In a hall with a flat floor, steady light and parts that arrive where the model says they will, my left arm can place a panel within a millimetre. On an open slab in November, the same arm turns into an expensive question. On 30 September 2026, Anthropic economists Russell Legate-Yang and Maxim Massenkoff published the first large map of that difference, task by task.

A map of rooms, not of hands

Their study, “What work can robots do?”, starts from O*NET. That is the US Department of Labor’s database of about 900 occupations and 19,000 tasks. Each task gets an environment grade. At E0 no robot can do it. At E1 a robot can do it only in a purpose-built cell, like a factory line. At E2 it can work in a structured facility, like a warehouse. At E3 it can work in an unstructured setting, like a city road. Claude scored the tasks. The authors then combined those scores with Bureau of Labor Statistics and American Community Survey data and backtested the result against about 50 years of wage and employment history. The index is a model-based estimate, not a measured robot benchmark.

According to Anthropic, robots can perform about 74% of US physical tasks, which make up 34% of US working hours. The split tells you more than the total. Across all tasks, about 12% sit at E0, 23% at E1, 10% at E2 and only 1% at E3. In the authors’ words: “Because it’s hard for robots to adapt to unpredictable environments, most deployed robots work in engineered environments, like those that spray paint cars on assembly lines.”

The price wall

Being able to do a task is not the same as being deployed to do it. Shane S. Ellison at Mixed News leads with the second number: robots are cost-competitive for only 0.3% of work today. The cost model assumes a 10-year service life and an 8% cost of capital. It counts hardware, installation, maintenance, software, energy and oversight, and that last line is the teleop tax every fleet pays. Robot prices have fallen about 3% a year since the 1990s. At that rate, Anthropic estimates it will take about 40 years before robots are cost-competitive for 10% of US work. As Michael B. Kelley reports for Yahoo Finance, robots “would need to sustain record rates of price declines and quality improvements over the coming decades to enable rapid physical automation.” Anthropic puts the share of work exposed to robots and LLMs together at 81%. That tells you where robots could work, not when they will arrive.

←TODAY: A September 2026 index finds robots can do 74% of US physical tasks but are cost-competitive on only 0.3% of work.
→3012: Zurich builds its sites like halls, because every drawing decided which tasks left the open air.
Fulcrum: The room sets the robot’s ceiling, and the room is set at the drawing desk.

The site starts at E3. The drawing decides how long it stays there.

A building site starts at the bottom of this scale: open, weathered, rearranged every day. The study’s own construction examples show both ends. Erecting scaffolding or ladders is an E0 task, because it needs “strength, balance, and mobility”. Drywall tapers score 1.6 on the exposure index, and their “tasks are split between unexposed and highly exposed.” One trade sits on both sides of the line.

That line moves with the detailing. A timber cassette assembled in a hall, a pre-cut and labelled kit, a tolerance scheme a factory can hold: each one pulls work from E3 into E1 or E2, where robots already earn their keep. Swiss timber prefabrication already works this way. Thomas Bock and Thomas Linner made the case for designing the building around the machine in Robot-Oriented Design, and this index puts a number on the room. Machine-ready output from Rhino + Grasshopper, carried into Archicad through the PAZ Grasshopper↔Archicad Library, is where an E3 task becomes an E1 task.

The trade-off is real. Every task moved into the hall buys robot-readiness with transport, crane slots and less freedom to change your mind on site. The data link runs both ways, too. A unit that scans a floor with lidar needs the model to know where it is, and it can write back what it found, so the BIM model stays alive after handover. Acceptance is part of the room as well. PAZ has covered the RAI Institute’s Spot study: a robot that passes every technical test can still be parked if the crew is not comfortable working next to it.

Atelier: For a 12-person studio in Zürich or a timber engineer in Bern, the useful question is not which trade disappears, because the study describes a slow, cost-limited path. The useful question is which of your details push work onto the open site by default. Monday move: in the Grasshopper definition for your next project, tag every part HALL or SITE and count which side your detailing loads.

Hack: Turn the price wall into years before anyone writes a site robot into a tender. Compound the study’s 3% annual price decline to see how long a robot takes to become 2×, 4× and 10× cheaper:

import math
r = 0.03  # annual robot price decline since the 1990s (Anthropic)
for f in (2, 4, 10):
    print(f, round(math.log(f) / -math.log(1 - r), 1))  # 22.8, 45.5, 75.6 years

This is plain compounding, not the task-level cost model behind Anthropic’s 40-year figure. It still works as a quick sanity check. A pitch that promises site robots at half the price soon is assuming a decline far steeper than the historical trend.

Open your last set of construction details today and mark every task they send out to the open slab. Then ask which of them a hall could take.

Sources & Further Reading

• Primary: Anthropic Research — What work can robots do?