Robots Can Do 74% of Physical Work. Only 0.3% of It Pays.

A chart from Anthropic has been circulating for the past week, and the number attached to it is 81%. That is the share of US work now considered exposed to AI once robots are counted alongside language models, up from roughly half for language models on their own.

The number is real. It comes from What work can robots do?, published by Anthropic on 30 September 2026. But 81% is close to the least useful figure in the paper, and the version circulating on social media has mostly dropped the finding that actually matters for anyone planning a budget.

Robots can technically perform three quarters of physical work in the United States. They are cheaper than the human doing it in 0.3% of cases.

What the report measured, and what it did not

The research rates roughly 7,594 physical job tasks drawn from O*NET, the US occupational database, against what robots can demonstrably do today. Tasks fall into four tiers:

Read that distribution carefully, because the headline compresses it into something misleading. Counting every tier, robots can do 74% of physical work. But half of that requires rebuilding the workplace around the machine, and only 2% of physical tasks can be done in the messy, unstructured settings where most work actually happens.

Exposure, in this report, means capability. It is explicitly not a prediction of job losses, and the authors are direct about this.

“Robots, which we define as autonomous physical machines that sense and act, can perform three-quarters of physical tasks in the US, making up 34% of working hours, but mostly in limited settings.”

Where the 81% comes from

The 81% figure appears once, in a single sentence describing Figure 6. That chart overlays robot exposure onto the language-model exposure measure from Eloundou et al. (2024), the study behind most AI exposure numbers you have seen since 2023.

The shift is dramatic in physical categories. Fewer than 15% of transportation and moving tasks are exposed to language models alone. About 90% are exposed once robots are included. Office and administrative support, which mixes computer work with light physical tasks, rises to nearly 100%.

Nine of the ten most exposed occupations are vehicle operators. Taxi drivers top the list. This is the genuinely new information in the report: AI exposure stops being a white-collar story and becomes a driving and warehousing story.

It is also worth noting that no occupation group falls below 40% exposure. The widely shared list of jobs that are supposedly safe is softer than it looks.

The number nobody is sharing

Here is the finding that should change how you plan.

“Though robots can theoretically do tasks that add up to 34% of all work time, they are cost-competitive for just 0.3% today.”

Capability and economics are not the same thing, and for robots the gap between them is enormous. Software exposure translates into adoption quickly because copying software costs almost nothing. A robot is hardware. It has to be bought, installed, integrated, maintained, powered, insured and eventually decommissioned.

The report runs the arithmetic on specific jobs. Packers and packagers are the largest occupation where robots are already cost-competitive, and the margin is thin: robots cost roughly $2,500 less per year than the worker. Welders need robots costing about five times a human welder. Robotaxis come in around $7,000 more per year than a taxi driver, and face regulatory barriers on top.

For robots to become cost-competitive across just 10% of human work, costs would need to fall about 70%. Robot prices have declined roughly 3% per year since the 1990s. At that rate the report puts the timeline at around 40 years.

Even an aggressive scenario, with costs falling up to four times faster and capabilities improving twice as fast, only reaches cost parity for half of physical work by 2050.

Four barriers, not one

Cost is the most quantifiable obstacle but not the only one. The report separates the constraints, and the proportions are instructive:

• Capability blocks adoption for around 70% of physical tasks. Manipulation is the specific weakness: half of physical tasks would not be automated at scale unless robots get substantially better at touching and handling objects.
• Cost would need to fall for nearly all tasks, as above.
• Human preference holds robots back for about a quarter of tasks, either because people would not trust a machine with the work or because the social interaction is the point.
• Regulation currently prevents robots from performing 14% of physical tasks, concentrated in healthcare, protective service and education.

Only about 8% of physical tasks are limited by planning and reasoning, which is the part that better AI models are most likely to fix. The bottleneck is hands, not brains.

Why this lands differently in the UAE

Most markets can treat a 40-year cost curve as somebody else’s problem. The UAE cannot, because public policy is deliberately bending that curve locally.

The Dubai Robotics and Automation Programme, launched under the Dubai Future Foundation, targets the deployment of 200,000 robots by 2032 and aims to raise the robotics sector’s contribution to Dubai’s GDP to 9% in the same period. The focus areas include logistics, production and manufacturing, healthcare and connected mobility.

