On 30 September Anthropic published a piece of economics research by Russell Legate-Yang and Maxim Massenkoff with a plain question in the title: what work can robots do? Their answer is that robots, defined as autonomous physical machines that sense and act, can already perform 74% of the physical tasks in US jobs, which works out at 34% of all working hours.
The second half of the answer merits attention. Robots are cost-competitive with human labour for just 0.3% of job tasks. At the rate robot prices have fallen since the 1990s, which Anthropic puts at roughly 3% a year from industry and government data, it would take about 40 years for that share to reach 10%.
This comes at a time when the wider picture could easily suggest the opposite. In Curated Links on 28 September we pointed at Bedrock Robotics running excavators with nobody in the cab on live sites in Texas and Nevada, for contractors including Sundt Construction and Zachry Construction.
On 24 September the International Federation of Robotics (IFR) reported a record five million industrial robots operating in factories, with more than 600,000 installed in 2025 and China accounting for 59% of global deployments. Put those stories next to each other and the site-labour problem looks like it is solving itself.
Anthropic's numbers say the near-term picture is different. The technical capability for a lot of trades work exists, in some setting, somewhere. The economics mostly do not. For a contractor or PM, that changes the question from 'can a robot do it' to 'at what unit cost, on which repetitive task, with what site-data requirement'.
How Anthropic counted
The method is unusual and works as follows:
The data: The analysis uses O*NET, the US database covering around 900 occupations and 19,000 task descriptions.
Identifying physical work: Claude scored each task for its physical, cognitive and interpersonal requirements. This produced 7,594 tasks classified as physical.
Robot assessment: Claude then searched for specific robots that could perform each task and rated their capabilities on a four-tier scale:
E0: No robot can do the task.
E1: A purpose-built environment, such as an assembly line.
E2: A structured workplace, such as a warehouse.
E3: An unstructured environment, such as a public road.
What counts as evidence: Only demonstrated capabilities count. Sources had to show actual deployments, sales or demonstrations.
How exposure is measured: A task is considered exposed at the least structured level where robots can perform at least half of its real-world examples. The calculation is weighted by the amount of time spent on each example.
Construction examples:
Dig trenches: Claude rated cutting a linear trench in open ground at E3, citing a control system that retrofits hydraulic excavators. But careful digging around buried pipes and cables was rated E0. Because more of the task falls into the second category, the overall task is classed as unexposed.
Erect scaffolding or ladders: This is rated E0.
Drywall tapers: They score 1.6 on the 0 to 3 index. Around 41% of their time is spent on tasks rated E3 because of an autonomous drywall finishing robot. Robots cannot perform another 26% of their work, including pressing paper tape into wet compound.
The cost calculation: For each exposed task, Claude estimated the full cost of deploying the cited robots. This includes integration, maintenance, software, energy, supervision, insurance and decommissioning, annualised over the hardware's useful life.
Comparing robot and labour costs: The analysis compares that figure with total occupation compensation from the Bureau of Labor Statistics, adjusted for the share of time spent on the task.
A final caveat: Anthropic describes the results as "suggestive" rather than definitive. The analysis uses US data and ratings produced by a language model, so readers in the UK, Ireland and Australia should be cautious about applying the findings directly to their own markets.
The cost gap is the story
The worked example is packers and packagers. A set of robots costing over $2m to buy and install replaces the yearly work of around 14 workers; spread over a roughly 10-year life at an 8% cost of capital, with operating costs added, that comes to around $45,000 per worker replaced against labour cost of about $49,000. That is the cost-competitive 0.3%.
Welders are the opposite case: robots can weld, but automating the positioning, climbing, checking and grinding around the weld costs around five times more than the welders. Cleaners and dishwashers are paid $25,000 to $30,000 less than welders and their robot equivalents are still several times more expensive.
