Physical AI + Robots : Three Quarters Capable, Nowhere Competitive.

Two Exposure Indices and the Dual-Speed Labour Shock — > Physical AI: Three-Quarters Capable, Nowhere Competitive.

Robots can already do three-quarters of America’s physical work. They are cheaper than the humans doing it for 0.3% of it. That gap, between what a machine can do and what it makes economic sense for a machine to do, is the most important number in the Physical AI debate, and almost nobody is talking about it.

I. The Second Index

On 30 September 2026, Anthropic’s economics team published What work can robots do?, by Russell Legate-Yang and Maxim Massenkoff. It is the physical twin of the March paper, Labor market impacts of AI: A new measure and early evidence, by Massenkoff and Peter McCrory.

Read together, the two papers give us something we have never had: a pair of exposure indices, one for language models and one for robots, built on the same task database by the same team. The March paper told us where the white-collar shock would land first. This one tells us where the blue-collar shock would land, and, more importantly, why it will not land for a very long time.

I have spent much of the past year arguing that we are mis-measuring AI’s economic footprint, that output is leaking into what I call Ghost GDP while the workforce quietly freezes in place. This paper is the strongest empirical support yet for the view that the physical economy runs on a completely different clock from the cognitive one. The labour shock is not one wave. It is two, travelling at very different speeds.

II. What the Paper Actually Measures

The method is elegant. Take O*NET, the US Labor Department’s database of roughly 900 occupations and 19,000 task statements. Have Claude separate physical tasks from cognitive and interpersonal ones. Then, for each of the 7,594 physical tasks, have Claude search for real robots already doing that work and rate the hardest environment in which they have been shown to do it.

The rubric is the clever part. It measures exposure not by whether a robot could do a task, but by how much the world has to be rebuilt around the robot for it to do so:

Tier	Where a robot can do the task	Share of physical work	Example
E0	Nowhere today	about 26%	Tinting hair, erecting scaffolding
E1	Purpose-built robotic cell	about 50%	Welding, assembling meal trays
E2	Structured human workplace	22%	Moving warehouse goods, delivering medication in hospitals
E3	Unstructured, open world	2%	Driving on city roads

Only demonstrated capabilities count, sourced from deployments, sales or demonstrations. The authors then estimate, task by task, what it would cost a robot to match a human’s annual output, including integration, maintenance, energy, insurance and supervision.

The headline findings:

• Capability: robots can perform 74% of physical tasks, equal to 34% of all US working hours.
• Combined reach: LLMs alone expose about half of all work; add robots and it rises to 81%. The unexposed fifth is highly interpersonal or needs physical skills robots lack.
• Economics: robots are cost-competitive for just 0.3% of job tasks, though that still covers around 300,000 workers for whom robots can do 95% of the job.
• Pace: at the historical 3% annual decline in robot prices, reaching 10% cost-competitiveness takes about 40 years. In an aggressive scenario, robots undercut humans on half of today’s physical work only by 2050.
• Validation: a 50-year backtest from 1977 shows that jobs more exposed to the robots of their day saw larger later declines in wages and employment. Robots have learned roughly 2% of previously impossible physical work each year.

The most exposed occupations are almost all vehicle operators, with taxi drivers top at an index of 2.2 out of 3. The largest occupation already facing cost-competitive robots is packers and packagers, some 560,000 US workers, where a robot line costs about $45,000 a year against $49,000 of human compensation. Nurses and general repair technicians sit at the other end.

III. The Mirror Image

Put the two Anthropic indices side by side and you get something close to a photographic negative. The people most exposed to language models are almost exactly the people least exposed to robots, and vice versa.

Dimension	LLM exposure (March 2026)	Robot exposure (September 2026)
Most exposed jobs	Computer programmers (75% task coverage), customer service representatives, data entry keyers (67%)	Taxi drivers (index 2.2 of 3), shuttle drivers, truck and tractor operators, packers, recycling sorters
Least exposed jobs	Cooks, bartenders, lifeguards, motorcycle mechanics (30% of workers at zero)	Nurses, general repair technicians, hairdressers, care workers
Typical exposed worker	Older, more often female, more educated, paid about 47% more	Male, 55 points less likely to hold a degree, paid about $30 an hour less
Labour market signal so far	No unemployment spike; job-finding for 22 to 25 year olds in exposed fields down roughly 14%	Exposed workers already face over twice the unemployment rate; packer employment down 22% since 2015
Binding constraint	Adoption and organisational change	Cost and manipulation dexterity
Clock speed	Software: near-zero marginal cost, diffuses in quarters	Hardware: capex, integration, diffuses in decades

Two observations follow. First, the much-repeated folk wisdom that “robots come for the blue-collar worker, AI comes for the white-collar worker” is now empirically grounded rather than anecdotal. Second, and more interesting, the two indices overlap in office and administrative support, which robots push to nearly 100% exposure because those roles combine screen work with light physical tasks: filing, mail, stock. That is the cohort facing a pincer from both sides.

