here is the method, stripped down.

They took the Labor Department's O*NET database, roughly 900 occupations broken into about 19,000 specific tasks. They flagged the physical ones, about 7,600 tasks that cannot be automated without a machine that moves through the world. Then, for each task, they asked one question: is there a real robot, deployed, sold, or demonstrated, that can do this today, and where?

The “where” is the clever part. They sorted every task into four tiers:

• E0: No robot can do it. Anywhere.
• E1: A robot can do it, but only in an environment built around the robot. Think fenced-off assembly cell.
• E2: A robot can do it in a normal, structured human workplace. Think warehouse or hospital hallway.
• E3: A robot can do it out in the messy, unpredictable world. Think city street or open field.

The results:

• 74% of physical tasks are robot-doable somewhere. That is 34% of all working hours in the US.
• But only about 2% of physical tasks are doable in the messy world (E3). Most robots still need the world rearranged for them.
• 9 of the 10 most exposed jobs are vehicle operators. Taxi drivers top the list. Light truck drivers, nearly a million of them, score 2.1. Forklift and industrial truck operators, about 774,000 jobs, score 2.0.
• Robots plus AI language models together cover about 81% of all work. The chatbots came for the cubicle. The robots cover most of what the chatbots could not reach.

And here is the part that keeps this from being a guess. They ran it backward. They rated robot exposure for jobs as they existed in 1977, then checked what happened over the following decades. The jobs that were more exposed to the robots of their day saw bigger drops in wages and employment later. The measure predicted the past. That gives it real teeth for predicting the future.

Can vs. Will: The Price Tag Is the Only Wall

If the study stopped there, it would be a panic headline. It does not stop there.

For every exposed task, they estimated what it would actually cost to have a robot do it for a year: the hardware, installation, maintenance, software, energy, insurance, and the human babysitter who still has to watch it. Then they compared that to what the worker costs.

Robots are cheaper than people for just 0.3% of the work.

That is the wall. Here is what it looks like up close:

Source: Figure 7 of the study. Worker cost is median total compensation, not just wages. Robots cover 89% of a taxi driver's tasks, so the fair comparison puts robotaxis about $7,000 a year above the driver, per the authors.

Look at the top row. Packers and packagers are already over the line. The study estimates the robots do that work for about $2,500 a year less than the person. That job has fallen 22% since 2015, and the Bureau of Labor Statistics projects it among the biggest job losers in the country through 2035.

Look at the second row. Robotaxis are basically at parity. What is holding them back, per the paper's own sources, is regulation, not engineering.

Now the forecast. Robot prices have fallen about 3% a year since the 1990s. At that pace:

• It takes about 40 years for robots to be cheaper than humans for even 10% of the work.
• Half of today's physical work does not cross the line until 2085.
• In the authors' fast scenario, with prices falling up to four times faster and capabilities improving twice as fast, half of physical work crosses by 2050.

So the honest read is: not tomorrow, and not never. The capability is mostly already here. What is left is a slow-moving price curve, and a handful of jobs that are already on the wrong side of it.

Who's Standing Behind That Wall

This is where it stopped being an economics paper for me.

The study compared the top fifth of most-exposed workers to the fifth with zero exposure. Here is who is standing on the wrong side of the wall:

• Hourly wage: $22.88 vs. $52.97. About $30 an hour less.
• Unemployment: 5.2% vs. 2.2%. More than double.
• High school diploma or less: 62.6% vs. 13.7%.
• On their feet most of the day: 79% vs. 18%.
• Frequently lifting or carrying weight: 61% vs. 3.5%.
• Working outdoors: 55% vs. 14%.
• Union members: 10.2% vs. 5.2%.

Read that last one again. The most robot-exposed workers are twice as likely to be in a union. The people in the blast radius are, statistically, the people most likely to already know how to bargain as a group. Hold that thought. It comes back.

The authors point out that this profile is close to the mirror image of who is exposed to AI chatbots. The chatbots threaten the college-educated office worker. The robots threaten the person who lifts, drives, sorts, and sweats. Between the two, almost nobody is outside the fence.

