A robot picks up a parcel, turns neatly and places it on a shelf. The demonstration looks excellent. Somewhere outside the camera frame, an operations manager asks the less glamorous question: how much does each correctly placed parcel cost?

That question explains a gap in the robotics debate. Being able to perform a task and being an economical way to perform it are different achievements. Recent research and factory reports make the gap worth examining.

Research brief: This article examines sources available on 17 October 2026. It distinguishes modeled estimates, industry installation figures and a manufacturer’s demonstration. The rupee example below uses invented numbers to explain the calculation.

What the striking 0.3% estimate actually means

In research published on 30 September 2026, Anthropic used Claude and web evidence to assess 7,594 physical tasks in US occupational data. Its model estimated that existing robots could perform tasks covering 74% of physical working time, equivalent to 34% of all working time, under some circumstances.

After considering costs, the estimate fell to 0.3% of all working time. That is not a claim that 0.3% of jobs will disappear. The calculation accounts for expenses such as equipment, installation, maintenance, energy and supervision. It is an LLM-assisted model of US work, not a field audit of every employer or a global forecast. Capability under suitable conditions also does not mean a machine can take over an entire occupation.

The useful lesson is the distance between the capability estimate and the cost estimate. A convincing movement is one ingredient in a working business case.

Robots are already working at substantial scale

The gap should not be mistaken for an empty factory floor. The International Federation of Robotics’ September 2026 release reports five million industrial robots operating in factories worldwide in 2025, up 9% from the previous year, with more than 600,000 new installations. It also reports roughly 10,500 installations in India, up 15%.

These are industry figures for industrial robots. They are not a count of general-purpose humanoids or proof that every installation made a profit. They show substantial adoption in real manufacturing settings; the occupational model asks a different question about where additional tasks may be economical.

A robot does not need to make tea, climb stairs and hold a meeting to earn its place beside a production line. If the job is placing the same component accurately thousands of times, a specialized machine may be exactly what the process needs. Versatility is useful when the work requires it. Otherwise, it can be an expensive item on the brochure.

What a humanoid demonstration does—and does not—show

In a June 2026 announcement, Figure described its Figure 03 robot performing a logistics workflow at BMW’s Spartanburg plant: handling parts and repositioning itself while pulling a cart. The company also said its earlier Figure 02 deployment contributed to assembling 30,000 cars in 2025.

That is a useful concrete example, attributed to the manufacturer. The announcement does not provide the full costs and sustained performance data needed to calculate an independently verified return on investment. Handling variation in parts and movement is an engineering milestone; deciding what to pay for that capability requires another set of records.

For a buyer, the next questions concern ordinary shifts: how many acceptable items were completed, how often someone intervened, what happened after a fault and how long the equipment waited for work.

A simple cost example, in rupees

Imagine a small workshop considering an installed robotic system. The following figures are entirely hypothetical; they are neither supplier quotes nor estimates of typical market prices. We spread the purchase over five years purely to illustrate planning arithmetic.

The machine did not become less capable in the last row. It simply had fewer useful hours over which to spread the same assumed annual cost. This is why utilization deserves attention alongside the purchase price.

A real comparison needs more detail: financing, changes to the workspace, training, repair terms, residual value and the actual effect of idle time on running costs. The example deliberately holds annual costs constant to isolate utilization. It is not an accounting or investment model.

Even then, hourly cost alone can mislead. Compare equivalent acceptable output: a slower process producing fewer defects may beat a faster process with expensive rework. Likewise, automating one task does not necessarily eliminate a whole paid position. Someone may still prepare materials, handle exceptions or do the rest of the job.

Four things to ask before believing the business case

• Show the complete workflow. Include loading, changeovers, cleaning and recovery from failures, not just the most photogenic movement.
• Use representative work. Test the awkward package, the changed product and the busy period as well as the ideal sample.
• Count human help. Record interventions and support time. A demonstration with a nearby specialist is not automatically a hands-off production process.
• Separate the benefits. Capacity, quality, reduced exposure to hazards and labor savings deserve their own assumptions. Avoid counting the same improvement twice.

These are questions for evaluating evidence, not a verdict against automation. The same principle appears in our guide to AI-agent reliability: the interesting test is what happens across repeated real tasks, including the inconvenient ones. Our AI workday article looks at how people and tools can share a workflow.

Robotics progress is worth watching closely. The best question after “Can it do that?” is “Can it keep doing that, here, at a cost that makes sense?” The demonstration earns attention. The production log and the invoice decide what happens next.

Editorial disclosure: Written with AI assistance, with primary sources checked for dates and scope. The illustration is AI-generated. No robot was tested for this article, and the cost example is invented for explanation. See our editorial policy.