Will Robots Take Blue-Collar Jobs? What Anthropic's 2026 Study Really Says (Anthropic, Tesla, Nvidia, Jobbit)

For a decade the public conversation about automation has run on a single number. In 2013 Carl Benedikt Frey and Michael Osborne at Oxford estimated that 47% of United States employment sat in occupations at high risk of computerisation within two decades, and the figure has been repeated in every speech about the future of work since. Thirteen years on, the robots that were supposed to empty the warehouses and building sites have arrived in demonstrations, in a few thousand factories and in a small number of taxis, and the people who were supposed to be replaced are still at work, many of them in shortage occupations that employers cannot fill. Something in the standard story was wrong, and until recently nobody had measured what.

On 30 September 2026 Anthropic's economics team published "What work can robots do?", by Russell Legate-Yang and Maxim Massenkoff. It is, as far as this review can establish, the first attempt to rate every physical task in the American economy against robots that actually exist, and then to ask the question the 47% never did: at what price. The answer reshapes the debate. Robots can already perform about three-quarters of physical work tasks in the United States, but they are cheaper than a human for only three in a thousand. Closing that gap at the pace robot prices have fallen since the 1990s would take about forty years to reach a tenth of work, and even an aggressive scenario puts half of physical work beyond 2050. This case study reads the paper closely, sets it beside the evidence on what AI is already doing to desk work, considers the objections, and draws out what the pattern means for a British tradesperson, a small-business owner and a company such as Jobbit that was built on the premise the paper now quantifies: the handover between machines and people is the central design question of the next twenty years.

Summary of findings

First, capability is no longer the binding constraint on most physical work. The Anthropic study finds that robots can perform 74% of physical tasks in the United States, which amounts to 34% of all working hours, using robots that have been sold, deployed or demonstrated rather than imagined.

Second, cost is. Robots are cost-competitive with a worker for 0.3% of job tasks today. For the most exposed large occupation, packers and packagers, the estimated annual saving is about $2,500 per worker against labour costs of about $49,000; for taxi drivers, the robot still costs around $7,000 more than the human.

Third, the gap closes slowly. Robot prices have fallen about 3% a year since the 1990s. At that rate it takes 40 years for robots to become cost-competitive for 10% of US work. In the paper's fast scenario, with costs falling four times faster and capabilities improving twice as fast, robots become cheaper than people for half of physical work around 2050, against 2085 on the historical trend, and automating 90% of it still takes 53 years.

Fourth, digital work faces the opposite picture. About 50% of US work is exposed to large language models alone and 81% to language models or robots combined, and the early labour-market evidence from payroll data shows employment among 22 to 25 year olds in the most AI-exposed occupations running about 19% below their less-exposed peers, with no equivalent movement yet for experienced workers or for the economy as a whole.

Fifth, the people most exposed to robots are the ones with the least cushion. Compared with unexposed workers they are 55 percentage points less likely to hold a degree, earn about $30 an hour less and have an unemployment rate more than twice as high.

Sixth, the barriers that remain are specific and measurable. About 70% of physical tasks are held back by capability, roughly half of them by manipulation and dexterity; 14% are blocked by current regulation and 25% by human preferences. Those are the tasks where a person will be needed for decades, and where the data to teach machines does not yet exist.

The case: why this study is the right laboratory

Automation forecasts have usually failed in one of two ways. Occupation-level studies such as Frey and Osborne's treated whole jobs as automatable units, which produced the 47% figure; when Melanie Arntz, Terry Gregory and Ulrich Zierahn at the OECD repeated the exercise in 2016 at the level of tasks, accounting for the mix of work inside each occupation across 21 countries, the share of automatable jobs fell to 9%. Expert-survey studies, meanwhile, asked specialists what machines would be able to do, which is a bet on technologies that may never ship.