The gap between the two clock speeds is the heart of the matter. A model update can reach a hundred million users in a weekend. A palletising robot needs a purchase order, a site survey, an integrator, a safety case and a ten-year depreciation schedule.

IV. Reading Across the Literature

The Anthropic robot index is the first major exposure study to price the machine, not merely rate it. Placed against the canon, it fills the one gap every previous index left open: the physical economy, measured against cost.

Study	Year	What it measures	Headline	Covers physical work?	Prices the machine?
Anthropic, What work can robots do?	Sep 2026	Demonstrated robot capability by environment, plus task-level cost	74% of physical tasks doable; 0.3% cost-competitive	Yes, exclusively	Yes
Anthropic, Labor market impacts of AI	Mar 2026	LLM capability weighted by real Claude usage	Programmers 75% covered; no unemployment spike yet	No	No
MIT and Oak Ridge, Iceberg Index	Nov 2025	Skills simulation of 151m workers	11.7% of US wage value (about $1.2tn) technically replaceable; visible tip only 2.2%	Partly	Partly
Yale Budget Lab and Brookings	Oct 2025	Change in occupational mix since ChatGPT	No discernible disruption after 33 months	Indirectly	No
Stanford Digital Economy Lab, Canaries in the Coal Mine?	Aug 2025	Payroll data on actual employment	13% relative fall for 22 to 25 year olds in most exposed jobs	No	No
ILO and NASK, refined global index	May 2025	GenAI exposure across 30,000 tasks, global	One in four workers exposed; 3.3% in highest tier; women twice as exposed	No	No
WEF, Future of Jobs 2025	Jan 2025	Survey of 1,000+ employers in 55 economies	170m jobs created, 92m displaced, net +78m by 2030; 58% of firms cite robots and autonomous systems as transformative	Yes, by employer expectation	No
IMF, Gen-AI and the Future of Work	Jan 2024	Occupational exposure and complementarity, 125 countries	40% of global jobs exposed; 60% in advanced economies	No	No
Eloundou et al., GPTs are GPTs	2023/24	Share of tasks an LLM could halve	About half of US work exposed	No	No
Acemoglu and Restrepo, Robots and Jobs	2020	Realised robot adoption, US commuting zones	Industrial robots measurably cut local employment and wages	Yes	Implicitly
Frey and Osborne	2013/17	Expert-rated automatability	47% of US jobs at high risk	Yes	No

Three patterns stand out.

Capability estimates keep rising, while realised impact keeps disappointing the doom-mongers. Frey and Osborne’s 47% became Eloundou’s half, which became Anthropic’s 81% once robots are added. Yet Yale finds no aggregate disruption and Anthropic’s own March paper finds no unemployment spike. The gap between “can” and “does” is the single most persistent finding in this literature.

The only place the evidence is unambiguous is at the entry level. Stanford’s payroll data and Anthropic’s job-finding rates both point at the same cohort: young people trying to get their first rung on the ladder in codified, screen-based work. This is my Frozen Workforce thesis in empirical form. Incumbents are protected by tacit knowledge; the door simply stops opening.

Cost is the variable everyone else left out. The IMF, ILO and Iceberg indices all measure exposure as a technical overlap. The Anthropic robot paper is the first to show, task by task, that exposure without economics is a misleading number. MIT’s own earlier computer vision work reached a similar conclusion for narrow vision tasks; Anthropic generalises it to the whole physical economy.

V. The Institutional Lens

The public institutions are converging on the same verdict as the lab: capability is racing, measured displacement is not, and the young are the exception. What they add is breadth, and in two cases, an explicit physical AI data point.

Stanford HAI, AI Index 2026. The ninth AI Index provides the most useful sanity check on physical AI I have seen this year. In embodied AI, robots succeed on real household tasks such as folding laundry or washing dishes only around 12% of the time (summary). That squares precisely with Anthropic’s finding that manipulation is the binding constraint for half of all physical tasks. The same Index finds generative AI reached 53% population adoption within three years, with Singapore at 61% against the United States at 28.3%. The cognitive wave is already everywhere; the physical one is still folding towels.

