74% of Jobs Can Be Done by Robots, But Only 0.3% Are Cost-Competitive | Anthropic
機械翻訳 / Machine-translated
A
AI Friends
1日前・19分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever worried about AI or robots taking your job?
Anthropic, the company behind Claude, has released a study that answers that question with hard numbers. The conclusion: robots can do a lot — but they're still too expensive. This article breaks down the findings and their implications for the way we work in Japan.
On September 30, 2026, Anthropic published a research report titled "Can we predict the jobs robots will do?"
The report was authored by Russell Legate-Yang and Maxim Massenkoff, both members of the company's economic research team.
Anthropic has previously studied how LLMs (AI that can write like a human) — such as Claude — affect the workforce.
This study extends that scope to include physical work: carrying packages, driving vehicles, cleaning, and other tasks that require bodily movement.
It's worth noting that it's unusual for an AI research company to publish an economic analysis focused on robotics. This reflects the growing interest in physical AI — AI that moves and acts in the real world.
The study drew on O*NET, the U.S. occupational database. It covered 7,594 physical tasks across roughly 900 occupations.
For each task, Claude assessed whether today's robots could perform it. The assessment used a four-level scale:
The assessments were informed by real-world examples of robots currently in use — such as Waymo and Tesla self-driving vehicles, Amazon's warehouse robot "Vulcan," and Boston Dynamics' unloading robot "Stretch."
Tasks were also weighted by how much time workers spend on them, so frequently performed tasks have a greater impact on the results.
The results show that 74% of physical tasks can be handled by robots in some capacity.
Measured by working hours, 34% of total labor time corresponds to tasks that today's robots can perform.
Combined with office tasks that LLMs can handle, roughly 80% of all work is affected by AI or robots. LLMs alone cover about half — adding robots dramatically expands the scope.
When narrowed to tasks where "hiring a robot is cheaper than hiring a human," the figure drops to just 0.3%.
The cost calculation includes not just the purchase price, but also installation, maintenance, software, electricity, supervision, insurance, and disposal costs.
Upfront costs such as the robot's purchase price are amortized over an assumed 10-year lifespan with an 8% annual cost of capital (similar to an interest rate on borrowed money).
Robot prices have fallen by around 3% per year since the 1990s. Even if that pace continues, the study estimates it would take approximately 40 years for cost-competitive tasks to reach 10%.
Among occupations, taxi drivers had the highest robot exposure, topping all professions with an index score of 2.2.
Shuttle bus drivers, chauffeurs, and warehouse workers also ranked near the top — because the majority of their working hours are spent on driving or moving goods.
Delivering medication within hospitals also scored high, given the existence of delivery robots like RelayRx.
On the other hand, tasks rated E0 (not possible for robots) include:
According to the study, around 70% of physical tasks are limited by the need for grasping and manipulation. A further 14% are constrained by regulation, and roughly 25% by human preference — people simply want a human to do the work.
The closest example of robots matching human cost is packaging workers. About 97% of packaging tasks were rated as robotically feasible.
A robotic system capable of replacing roughly 14 workers costs over $2 million to implement, but that works out to approximately $45,000 per worker per year.
Since packaging workers earn around $49,000 annually, robots come out roughly $2,500 cheaper per year.
By contrast, robots cost about five times more than humans for welding. For dishwashers and cleaners — where wages are lower — robots are several times more expensive.
Imagine a fulfillment center for an e-commerce company shipping tens of thousands of items a day.
In this setting, robots are already slightly cheaper on paper. If labor shortages are severe, the decision to adopt them could come even sooner — though the upfront investment of several hundred million yen means larger companies will likely move first.
Now picture a neighborhood hair salon. Coloring requires fine motor skills, and customers tend to prefer having a human do it.
The study rated this E0 — not suitable for robots. It seems unlikely that robots will be replacing hairstylists anytime soon.
Finally, consider a regional taxi company. Driving is technically well-suited for robots, and Waymo has announced plans to operate driverless taxis in Tokyo as early as 2027.
In Japan, however, traffic regulations and public acceptance present significant barriers — precisely the "regulation" and "preference" walls the study identifies.
A similar conclusion had actually been reached before. In 2024, MIT's CSAIL (Computer Science and Artificial Intelligence Laboratory) published a comparable study.
MIT focused on image recognition AI and found that only 23% of tasks that could technically be automated by AI would actually be cheaper to automate — just about 0.4% of the overall U.S. economy.
| Anthropic (2026) | MIT CSAIL (2024) | |
|---|---|---|
| Subject | Robots (physical tasks) | Image recognition AI |
| Technically feasible | 74% of physical tasks | 1.6% of wages |
| Cheaper than humans | 0.3% of tasks | ~0.4% of the economy |
| Method | Claude assessed 7,594 tasks | Researchers modeled estimates |
What both studies share is a large gap between what's technically possible and what's economically worthwhile.
Anthropic's study also cross-referenced 50 years of employment data. It confirmed that since 1977, jobs with higher robot exposure have seen greater wage and employment declines — lending some credibility to the rating methodology.
This study is based on U.S. data, so it doesn't translate directly to Japan.
That said, Japan already has a strong track record of factory robotics, and labor shortages in logistics, caregiving, and construction are becoming increasingly serious due to an aging and shrinking population.
In sectors where workers are scarce, rising wages close the cost gap with robots faster. Labor shortages could accelerate robot adoption in Japan more quickly than in the U.S.
The areas most likely to see an impact include e-commerce packaging and sorting, intra-warehouse transportation, and in-hospital delivery. Conversely, beauty services, hands-on caregiving, and on-site assembly work are likely to remain human jobs for the foreseeable future.
For workers, the ability to work alongside robots will become an increasingly valuable skill — whether that means operating machines, monitoring their performance, or handling malfunctions.
No. The 74% represents tasks that are technically feasible for robots. Since only 0.3% are actually cheaper than human labor, large-scale replacement is not imminent.
Taxi and shuttle bus drivers, warehouse workers, and packaging workers are among the most exposed — because practical robots for driving and goods handling already exist.
The research team validated the results against 50 years of employment data. Jobs with higher robot exposure were confirmed to have experienced actual declines in wages and employment, lending the methodology a degree of reliability.
Potentially. The report acknowledges that humanoid robots and new manufacturing methods could bring costs down further. However, the timing is uncertain, and the barriers of regulation and human preference remain.
The study is based on U.S. data, so the specific numbers don't directly apply. However, given Japan's serious labor shortages, robotization in sectors like logistics may accelerate faster than in the U.S.
As a practical exercise, try listing how much of your own workday involves driving, transporting goods, or packaging — it may give you a clearer sense of your exposure.
This article is a cross-post from AI Friends.