The Anthropic paper avoids both traps by design. It starts from the roughly 19,000 task statements for about 900 occupations in O*NET, the US Department of Labor's occupational database, and from Bureau of Labor Statistics employment and wages. It uses the Claude model to score each task on its physical, cognitive and interpersonal content, keeps only the physical tasks that could not be automated by ordinary software, and then searches the web for specific robots that have demonstrably performed each task, requiring commercial deployment, sales or a significant demonstration as evidence. Each task is placed on a four-tier scale of environmental control: E0, no robot can do it; E1, a robot can do it in a purpose-built environment such as a factory line; E2, in a structured human facility such as a warehouse; E3, in an unstructured environment such as a city road. Most deployed robots live in engineered spaces, which is why the authors treat the tier as a good measure of near-term risk.

Cost is modelled rather than assumed. For each exposed task the model estimates the annual cost of the cited robot, including purchase and installation, maintenance, software, energy, oversight, insurance and decommissioning, with fixed costs annualised at an estimated cost of capital, and compares it with the worker's total compensation scaled by the share of their time spent on that task. A task counts as cost-competitive only when the robot's annual cost falls below the labour it replaces. The authors then validate the whole apparatus backwards: rating tasks against the robots that existed in 1977 and later decades, they show that occupations more exposed to the robots of their day went on to lose more wages and employment, which is what an exposure measure must do to be worth anything.

The study is American. That is the imperfection the reader should hold onto. British wages, energy prices, regulation and the mix of occupations differ, and nobody has yet run the method on UK data. But the stakes here are if anything higher. The Construction Industry Training Board's June 2026 forecast says the sector needs an average of 41,200 additional workers a year between 2026 and 2030, more than 206,000 in total, to keep pace with demand, with 40% of the existing workforce over 45 and more than 200,000 EU workers lost since Brexit. The Department for Education's Occupations in Demand release of December 2025 counts 10.9 million people, 32.9% of the workforce, in 125 elevated-demand occupations, and names electricians and electrical fitters (201,000 jobs) and vehicle technicians and mechanics (183,000) among the largest skilled trades on that list. If robots were about to replace the trades, the shortage would be a passing problem. The evidence below says it is not.

A timeline of the evidence, 2013 to 2026

Finding 1: robots can do most physical work and almost none of it pays

The first number in the paper is the one that will be quoted, and it deserves to be stated precisely. Of all US work time, 54% is cognitive or interpersonal and 46% is physical. Within the physical share, 12 percentage points cannot be done by any robot the authors could find (E0), 23 can be done in purpose-built environments (E1), 10 in structured human facilities such as warehouses (E2), and 1 in unstructured environments (E3). Put together, robots can perform 74% of physical tasks, which is 34% of all working hours. The examples are concrete: autonomous taxis on city roads, autonomous mobile robots moving stock in warehouses, AI-guided welding cells in fabrication shops, trench excavators that dig straight lines without a driver.

The second number is the one that matters. When the model prices the cited robot for each task and compares it with the labour it would replace, robots come out cheaper for just 0.3% of job tasks. Roughly 300,000 American workers are in occupations where robots could do 95% of the work; the largest group, about 560,000 packers and packagers, is the clearest case of cost parity, with the robot estimated at about $2,500 a year less than a human whose labour costs about $49,000. At the other end, the robot taxi, technically the most exposed occupation in the data with an exposure index of 2.2 out of 3, is estimated to cost around $7,000 a year more than the driver it would replace, before regulation is considered. Nine of the ten most exposed occupations are vehicle operators.

The interpretation the authors offer, and the one this review finds most defensible, is that the physical frontier has moved much further than the economic one. Over the past 50 years robots became able to do about 2% of physical work each year that they could not do before: in 1977 they could not perform 62% of physical tasks, today they cannot perform 24% of those same tasks. Capability has compounded. Cost has not followed, because a robot that can carry a box is still a capital asset with maintenance, energy, oversight and insurance attached, competing against a worker who is paid only for the hours worked.