Anthropic's sensitivity test is the useful bit for anyone building a business case. If robots cost 20% less than today, they would be cost-competitive for the physical work of 2.8 million US workers, which is 0.8% of all working time. Reaching 10% needs a cost decline of about 70%.
Even in Anthropic's fast scenario, with quality-adjusted costs falling up to four times faster than history and capabilities improving twice as fast, robots become cost-competitive for half of physical work by 2050. The authors' own summary is that robots would need to sustain record rates of price declines and quality improvements for rapid physical automation.
Humanoids are not yet bending that curve. IFR figures reviewed by Reuters put humanoid sales at around 7,000 units worldwide last year for industrial and professional service use.
Susanne Bieller, General Secretary of the IFR, told Reuters that many were bought by research institutions or by companies generating training data rather than doing productive work, and that carmakers are piloting 'single-digit or sometimes double-digit numbers of robots in their plants'.
For sites, that means the economics of the next few years are set by retrofits and task-specific machines, not general-purpose labour.
The new mark of five million robots operational in factories worldwide is more than double the number seven years ago.
Heffner's point stands, but note where those five million are: factories, which is Anthropic's E1 environment. The doubling happened in the setting robots have always suited.
The office gets there first
The same paper gives the comparison that matters for delivery teams. Around half of all work is exposed to large language models alone; adding robots lifts that to 81%. Office and administrative support jobs, which mix computer work with light physical tasks, come out close to 100% exposed on the combined measure, whilst construction, repair and material-moving tasks are among those where Anthropic finds regulation is a smaller barrier than capability and cost.
The Stanford Digital Economy Lab working paper we covered on 29 September, which analysed 1.25 billion job postings and 154 million employment records across 41 countries, found that foreign affiliates of AI-adopting companies reduced the junior share of their workforce, mainly through growth in senior employment. Elsewhere this week we look at Antony Slumbers building a £10m asset-strategy tool in 78 minutes. The pressure is arriving at the estimator's desk and the planner's model before it reaches the bricklayer.
The aim is 'allowing our skilled crews to focus on the work where their experience matters most'.
For a contractor, that is the honest version of the pitch: fewer hours on the repetitive cut-and-fill, not fewer operators. Anthropic's footnotes add a note for non-US readers: a November 2025 survey of US executives found 13% of firms using robotics and 22% expecting to within three years, with expected use described as broadly similar in the UK, Germany and Australia.
Five questions for a robotics vendor
What is the all-in annual cost per unit of output, say per cubic metre moved or per square metre finished, including integration, supervision, maintenance, insurance and decommissioning, so it can be set against the loaded labour cost for that task alone?
Which specific versions of the task has the machine completed on a live site, and which versions still go back to a person, such as hand digging around services or the first tape coat on a joint?
What does the site need to look like for it to work, and what survey, model or control data must exist before it starts, and who is responsible for producing and maintaining that data?
What happens when the machine stops, who handles the exception, and is that human time included in the quoted rate?
What evidence is there that the price of this product line is falling, given that the business case for anything beyond today's 0.3% depends on it?
On the data question, one small pointer. Niantic Spatial released a Places Library on 25 September: 100 simulation-ready 3D environments for training robot policies, covering industrial, logistics, commercial and residential settings. Construction is on the request list, not in the catalogue. That tells you where the training data currently is.
Takeaway
Anthropic has given the industry a way to stop arguing about whether robots can do site work and start arguing about price, which is where procurement is good. The 74% figure is a capability inventory, rated by a language model from demonstrations and sales, and the 0.3% figure is a cost comparison with American wages.
Neither is a forecast, and the authors say so. The practical reading is that a contractor should treat autonomous plant the way it treats any other item of equipment: by task, by unit rate, with the data and supervision cost written in. Meanwhile the people most exposed in the next two to three years are the ones producing the drawings, the programme and the cost plan, and that is where the retraining budget belongs first.
If you want the unit-cost view of AI in construction every week rather than the demo-video view, the Project Flux newsletter at projectflux.ai is where we keep score.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
