Anthropic’s robot exposure index rates operating engineers at 1.6 out of 3 and electricians at 0.33

What the index measures

Robots can do 1.9% of physical work in places such as a construction site

Each physical task gets a tier by where a robot can do it. E0 means nowhere. E1 means a setting built around the machine, such as a fenced cell where fixtures hold each piece in a known position. E2 means a structured workplace built for people, such as a warehouse or a hospital. E3 means a place taken as found, and the rubric names “a construction site” as an example. An occupation’s index averages its task tiers, with E1 counting 1 and E3 counting 3, weighted by the share of work time each task takes.

Claude splits each task into typical instances, rates each one, and gives the task the highest tier that covers at least half the weight. For “Dig trenches,” Claude rated straight trenching in open ground, 25% of the task, at E3, citing the Built Robotics Exosystem kit that lets a hydraulic excavator dig on its own. Hand digging around buried pipes and cables, another 20%, is E0. More of the task looks like the careful digging, so the task is E0.

Robots do the least in E3 settings. Weighted by work time, 26.3% of physical work is E0, 49.8% is E1, 22.0% is E2 and 1.9% is E3. Counting every tier, the authors report that robots can do 74% of physical work somewhere, which is 34% of all work hours.

Where construction occupations land

Paving operators score 1.85 and electricians 0.33 on the same 0 to 3 scale

The data covers 49 construction occupations in SOC group 47, the federal code for construction and extraction jobs, leaving out extraction. Their median score is 0.88. Seventeen score 1.0 or higher, and eight score below 0.5. The top of the whole index is taxi drivers at 2.22.

Equipment and finish work lead. Paving, surfacing and tamping equipment operators score 1.85, and the authors cite an XCMG fleet of unmanned pavers and rollers. Operating engineers score 1.6, tapers 1.59 and brickmasons and blockmasons 1.42. The mechanical and electrical trades sit low, with plumbers and pipefitters at 0.51, HVAC mechanics at 0.46 and electricians at 0.33. Carpenters score 0.70. Cost estimators score 0.07 and construction managers 0.

Pile driver operators score 0.09, which looks odd next to the appendix’s mention of a pile-driving robot on solar farm sites. This post uses the occupation file, which gives one score per occupation without the task detail.

Inside one trade

Tapers spend 41% of their time on work robots can do on site and 26% on work no robot can do

A trade can score high while much of its work stays manual. Pressing paper tape into joint compound is E0, since mud and tape are messy materials for a robot. Applying the later coats is E3, citing the Canvas drywall finishing robot, which scans walls, sprays the coats and sands them. By work time, 41% of tapers’ tasks are E3 and 26% are E0. For carpenters, no robot can yet install windows, frames, flooring or trim.

Helpers and their trades

Helpers of electricians score 0.85, more than twice the 0.33 for electricians

Helpers score higher than the trades they work for. Helpers of carpenters score 1.28 against 0.70 for carpenters, helpers of electricians 0.85 against 0.33, and helpers of roofers 0.79 against 0.51. Those are the three helper categories that cover a single trade. One reading is that the robot-capable work on a crew sits with the entry-level positions, where firms train their next journeymen.

What the costs say

Robots doing a welder’s work would cost about five times the $66,100 of labor they replace

The authors estimate cost in a separate step. Claude estimated the full annual cost of the robots cited for each task, and the authors compared it with BLS total compensation for the share of the job robots can do. On that basis, robots cost less than people for 0.3% of all work.

Welders are the closest occupation to construction in the cost table. Robots can do 90% of their work by time, which at median total compensation of $73,400 is about $66,100 of labor a year. The robots would cost an estimated $334,600 a year, about five times as much, and the example system is a Zeman steel-beam assembler line at $2.7 million upfront.

Robot prices have fallen about 3% a year since the 1990s, in the data the authors use. At that rate, robots would be cheaper than labor for 0.8% of all work in about seven years and for 10% in about 40 years. The authors find regulation matters less for construction tasks than in healthcare or education.

Evidence so far

In the authors’ backtest, fully exposed occupations lost 34.2% of their jobs relative to unexposed ones

The evidence on the index comes from its authors at Anthropic, which makes the model that did the rating. They rated 1977 job tasks against 1977 robots and followed each occupation through 2024. Compared with an unexposed occupation in the same industry, one whose tasks robots could all do had wages 7.1% lower 20 years later and lost 34.2% of its jobs, with a 95% confidence interval of 16.4% to 52.0% on the job figure (the range the data support at that level of certainty).

Claude could count a robot that does a related task, such as the solar pile driver. Dropping those ratings cuts the exposed share of physical work from about three-quarters to about half. The appendix also notes that a bricklaying robot now spreads the mortar bed for block, a task no robot could do in 1977.

Availability and cost

The data covers 923 occupations and 18,796 tasks under a CC BY 4.0 license

The paper and data are free. The data is in Anthropic’s Economic Index dataset on Hugging Face under a CC BY 4.0 license, with the index for 923 occupations and ratings for 18,796 tasks, including Claude’s reasoning and the robots it cited.

This page’s data: occupations, tiers, tasks, tapers, helpers, costs, cost path (CSV files)

Caveats

The index records where robots can do the work as of 2026

The index counts a task as exposed when a robot can do it anywhere, as of 2026. Adoption on a given job depends on cost and site conditions. The counts in this post weight each occupation equally, and the cost figures are model estimates that the authors call suggestive.