Finding 2: the timeline is decades, even in the optimistic case

The paper's most consequential contribution is the scenario analysis, because it turns "robots are expensive" into dates. Robot prices, drawing on Bank of Japan series and other sources, have declined roughly 3% a year since the 1990s. Hold that rate and it takes about 40 years before robots are cost-competitive for 10% of US work. A 20% fall in costs, about seven years at the historical rate, would make robots competitive for the physical tasks of about 2.8 million workers; a 70% fall is needed to reach 10% of all work.

The fast scenario is generous by design. Quality-adjusted costs fall up to four times faster than history, and robots learn new tasks twice as quickly. Even then, robots become cost-competitive for 50% of physical work only around 2050, against 2085 on the historical trend, and automating 90% of physical work takes 53 years. The sentence the authors use is worth quoting because it is unusually plain for an economics paper: "Robots would need to sustain record rates of price declines and quality improvements over the coming decades to enable rapid physical automation."

Here the evidence should be separated from the interpretation. The dates are not forecasts; they are the consequences of assumptions applied uniformly to every task, which the authors say themselves may not reflect how robotics companies actually prioritise. A firm that pours its effort into one high-value task, warehouse picking, say, could drive that task's cost down far faster than 3% a year while the rest of the economy stands still. What the scenarios establish is the shape of the problem, not its exact calendar: physical automation is a long, uneven process governed by hardware economics, in contrast to software, where the marginal cost of one more task is close to zero.

The comparison with past technology supports that reading. Acemoglu and Restrepo's 2020 study of industrial robots, the previous generation of this technology, found that each additional robot per 1,000 workers between 1990 and 2007 reduced the employment-to-population ratio in the affected American commuting zones by 0.2 percentage points and wages by 0.42%. Real, measurable, and slow: the effects accumulated over seventeen years in the sectors where robots were already cheapest, and nowhere else.

Finding 3: digital work is the opposite story

The same research group measures the other frontier, and the contrast is the heart of this case study. Anthropic's earlier exposure work and the paper's own comparison put about 50% of US work within reach of large language models alone, and 81% within reach of language models or robots combined. The pattern by occupation is instructive. Transportation and material moving is under 15% exposed to language models but about 90% exposed once robots are included; office and administrative support is already highly exposed to language models and reaches almost 100% with robots; personal care and service work sits at roughly 40% under both, because it is interpersonal in a way neither technology addresses.

The language-model estimates have a longer pedigree. In 2023 Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock at OpenAI and the University of Pennsylvania found that 80% of American workers were in occupations with at least 10% of tasks exposed to language models and 19% in occupations with more than half their tasks exposed, with writing, programming and data analysis most exposed and physical and outdoor work least. The Anthropic Economic Index, which maps anonymised Claude conversations to O*NET tasks, shows what people actually do with the tools: in the 2026 reports augmentation ran slightly ahead of automation, 52% to 45%, and use diversified, with the ten most common tasks falling from 24% of conversations to 19% between November 2025 and February 2026. The Index also reports that a 1% rise in GDP per capita is associated with a 0.7% rise in usage per person, which is a reminder that adoption follows wealth.

Whether exposure has become displacement is the question the payroll data answers, cautiously. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford's Digital Economy Lab, using ADP records for more than 25 million US workers, reported in 2025 that early-career workers in the most AI-exposed occupations had seen relative employment declines while experienced workers in the same jobs had not; their August 2026 revision, with data through the middle of the year, puts employment among 22 to 25 year olds in those occupations about 19% below where it would be had it tracked less-exposed peers, with the adjustment running through hiring rather than pay and concentrated where AI automates rather than augments. At the same time, Anthropic's own economists Maxim Massenkoff and Peter McCrory, as Yahoo Finance reported on 1 October, find no systematic increase in unemployment for highly exposed workers since late 2022, and Anthropic's chief executive Dario Amodei's warning that AI could remove half of entry-level white-collar jobs remains, on the data so far, a warning rather than an observation.

