Two reports released on September 30 put a difficult question at the center of the AI robotics debate: how can robot sales grow strongly while much of the work machines can technically perform remains uneconomic to automate? The answer starts with recognizing that a machine shipped, a task demonstrated and an hour of human work replaced are different units of progress.
Anthropic’s new research estimates that robots can perform tasks representing 34% of US working time under some conditions, but are cost-competitive for only 0.3% of working time. Separately, the International Federation of Robotics reports that worldwide professional service robot shipments rose 24% in 2025, reaching almost 250,000 units. Neither figure, by itself, measures the number of jobs eliminated.
A small share can support a growing market
TENS Magazine’s analysis centers on concentration: repeated purchases for a narrow range of economically attractive activities could support a growing robotics market without implying broad job replacement. The relevant questions are how widely useful applications spread and how much work changes at each adopting site. Sales growth alone cannot answer either question.
Anthropic uses Claude to assess occupational tasks, supporting evidence and operating conditions. Its categories distinguish purpose-built robotic environments, structured human workplaces and unstructured settings. The analysis also relies on estimated task time and deployment costs. These are model-assisted research estimates, rather than a census of machines replacing workers. A task’s classification therefore carries assumptions about where and how it is performed.
Consider a hypothetical warehouse deciding whether to automate movement between two fixed stations. Success on that route would establish something valuable but narrow. It would not automatically demonstrate an ability to unpack damaged deliveries, resolve inventory discrepancies or reorganize the shift. Those responsibilities would need their own evidence before anyone converted a transport demonstration into a claim about an entire role.
Why growing sales do not settle the cost question
The federation’s report places transportation and logistics first among professional service applications, with 117,500 units sold in 2025, or 47% of the market. It also reports about 7,000 full-size humanoids above 140 centimeters sold for commercial and professional applications beyond research, development and entertainment. It says many current humanoid applications remain specialized and often involve human teleoperation.
The two reports use different geographic and measurement boundaries: global unit shipments in the federation’s release and US work-time estimates in Anthropic’s research. They cannot be combined into a robot-per-job conversion rate. Nor does a high shipment growth rate demonstrate that a broad share of all occupations has become economical to automate.
TENS reads the apparent tension as a reason to examine concentration. A market can expand through repeated purchases for a limited set of suitable workflows. That would be compatible with a much smaller opportunity across the full range of work. This is an interpretation of how the measures can coexist, not a finding that every robot purchase meets a common profitability threshold.
The reverse mistake would be to use a small economy-wide estimate to dismiss a particular workplace’s experience. An aggregate figure cannot determine whether one site has the layout, volume and staffing arrangements for a useful deployment. The relevant comparison must return to the actual process and the same quantity and quality of completed work.
The arithmetic of output and working hours
A simple calculation shows why employment conclusions need their own evidence. Imagine a warehouse completing 1,000 deliveries using 100 staff hours in a defined period. Its output is ten deliveries per staff hour. These numbers are invented for illustration; they do not describe a company, a robot trial or either report.
If automation enables 1,200 comparable deliveries with the same 100 hours, deliveries per staff hour rise to twelve, a 20% increase. Output has grown while staff hours remain unchanged. If instead the warehouse completes the original 1,000 deliveries using 80 hours, the rate becomes 12.5, a 25% increase, while staff hours fall 20%. Neither calculation determines the number of people employed.
The example also shows why the accounting boundary matters. Suppose the second scenario requires 20 additional hours from a separate support team. Total staff time returns to 100 hours, and the apparent improvement disappears on that broader measure. A department can look more productive simply because supporting work is counted elsewhere. This is a hypothetical accounting effect, not an allegation about robot vendors.
TENS would therefore pair any claim of labor savings with comparable output, all associated staff hours and a description of where the remaining work goes. Quality must be held comparable too: more deliveries that need correction are not equivalent to more accepted deliveries. Even a valid reduction in hours would leave open whether employers reduce overtime, reassign people or change headcount.
The September reports are most informative when read as evidence about different scales of change. A growing market can be concentrated in suitable applications; a useful installation can change output without reducing employment. Following both the spread of adoption and the allocation of working hours would give readers a firmer basis for judging robotics than turning either headline into a universal forecast.
Image: NASA’s Valkyrie R5 research robot, photographed in 2013; an archival illustration, not a robot tested in these reports. Credit: NASA/Bill Stafford, James Blair, Regan Geeseman via Wikimedia Commons. Public domain, United States government work. Top-aligned landscape crop and resize; no generative edits.