International AI Safety Report 2026 (chaired by Yoshua Bengio, Mila). The second edition, published on 3 February 2026 by more than 100 experts with a secretariat at the UK AI Security Institute, finds no significant effect on overall employment to date while flagging junior workers in exposed occupations such as software engineering and customer service. It treats labour disruption as a systemic risk and explicitly models scenarios to 2030 in which progress plateaus, holds steady or accelerates dramatically.

UK AI Security Institute. AISI’s Frontier AI Trends Report is not a jobs study, but it supplies the cognitive clock speed that the Anthropic robot paper contrasts against. Frontier models now complete hour-long software tasks with over 40% success, against under 5% in late 2023. Set that against robots learning about 2% of previously impossible physical work a year, and the asymmetry is stark.

World Economic Forum. The Future of Jobs Report 2025 projects 170 million roles created and 92 million displaced by 2030. Employers rank robots and autonomous systems as the second most transformative technology after AI itself, cited by 58% of firms. The Anthropic cost curve suggests that expectation is running well ahead of the economics.

Singapore. Our own data tells the augmentation story most clearly. The Ministry of Manpower found firms were roughly three times more likely to redesign jobs than cut headcount because of AI, with only about 6% reporting AI-driven headcount reductions. Adoption is deeply uneven: 27.2% of firms under 200 staff versus 76.4% of those above 500. MOM has also admitted it cannot yet isolate AI as a cause in retrenchment data, which is my Measurement Gap in an official sentence. Budget 2026 launched National AI Missions including Advanced Manufacturing, where physical AI will be tested first. As far as I can find, no Singapore body has yet published a robot exposure index of the Anthropic kind. It should.

The multilaterals. The ILO and IMF both emphasise transformation over replacement, and both note the exposure skew towards richer economies and, in the ILO’s case, towards women. Neither yet prices robots.

VI. My Usual Panel

As ever, I have put the paper in front of the voices I track most closely. I have not invented their words: each position below is drawn from their public record and linked, and I set it against what the robot index actually shows. The panel splits neatly into those the paper vindicates and those it challenges.

Vindicated

Geoffrey Hinton. His career advice on The Diary of a CEO was that machines will lag us at physical manipulation for a long time, so a good bet would be to be a plumber, while paralegals and call centre staff face early displacement. The Anthropic data backs him precisely: manipulation is the missing capability for half of all physical tasks, and repair work sits near the bottom of the exposure ranking. My one caveat is that his advice is a timing claim, not a permanence claim, and the paper’s 2% a year capability creep is the clock on it.

Erik Brynjolfsson. His Canaries paper found that the damage concentrates where AI automates rather than augments, and that experienced workers are protected by tacit knowledge that is never written down. The robot paper extends that insight into the physical world: tacit physical skill, untangling wires or digging around buried pipes, is exactly what keeps E0 tasks unexposed. Tacit knowledge is the moat on both sides of the collar line.

Fei-Fei Li. In From Words to Worlds she argues that language models are, in effect, blind to the physical world, and that spatial intelligence is the next frontier, with robotics a mid-term horizon rather than an imminent one. The paper agrees on timing. Where she would push back is on Anthropic’s finding that planning and reasoning limit only 8% of physical tasks: world models are as much about perception as reasoning, and perception is one of the four bottlenecks the paper names.

Yoshua Bengio. The International AI Safety Report he chairs treats labour disruption as a systemic risk while finding no significant effect on overall employment to date. Bengio has also noted that agents still lack the long-horizon planning needed to eradicate jobs wholesale. The robot index is precisely the kind of evidence-first instrument his report calls for.

Challenged

Dario Amodei. His warning that AI could disrupt half of entry-level white-collar work was always about the cognitive economy, and nothing here contradicts it. What is striking is that his own company’s economists now publish the most sober evidence in the field: a physical labour market that, on cost, barely moves for decades, and a colleague, Peter McCrory, who wrote in July that he does not expect AI to push unemployment noticeably higher within a year. That internal plurality is healthy. It is also a reminder that a chief executive’s warning and a research team’s measurement answer different questions.

Elon Musk. Musk has said 80% of Tesla’s value will be Optimus, targeting a million units a year and projecting that humanoids will outnumber humans by 2040. The Anthropic cost curve is the most rigorous counterweight to that thesis I have seen. Even its fast scenario, with costs falling four times faster than history, reaches cost parity on half of physical work only in 2050. Musk’s bet requires humanoids to break the historical cost curve by an order of magnitude. Possible, but it is a bet on manufacturing economics, not on intelligence.

Mo Gawdat. Gawdat calls the claim that AI will create new jobs 100% crap, pointing to a start-up built by three people that would once have needed 350. For software, his arithmetic is plausible. For atoms, the robot paper shows it does not transfer: you cannot replicate a palletiser the way you replicate a model, and hardware does not enjoy software’s zero marginal cost. Gawdat’s world arrives, if it arrives, at two very different speeds.