The honest synthesis is this. The first measurable labour-market effect of the current wave of AI is not the robot in the warehouse; it is the missing graduate hire in the office. Physical work is protected by economics; desk work is not.

Finding 4: who is exposed, and what still blocks the rest

The demographic profile of robot-exposed work is the finding that should shape policy. Compared with workers in unexposed occupations, those in highly exposed ones are 20 percentage points less likely to be women, 16 points more likely to be Hispanic, 55 points less likely to hold a bachelor's degree, paid about $30 an hour less, and more than twice as likely to be unemployed. They are also far more likely to carry weight, work in extreme heat or near hazardous substances. The tasks robots can already do are disproportionately the hard, dangerous, poorly paid ones, which is both the strongest argument for automating them and the reason it will hurt if the transition is unmanaged.

What blocks the rest is equally specific. About 70% of physical tasks that robots cannot yet do are limited by capability, and roughly half of those by manipulation and dexterity: hair colouring, hand-digging around buried pipes, erecting scaffolding, untangling wires. Planning and reasoning shortfalls account for 8%. Regulation blocks 14% of physical tasks and human preference 25%, the two together covering most of the work where a robot would be tolerated in a factory but not in a bedroom or a classroom. Two occupations in the paper show how mixed the picture is within one job: drywall tapers score 1.6 on the exposure index, with 41% of their time exposed to an autonomous spraying robot and 26% unexposed because the messy finishing work is beyond it; recycling sorters score 1.7, with over 75% of their time exposed to computer-vision grippers on conveyor belts and only 8% unexposed.

The authors' summary is the one a tradesperson should remember: unexposed work is "highly interpersonal or requires physical skills that robots today don't have", and "if the past is any guide, taxi drivers and warehouse packers will see changes sooner than nurses and mechanics".

Mechanisms: why the pattern looks like this

Four mechanisms explain the findings and predict what comes next.

Environments, not tasks, decide deployment. The four-tier scale is the paper's quiet insight. A robot that welds flawlessly on a factory line (E1) is useless on a domestic extension where every joist is a slightly different distance from the last. Most of the physical economy, and almost all of the trades, happens in E3 conditions: unstructured, improvised, around people. Robots have conquered E1 and are advancing into E2; E3 holds 1% of tasks after fifty years.

Hardware economics are not software economics. A language model, once trained, performs its next task at a cost close to zero, which is why exposure in desk work has become displacement in a few years. A robot is a depreciating asset that must be bought, installed, insured, maintained, powered and supervised for each site it works on. The 3% annual price decline is the signature of that difference, and the paper's scenarios show how little even a fourfold acceleration changes the decade in which parity arrives.

Capability compounds, but through data. Robots gain about 2% of physical work a year, and the half of remaining tasks that need dexterity are precisely the ones for which training data is scarcest. A model learns to write from the written web; there is no equivalent corpus of a plumber's hands. Progress on the remaining 24% depends on capturing real work, in real places, with the people who do it, which is slow by nature and raises consent and ownership questions that the software wave never faced.

Preferences and rules are a floor, not a lag. The 25% of tasks limited by preference and 14% by regulation are not waiting for a better robot. A patient's willingness to be lifted by a machine, a council's rules on autonomous vehicles and a customer's wish to speak to the person who fitted the kitchen change on political and cultural timescales, which is why the authors treat them as separate from capability and cost.

A sceptic should hold three alternative explanations alongside these. The exposure ratings are produced by a model reading terse task descriptions, which, as the authors note, "omit details that may be easy for humans but hard for robots", so capability may be overstated at the task level even as it is understated for the occupation as a whole. The cost estimates are approximations of a market with few public prices; a humanoid price survey in August 2026 found that only 12 of 58 commercial humanoids had a verified price at all. And the scenarios assume the same decline for every task; a breakthrough in general-purpose manipulation, the kind Nvidia and Tesla are betting on, could, in the authors' words, "leapfrog our scale". The dates are robust in shape and soft in detail.

