Only the whole job is for sale

"Taxi drivers lead exposure with an index of 2.2." That is the most robot-exposed occupation in the United States, according to an index Anthropic published on 30 September, and the sentence that explains why a person still drove you to the airport last week sits further down the same page: a robotaxi is "estimated to cost only around $7,000 more than taxi drivers, but face regulatory hurdles." Not ten times more. Seven thousand dollars a year, and a permit.

The study behind those sentences scores what today's robots can do across around 900 occupations and around 19,000 task statements, and it arrives at two figures that look like they belong to different papers. "Counting any exposure level, 74% of physical work, or 34% of all work, can be done by robots in some circumstances." And: "Robots are cost-competitive for just 0.3% of job tasks."

The reflex is to assume the capability figure is inflated and the cost figure is the real one. I think the gap is telling you something more specific, and more useful. The 0.3% is not a measure of what robots cannot do. It is a measure of how rarely a whole person's day is made of the one thing a machine can be bought to do. Capability is counted task by task because that is how work is described. Money is spent on an installed machine, in a building, against a shift. The two are not denominated in the same unit, and the 0.3% is the exchange rate.

How the index is built

The input is O*NET, "a database of around 900 occupations linked with descriptions of around 19,000 job tasks". Each task is a sentence of prose. To score it, the authors "instruct Claude to search for specific robots relevant to each task and assess their capabilities and operating environments, quoting sources directly". So the capability judgment is not a benchmark result; it is a literature search over existing commercial and research robots, resolved into a rating.

What makes the rating worth reading is the second axis. A task is not scored as doable or not, but by the kind of environment a robot needs in order to do it. E0 is a task no robot can do. E1 needs "a purpose-built robotic work environment, like a factory assembly line". E2 works "in a structured human work facility, like a logistics warehouse". E3 works "in an unstructured environment, like a city road".

Now the headline decomposes. "As a share of all tasks, around 12% cannot be done by any robot today (E0), 23% can be done by robots in specially built environments (E1), 10% in structured human workplaces (E2), and 1% in unstructured environments (E3)." Those four add to 46%, which is how much of American working time is physical at all; the 34% is the last three added together, and the 74% is 34 divided by 46.

Read it that way and two thirds of the capability claim turns out to be a claim about buildings. Twenty-three of the thirty-four points require an environment constructed around the machine. Only one point in a hundred of all working time is a task a robot can do in the world as it already is. That ordering is not an accident of this study. It is the standing shape of robot learning, where a policy is trained against a specific body in a specific setting and does not carry: the field's attempt to pool data across machines, Open X-Embodiment, exists to "provide all open-sourced robotic data in the same unified format, for easy downstream consumption", and the LeRobot library's hardware list names thirteen distinct platforms before it reaches third-party arms. Generality across environments is the open problem, not a detail of deployment.

The cost side is ordinary capital budgeting, done carefully. For packers and packagers the authors count several machines, for packing containers, moving materials, inspecting goods, erecting boxes, labelling and sealing. "These robots are estimated to cost over $2 million to purchase and install, but replace the yearly work of around 14 workers. Fixed costs are spread over a roughly 10-year service life at an 8% cost of capital. Operating expenses, such as maintenance, part-time human supervision, and energy, bring annual costs to around $45,000 to replace one human worker per year." That annual figure is then set against the occupation's compensation scaled by the share of time spent on the task, and the task counts as exposed when the robot is cheaper.

Where the arithmetic actually closes

The 0.3% is stated as a share of job tasks and the 34% as a share of working time, so I would not lean on the ratio between them. The worker counts are unambiguous and they make the point better. Of the cost-competitive set, the authors note: "that includes roughly 300,000 workers for whom robots can do 95% of their tasks."

Ninety-five percent. That is the condition under which the sums work. Packers and packagers are the worked example, and the reason they qualify is given in one clause: "Since packers and packagers spend 97% of their time on tasks that robots can do and cost around $49,000, robots cost about $2,500 less per year to do that work." A margin of two and a half thousand dollars, on a job that is 97% one thing.