Where the panel converges

Strip away the temperament and every voice agrees on three things. The entry level is where the pain lands first. Tacit, embodied skill is the last moat. And nobody, optimist or pessimist, has a reliable model of how fast manipulation will improve. That last gap is where the next decade of argument will be fought.

VII. The Cost Curve Is the Story

The most important sentence in the paper is the one most commentators skipped: robots would need to sustain record rates of price decline and quality improvement for decades to enable rapid physical automation. Capability is no longer the headline constraint. Unit economics is.

Milestone (share of today’s work)	At historical pace (3% a year cost decline)	Fast scenario (4x cost decline, 2x capability gain)
Cost-competitive today	0.3% of all tasks, about 300,000 US workers	Same
20% cheaper robots	About 7 years; 2.8m workers spend 42% of their time on newly economic tasks	Sooner
10% of all work cost-competitive	About 40 years (needs a 70% cost fall)	Faster
Half of physical work cost-competitive	2085	2050

This maps directly onto three ideas I keep returning to.

Ghost GDP. I have argued that AI’s productivity is leaking out of the national accounts: value created in software that never shows up as measured output. The physical economy has the opposite problem. Its gains are visible, capital-heavy and slow, which means the macroeconomic story of the next decade will be dominated by cognitive automation, with the physical economy acting as what the paper’s cited growth literature calls a weak link. Jones and Tonetti’s point, which Anthropic cites, is that automation boosts productivity only once machines are much cheaper than humans, not merely at parity.

The Frozen Workforce. The cognitive shock freezes the entry level: graduates who never get hired. The physical shock, when it comes, will look different, because robot-exposed workers already earn about $30 an hour less and face more than twice the unemployment rate. There is no cushion. A 22% fall in packer employment since 2015 is the early warning, and the BLS projects packers among the largest job losses to 2035.

The Fork and Long-AND, not Short-OR. The Star Trek path and the Mad Max path diverge not on what robots can do but on who captures the cost decline. If a two-speed transition lets societies retrain physical workers over decades while cognitive gains fund the transition, we get abundance. If the slow physical curve lulls policymakers into inaction while the fast cognitive curve hollows out the middle, we get the other film. The right posture is long on both robots and people, not a short bet against either.

VIII. The Asia Pacific Reading

The Anthropic paper is built on US tasks, US wages and US robot prices. Read from Singapore, the most important variable changes sign: the cost-competitiveness threshold depends on local wages, so the same robot crosses it years earlier in Seoul or Singapore than in Jakarta or Ho Chi Minh City.

The region is already the world’s robot laboratory. According to the IFR’s World Robotics 2025, Korea runs 1,220 industrial robots per 10,000 manufacturing employees, the world’s highest, and Singapore is second at 818, growing 13% a year since 2019. China installed 54% of all robots worldwide in 2024, some 295,000 units, even though its density reads only 166 after a revision to its labour statistics. The paper itself notes that robots in Japanese convenience stores already restock fridges, a task US retail robots cannot yet do.

Three implications follow for our part of the world.

1. High-wage, ageing economies hit cost parity first. Singapore, Japan, Korea, Taiwan and Australia combine high labour costs, shrinking workforces and, in Singapore’s case, a foreign-worker levy that raises the effective human cost. The 0.3% US figure understates near-term physical automation in these markets.
2. Low-wage manufacturing economies face a different risk. For Vietnam, Indonesia or Bangladesh, the threat is less that robots replace local workers and more that cheap robots in high-wage markets make reshoring viable. Exposure becomes a trade question, not a labour one.
3. Preferences differ. The paper finds human preference blocks robots in a quarter of physical tasks. Cultural acceptance of service robots in Japan, Korea and China is markedly higher than in the US, which shortens that barrier here.

Singapore sits at an interesting junction. It has the world’s second-densest factory robot base and, per the Stanford AI Index, generative AI adoption of 61%, far above the United States. Yet MOM’s data shows firms redesigning jobs far more often than cutting them, and an AI adoption gap of almost 50 points between large firms and SMEs. Our two-speed problem is not only cognitive versus physical. It is also large firm versus small.

IX. What Leaders Should Do

Exposure indices are only useful if they change decisions. My test for any recommendation is the enforceability test: can someone be held to it, with a number, by a date? Here is what passes.