The Wall Is Made of Policy

Here is the part the press coverage is going to skip.

That price wall is not just physics and supply chains. It is also tax law. And tucked into this study's own reference list is a 2020 Brookings paper by MIT economist Daron Acemoglu and his colleagues Andrea Manera and Pascual Restrepo with a title that does not bury the lede: “Does the US Tax Code Favor Automation?”

Their answer was yes. By their estimates:

• Labor faced an effective tax rate of roughly 25% to 33%, once you count payroll and income taxes.
• Equipment and software, the exact stuff robots are made of, faced about 10% in the 2010s, falling to about 5% after the 2017 tax law, mostly thanks to generous depreciation rules.
• They estimated that removing that tilt would raise employment by about 6.5% and nudge labor's share of income up by about 1.1 points.

And that tilt did not fade. The federal tax law signed in July 2025 made 100% bonus depreciation permanent for equipment acquired after January 19, 2025, so a company can still write off the whole robot in year one.

To be fair, other economists have pushed back on how sensitive those numbers are to the assumptions underneath them. The size of the tilt is debatable. The direction is not.

Now put that next to the packers row in the table above. The robot saves about $2,500 a year per worker replaced. That is not a productivity revolution. That is a rounding error on a $2 million equipment purchase. Acemoglu and Restrepo have a name for exactly this: “so-so automation.” Machines that are just barely cheaper than the people they replace, produce no real leap in output, and mostly just move money from wages to whoever owns the machine.

The robot study itself makes a related point: automation tends to boost productivity only once machines are much cheaper than people, not merely at break-even. Break-even automation is a transfer, not a breakthrough.

So when a tax code makes the machine look cheaper than it really is, it does not just speed up the robots. It speeds up the worst kind of robots first. Same move I keep writing about. This was not weather. It was written. And anything written can be rewritten.

Who Graded the Homework

Same rule I applied to AIR's agent-skill demo in the cyber series: credit the work, then name who graded it.

Anthropic makes the Claude AI model. In this study, Claude did a lot of the heavy lifting: it classified which tasks were physical, estimated how much time workers spend on each one, searched for robots that could do each task, rated the exposure tier, and estimated the robot costs. The authors are open about this, they published the prompts, and they released the full dataset with Claude's reasoning and cited sources for every rated task. That is more transparency than most vendor research gives you.

But it is still an AI company using its own AI to estimate how much work AI-powered robots can do. Read the precise numbers as estimates with a method attached, not as gospel. Full disclosure on my end: I use Claude in my own work, so weigh my read accordingly.

What holds up regardless of who did the rating:

• The ranking is the real product. The authors show that factoring in cost barely changes which jobs come out most exposed. Drivers and packers first, nurses and mechanics much later. That pattern does not depend on any single dollar figure being perfect.
• The 50-year backtest checks the method against what actually happened to real jobs since 1977.
• The demographic table comes from Census and Labor Department survey data, not from the model.

The Work They Can't Touch Yet

Here is where this ties back to the grill brush.

In the first Building Anyway post, I wrote about Destin Sandlin and the tool-and-die wizard named Roger whose knowledge died with him, and Destin's advice to a room full of billionaires: give your wizard an apprentice, now.

This study turns that from nostalgia into strategy. The jobs robots can barely touch today share two traits: they are hands-on in messy conditions, and they are face-to-face. Installation and repair sits near the bottom of the exposure chart. The authors name nursing and general repair as jobs robots can do little of, even in a perfectly controlled environment.

And the single biggest thing holding robots back is not intelligence. It is hands. The study estimates that about half of all physical tasks will not be automated at scale until robots get much better at touching and handling things. Planning and reasoning, the part AI is getting good at fast, is the main bottleneck for only about 8%.

Think about what that means. The skill this economy spent forty years treating as low-status and disposable, a person who can put their hands into a broken thing and make it work again, is the skill the most expensive robots on Earth still cannot copy.

That is why my whole bet with 3D Nomadic is reprint, swap, repair. Not because it is quaint. Because the person who can fix the thing is the last person standing.