This is what indivisibility does. Suppose a robot could do a clean 40% of a nurse's shift. You cannot buy 40% of a two-million-dollar cell, and more importantly you cannot stop paying 40% of a nurse. The remaining 60% still requires a person present, scheduled, and paid for a shift, so the automated fraction yields no headcount and the saving has nowhere to land. The task-level comparison assumes the fraction you automate is a fraction you can capture. Labour is not sold in fractions. It is sold in people.

Set that against the other half of the same sentence in the study: "Overall, about 80% of job tasks by working time are exposed to either robots or LLMs." The word "exposed" is doing two different jobs there. For a language model, the unit of purchase is a token. The Usage object the API returns on every request carries input_tokens and output_tokens, and the docstring is explicit that output_tokens "remains the inclusive, authoritative total used for billing"; at the published Haiku 4.5 rate of $1 per million input tokens, the smallest purchasable quantum of model work costs a millionth of a dollar. You can buy 3% of a task. You can buy it for one request, once, and stop. Nobody sells 3% of a robot, and nobody sells 3% of a warehouse built for one.

So exposure, for software, is close to a purchase order. Exposure, for machines, is a wish list conditional on capital, on a building, and on the rest of the bundle the worker is still holding.

The number that moves

The useful consequence is that this makes displacement a threshold, not an average, and thresholds behave badly. The study gives the sensitivity directly: "if robots cost 20% less today, they would be cost-competitive for the physical work done by 2.8 million workers." Three hundred thousand to 2.8 million. A ninefold change in the exposed workforce from a 1.25-fold change in price.

Two honest qualifications. The 300,000 figure carries the 95%-coverage condition and the 2.8 million figure is not stated with it, so part of that jump is a difference in framing rather than a pure price response. And the full paper sits on a host this environment could not reach, so every number here comes from the authors' own summary of their work rather than from the appendix that would settle it. But the mechanism is not in doubt, because it is the same mechanism as the packers: when the margin is $2,500 on a $49,000 job, small movements in price move whole occupations across the line at once.

That is the strongest case against reading the 0.3% as reassurance. It is also the strongest case against reading it as a countdown, which brings me to the objection I take most seriously.

The authors are not over-claiming. Their own forward number is slow: "For robots to be cost-competitive for 10% of human work today, costs would need to decline about 70%", and at the historical 3% a year "that would take around 40 years", with "automating 90% of physical work today still takes 53 years". A reader could fairly say the per-task accounting already handles fractional automation, since a robot that covers 30% of a task mix is charged against 30% of the pay; and that services fix the indivisibility anyway, because a shared cell or an hourly rental turns capital into an operating expense you can buy by the hour.

The accounting is right. The claim I am making is about what the number will be used for. Robots-as-a-service divides the capital; it does not divide the shift, and it does not divide the purpose-built environment that twenty-three of the thirty-four points depend on. The 34% will be quoted as a statement about capability, because it reads like one. It is a statement about bundles, and the bundle is doing most of the protecting.

For anyone learning to read these studies, the transferable move is small and worth keeping. When you meet an exposure figure, ask two questions before you ask whether the capability is real. What is the smallest unit the buyer can purchase, and what is the smallest unit the incumbent can stop paying for? Where those two match, as they do for tokens, exposure converts into use almost immediately. Where they do not, exposure accumulates quietly for years and then arrives in one occupation at a time. And ask for the distribution over workers rather than the mean over tasks, because the quantity that predicts a layoff is not how many tasks are covered on average but how many people are 95% covered.

The last detail in the study is the one I cannot stop thinking about. "The US currently employs around 560,000 packers and packagers." And: "Employment for packers and packagers has fallen 22% since 2015." The occupation sitting closest to the threshold, the one where the robot is finally $2,500 cheaper, has already lost more than a fifth of its people for reasons that have nothing to do with any of this. By the time the machine is unambiguously the cheaper option, a good deal of the work it was priced against may no longer be there to take.