For enterprises and operations leaders

• Run your own E0 to E3 audit. Take your top 20 physical roles, list their tasks, and rate each by the environment a robot needs. Most firms will find their exposure sits in E1, meaning automation requires redesigning the workplace, not buying a robot.
• Model cost parity, not capability. Use the paper’s method: robot annualised cost against fully loaded labour cost scaled by time on exposed tasks. In Singapore, include levies and the cost of not being able to hire at all.
• Start where Anthropic says the foothold already is: packing, palletising, intralogistics, sorting and inspection. These are the tasks where cost parity is closest and the paper’s backtest shows impact arrives first.
• Protect the entry pipeline in cognitive roles. The physical shock is slow; the white-collar entry-level shock is already measurable. Ring-fence graduate intake and redesign junior roles around supervising AI rather than competing with it.

For supply chain and IBP planners

• Treat automation as a capacity variable with a cost curve, not a binary. A 3% annual decline in robot cost is a planning assumption you can put in a scenario, alongside a fast case.
• Plan the human residual. The paper’s cost estimates already include supervision, exception handling and repair. Those are the jobs your network will need more of, not fewer.

For policymakers

• Publish a national robot exposure index. The method is open, the data release covers 923 occupations, and adapting it to local wages and task structures is a modest exercise for a ministry of manpower or statistics office.
• Fix the measurement gap. MOM’s own admission that AI cannot yet be isolated in retrenchment data should become a statutory reporting line: firms over a size threshold report headcount changes attributable to automation.
• Fund the transition from the fast curve. Cognitive AI productivity gains arrive first and accrue to capital-light firms. Some of that surplus should finance retraining for the physical workers whose turn comes later, before it comes.

X. Caveats and Open Questions

I admire this paper, which is exactly why it deserves an adversarial read.

• An AI rating AI. Claude classifies tasks, searches for robots and estimates costs. The authors validate against 50 years of history and release every rating with its reasoning, but model judgement is still doing a great deal of work. Independent replication with human raters would strengthen it.
• A lab measuring its own field. The publisher builds frontier models. To Anthropic’s credit, the conclusions are notably cautious, which is the opposite of what a commercial incentive would predict, but readers should hold the question in mind.
• United States only. Wages, robot prices, regulation and preferences all differ across Asia Pacific, as Section VIII argues. The 0.3% figure is not portable.
• Demonstrated capability only. By design, the index cannot see a leap. If humanoids or foundation models for robotics crack dexterous manipulation, exposure could jump faster than the 2% annual creep implies. The authors acknowledge this.
• Partial equilibrium. Costs are held against today’s wages. In reality, wages fall as automation rises, which would delay cost parity further, while mass production of robots could accelerate it.
• What it leaves out. AI can transform physical work without any robot, through predictive maintenance, scheduling or computer vision on existing lines. Those effects sit outside this index entirely.

The open question I would most like answered next is a joint one: what happens to occupations like office and administrative support, where LLMs and robots together push exposure to nearly 100%? That pincer cohort is where I would place the next study, and the next policy pilot.

Luke Soon is an AI futurist, ethicist and author of Genesis: Human Experience in the Age of Artificial Intelligence. He writes at GenesisHumanExperience.com.

Sources

• Legate-Yang, R. and Massenkoff, M., What work can robots do?, Anthropic, 30 September 2026
• Massenkoff, M. and McCrory, P., Labor market impacts of AI: A new measure and early evidence, Anthropic, 5 March 2026
• Brynjolfsson, E., Chandar, B. and Chen, R., Canaries in the Coal Mine?, Stanford Digital Economy Lab, August 2025
• Gimbel, M. et al., Evaluating the Impact of AI on the Labor Market, Yale Budget Lab and Brookings, October 2025
• MIT and Oak Ridge National Laboratory, Iceberg Index, coverage in Fortune, November 2025
• ILO and NASK, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, May 2025; gender findings
• IMF, Gen-AI: Artificial Intelligence and the Future of Work, SDN/2024/001, January 2024
• World Economic Forum, Future of Jobs Report 2025, summary data and technology drivers
• Stanford HAI, AI Index Report 2026, April 2026; takeaways summary
• International AI Safety Report 2026, summary and Australian Government page
• UK AI Security Institute, Frontier AI Trends Report, December 2025
• IFR, World Robotics 2025 robot density
• Singapore Ministry of Manpower, Labour Market Report 1Q 2026 coverage, Response to Motion on AI, retrenchment data review
• Ministry of Trade and Industry, Oral reply on AI capex and jobs, 9 September 2026
• Panel sources: Hinton, Fei-Fei Li, Bengio, Amodei, McCrory, Musk, Gawdat
• Classic references cited via the Anthropic paper: Eloundou et al. (2024), Acemoglu and Restrepo (2020), Frey and Osborne (2017), Jones and Tonetti (2026)