I Refuse to Leave You in the Dark

Same as always: the levee is what changes the structure. The sandbags keep you dry this week. Do not mistake one for the other.

The levee (what we should be demanding):

• Stop subsidizing so-so automation. Tax a dollar spent on a worker and a dollar spent on a machine on level ground. Let robots win where they are genuinely better, not where the depreciation schedule makes them look better.
• Put workers in the room before the robot shows up. The most exposed workers are already the most unionized. Automation terms belong in the contract, not in a press release after the purchase order clears.
• Fund the trades robots can't touch. Repair, installation, maintenance, the work that needs hands in messy places. Apprenticeships, not just factories.
• Right to repair applies to robots too. If the autonomous tractor can only be fixed by the dealer, the lock-in just moved up the chain. Whoever deploys the machine should be able to keep it running.
• Watch the canaries in public. The study itself says drivers and warehouse packers are where disruption shows up first. Track them openly, with real numbers, so nobody can claim surprise.

The sandbags (what you can do Monday):

• Look up your own job. Anthropic released the full dataset, task by task. If your work is mostly E2 or E3, you are not paranoid. You are early.
• Move toward the hands. Inside almost any job, there are tasks robots fumble: fixing, troubleshooting, messy conditions, dealing with people. Volunteer for those. They are the moat.
• Learn to fix one real machine. Same advice as last time, now with data behind it. A repair skill is the one asset nobody can meter or automate cheaply.
• Join, or build, the union. The table says the people most at risk already have the most practice organizing. Use it before you need it.

Apps, side hustles, and clever career moves are sandbags. They protect you. They do not change who sets the price of a robot.

Why I'm Telling You This

I have spent most of my career on the implementation side of automation. Endpoint management is, at heart, the business of making one person do the work that used to take a room full of people. I know how fast a budget meeting greenlights headcount reduction once the tooling works. Nobody in that room ever called it replacing people. They called it efficiency.

So when I read a study that says the machines can already do most of the work, and the only thing holding them back is price, I do not read it as a forecast. I read it as a countdown that somebody else controls.

Despair says the robots are coming and there is nothing to do but wait. Fury says the clock is set by tax law, labor contracts, and repair rights, all of which were written by people, which means they can be rewritten by people.

The robots can already do your job. You are just still cheaper. The question worth fighting over is who gets to decide when that stops being true.

Read the study. Learn to fix something. Then get loud with me.

Sources & Further Reading

• Anthropic (Russell Legate-Yang & Maxim Massenkoff): “What work can robots do?” (Sept. 30, 2026), plus the full PDF and the task-level dataset. Source for the exposure tiers, the 74% / 34% / 0.3% figures, Figures 4, 5, and 7, the cost-decline scenarios, the 1977 backtest, and the John Deere citation.
• Acemoglu, Manera & Restrepo: “Does the US Tax Code Favor Automation?” (Brookings Papers on Economic Activity, Spring 2020). Source for the effective tax rates on labor vs. equipment and software.
• Brookings: summary of the tax-code paper, including the 6.5% employment and 1.1-point labor-share estimates.
• MIT Work of the Future: “Taxes, Automation, and the Future of Labor”, the authors' plain-language version.
• Eric Zwick (discussant comments): critique of the tax-code paper's assumptions.
• Smarter Every Day: “The PROBLEM with Capitalism”, the source of the Roger and apprentice story from Building Anyway #1.

Sourcing note: every robot-exposure, cost, and demographic number comes from the Anthropic study's own text and figures. The robot cost estimates and task time shares were produced by Anthropic's Claude model, so treat exact dollar figures as estimates. The taxi comparison uses the authors' exposure-adjusted figure, not a raw subtraction from the table. The tax figures come from a 2020 paper and reflect the tax code at that time; check current law before repeating them. The “so-so automation” term belongs to Acemoglu and Restrepo. Verify anything you plan to repeat.

Connect with Casey

Have a story, a question, or want Casey to write about a specific topic? DM me and tell me which story you want next.

Casey writes about economic policy, nomadic life, cybersecurity, and navigating the world as a late-diagnosed AuDHD adult. New posts drop on my professional website.