How the industry responded

The reaction split along the line the paper draws. In the hardware camp, Nvidia's chief executive Jensen Huang had predicted a "ChatGPT moment" for robotics at CES in January 2026, and Elon Musk has said Tesla's Optimus humanoid will go on sale as early as 2027 with a long-term target price under $20,000 at scale, though no official price exists; Morgan Stanley's 2025 outlook puts the humanoid market above $5 trillion by 2050 with adoption accelerating from the late 2030s. Yahoo Finance framed the Anthropic paper, on 2 October, as the projection those bets will test. The shipping reality is more modest: Unitree's G1 lists at $13,500 and the company shipped more than 5,500 humanoids in 2025, more than every rival combined; Agility's Digit, the most deployed humanoid in US warehouses, is priced above $250,000 per unit or around $8,500 a month as a service; Figure, valued at $39 billion, sells to enterprises without a public price.

Economists mostly accepted the timeline. Apollo's chief economist Torsten Sløk, quoted by Yahoo Finance, said that even under aggressive projections widespread physical displacement would take decades, citing bottlenecks in fine manipulation, regulation and preferences. The sharper debate is on the digital side, where the Stanford payroll evidence and Anthropic's own unemployment analysis point in different directions on how much displacement has already happened, and where OpenAI, Anthropic, Google, Microsoft, Manus and Meta with Muse are shipping agents that browse, fill forms and complete multi-step office tasks, the category where exposure is highest and cost lowest.

Jobbit's position in this landscape is a design choice rather than a forecast, and the paper happens to describe it. The company's AI agent handles the digital, reviewable work the evidence says is cheap to automate: research, documents, websites, automations, images and the admin that surrounds a job. When a task needs hands, on-site judgement or someone accountable, which the paper's numbers say is most physical work for decades, the job passes to a vetted human professional on pro.jobbit.uk, with payment held in escrow and the customer able to watch the agent and take control at any point. Jobbit Labs, the company's research division, works on the slower problem the paper identifies in its data gap: capturing real-world task data from people doing real work, with consent, to train and verify world models for physical AI. Jobbit has not published trial data on the outcomes of this model and makes no claim here that it has; the claim is narrower, that an agent-plus-human-network architecture is the one the evidence supports while cost parity remains decades away. The practical side of that split is set out in 12 Jobs an AI Agent Can Do for Your Business This Week and Should Your Next Hire Be an AI Agent?, the freelance evidence in Do AI Agents Replace Freelancers?, and the company itself in What Is Jobbit?.

What a business should do with this evidence

The evidence supports a delegation rule that a small business can apply without a strategy consultant.

Automate the digital layer around physical work first, because that is where cost and capability already meet. A plumber's quotes, invoices, website, review requests, supplier research and scheduling are the 50% of work exposed to language models; the paper's own numbers say the fitting, the diagnosis in a damp cupboard and the conversation with the homeowner are not going anywhere. The margin in a trade business over the next decade comes from spending the saved desk hours on more site hours, not from waiting for a robot.

Treat physical automation as a procurement decision on a task, not a workforce decision on a job. The paper's cost model is the template: price the robot's whole annual cost against the labour on that one task, in your environment tier, and expect parity only in E1 and E2 conditions, repetitive packing, moving, welding and sorting, for years to come. For almost everything done in a customer's home, parity is not on the calendar.

Hire and train for the unexposed core. The 24% of physical tasks robots cannot do and the 25% people do not want them to do are the durable part of a trade: dexterity, judgement in unstructured spaces and the human relationship. CITB's 206,000-worker gap by 2030 says the market will pay for them.

Watch the handover, not the headline. The useful question for any job is where the machine should stop and the person should start, and that line moves task by task. A business that writes it down, who reviews the agent's output, who signs off the site work, who is accountable to the customer, captures the gains without the failures.

