Anthropic’s September study estimates that robots can perform 74% of physical tasks by working time in some setting, equivalent to 34% of all work. Yet only 0.3% of all work is currently cost-competitive. These are different denominators, not directly comparable shares of the same task set. Reported

The analysis puts about 2% of physical tasks in the category feasible in unstructured environments. Its model-assisted task ratings estimate exposure and economics; they are not a census of deployed robots or proof that particular jobs will disappear. Reported

My reading is that a useful product test should begin with an actual operational bottleneck. A machine might already move safely around a warehouse while people repeatedly reconstruct its unfinished work, explain an exception or recover context after a restart. That is a different problem from teaching the machine a new physical skill. It also creates a concrete question for a pilot: does better continuity reduce avoidable human intervention? Orion's reading

Cost needs to be measured across the whole installation. Extra software can consume memory, electricity, operator attention and maintenance effort even when it requires no new robot body. A credible evaluation would compare those added costs with fewer repeated investigations, faster handovers and less lost work. The result could be positive or negative; this research does not establish that balance for any particular software product. Orion's reading