Finally, read the early signals in the right place. The paper suggests tracking exposed occupations, drivers and warehouse packers, for the first signs of physical disruption, and the payroll data says the digital disruption is already visible in graduate hiring. Plan for the second now and the first later.

Limitations and open questions

The study is American, model-rated and static. The exposure and cost ratings come from Claude reading task descriptions and web evidence about robots, checked against history but not against field trials; a human expert panel might rate dexterity differently. The cost estimates omit adoption frictions such as borrowing limits, and the scenarios set aside feedback effects, including the possibility that falling wages in exposed occupations slow adoption further. The paper does not examine how AI could change physical work without robots, for example through predictive maintenance, and it does not model the UK, where energy costs, regulation and the structure of the trades differ.

The digital evidence has its own limits. Payroll data shows relative employment of young workers in exposed occupations, not causation, and the Stanford authors describe their findings as consistent with "no widespread displacement so far"; Anthropic's unemployment analysis and the Stanford hiring analysis may both be right about different margins. Usage data from one assistant is not the economy. And the humanoid market is moving fast enough that any price quoted here is a snapshot of early October 2026.

The open questions are the ones that will decide the next decade: whether general-purpose manipulation improves faster than 2% a year once robots learn from real work rather than engineered demonstrations; whether the regulatory and preference floors move; and, as the authors put it, how work itself changes, "perhaps becoming more social and interpersonal as robots and AI advance".

Conclusion

Will robots take blue-collar jobs? Not in the way the 47% suggested and not on any timetable a working tradesperson needs to plan around. Anthropic's study shows that robots can already do most physical tasks and can afford almost none of them; that the economics close at a few per cent a year; and that even a world of record-breaking robot progress leaves half of physical work in human hands past 2050. The disruption that is measurable today is in digital work, where exposure is highest, cost is lowest and the first graduates are already missing from the payrolls. For a British trade or a small business, the evidence points to one strategy: automate the desk, protect the hands, and be deliberate about where one hands over to the other. That handover is where Jobbit works, and the paper is the clearest statement yet of why the gap it is built in will be there for a long time.

Frequently asked questions

Will robots replace blue-collar workers?

Not for decades, according to Anthropic's September 2026 study. Robots can already do about 74% of physical tasks in the United States but are cheaper than a worker for only 0.3% of them, and at the historical 3% a year fall in robot prices it would take about 40 years to reach cost parity on 10% of work. Even an aggressive scenario puts half of physical work beyond 2050.

Which jobs are most exposed to robots?

Vehicle operators and warehouse packers. Taxi drivers top the exposure index at 2.2 out of 3, nine of the ten most exposed occupations are drivers, and packers and packagers are the clearest case where a robot is already slightly cheaper than a person. Nurses, mechanics and most skilled trades sit at the other end because their work needs dexterity, judgement in unstructured spaces or human contact.

Is AI taking white-collar jobs faster than robots take physical ones?

The early evidence says yes, at the entry level. Stanford's analysis of payroll records for over 25 million workers finds employment of 22 to 25 year olds in the most AI-exposed occupations about 19% below less-exposed peers, with no equivalent change for experienced workers, while Anthropic's economists find no systematic rise in unemployment for exposed workers overall. Digital tasks cost almost nothing to automate once a model exists; physical tasks need hardware that is still expensive.

What should a tradesperson do about automation?

Automate the office side of the business, which is where AI is already cheap and capable, and invest in the skills robots cannot match: dexterity, diagnosis in unstructured spaces and the customer relationship. The UK construction sector alone needs more than 206,000 extra workers by 2030 according to CITB, so the human part of the trade is in shortage, not surplus.

What is the gap Jobbit is built for?

The handover between automated digital work and human physical work. Jobbit's AI agent does the research, documents, websites and admin that the evidence says are cheap to automate, and passes work that needs hands, a site visit or accountability to a vetted professional on pro.jobbit.uk, with payment held in escrow. Jobbit Labs works on the real-world task data that the study identifies as the missing ingredient for physical AI.