That matters because of what the report says about where robots work today. Robots are strongest in structured environments: warehouses, ports, assembly lines, fulfilment centres. The UAE has concentrated exactly that kind of infrastructure at Jebel Ali, Dubai South and Khalifa Industrial Zone, and it is deploying public capital to accelerate adoption rather than waiting for hardware prices to drift down.

The practical implication is that the global cost timeline is the wrong planning input for a UAE operator. Subsidy, scale and policy can pull specific sectors forward well ahead of the 40-year curve, while leaving others exactly where they are.

What to do with this

The temptation after a report like this is either to panic or to dismiss it. Both are wrong, and both come from treating exposure as a forecast.

Do not reorganise around the 81%. It describes what is technically possible, not what is economically sensible. An organisation that restructures its workforce on an exposure number is responding to a capability demonstration, not a business case.

Separate your exposure analysis into two questions. First, can the technology do this task at all? Second, is it cheaper than the current arrangement, including integration, supervision and downtime? The report found that around 300,000 US workers are in roles where robots can do 95% of the tasks. That is a real number, and it is also 0.2% of the workforce.

Watch structured environments first. If you run warehousing, fulfilment, port logistics or packing operations in the UAE, your exposure is live now rather than theoretical. If your operation is unstructured, variable and hands-on, you have considerably more time than the headline suggests.

Treat language-model exposure as the nearer risk. For most businesses reading this, the cognitive half of the chart is where change arrives first, because software deployment has none of the hardware economics that slow robots down. The office and administrative support category is exposed on both axes, which makes it the clearest near-term pressure point.

The honest summary

Robots can do far more work than most people assume, and far less of it profitably than the headline implies. Those two statements are both in the same report, and only one of them travelled.

This is the recurring pattern in AI adoption, and it is the same one I keep writing about from a different angle each time. The technology works as specified. The returns disappoint because capability was mistaken for readiness, and nobody did the arithmetic on what it would cost to actually deploy.

The 0.3% is the arithmetic. It is a far more useful number to plan around than the 81%.

Frequently asked questions

What does AI or robot “exposure” actually mean?
Exposure measures whether current technology could perform a task, not whether it will replace the worker doing it. Anthropic’s robot exposure index rates whether a robot can do a task today and in what kind of environment, from purpose-built factory floors through to open city roads. It excludes cost, regulation and human preference, all of which are handled separately in the report.

Are 81% of jobs about to be automated?
No. The 81% figure covers tasks that AI or robots are technically capable of performing. The same report finds robots are cheaper than the human worker for only 0.3% of tasks, and estimates that reaching 10% would take around 40 years at historical rates of robot price decline.

Which jobs are most exposed to robots?
Nine of the ten most exposed occupations are vehicle operators, with taxi drivers ranked highest. Transportation and moving work rises from under 15% exposure with language models alone to about 90% once robots are included. Warehouse and packing roles follow closely.

Which jobs are least exposed?
Personal care and service, installation and repair, healthcare support, and community and social service show the lowest exposure, because they combine in-person interaction with delicate physical handling. However, no occupation group overall falls below 40% exposure.

What is the biggest barrier to robots replacing physical work?
Capability, specifically manipulation. Around 70% of physical tasks are blocked by what robots cannot yet do, and half of all physical tasks would require robots to become substantially better at touching and handling objects. Cost is the second constraint, with human preference and regulation accounting for roughly 30% between them.

Source

Legate-Yang, Russell and Maxim Massenkoff. What work can robots do? Anthropic, 30 September 2026. Language-model exposure data from Eloundou, Tyna, Sam Manning, Pamela Mishkin and Daniel Rock, “GPTs are GPTs: Labor Market Impact Potential of LLMs,” Science 384 (6702), 2024. UAE programme detail from the Dubai Robotics and Automation Program, UAE Government portal.

Further reading

Related: AI Adoption in the UAE covers what adoption actually involves beyond the capability headlines. Enterprise AI Adoption covers the governance side at scale. AI Adoption for UAE SMEs covers where smaller organisations should begin. Why AI Assistant Costs Run Away covers the same capability-versus-economics gap on the software side.

Martin Alva

Marketing, Brand & Digital Transformation Leader, United Arab Emirates

Senior Manager, Employee Experience (Brand Positioning, Digital Platforms & AI) at Space42 (A G42 & Mubadala Company). 15+ years in marketing, brand and digital transformation across MENA and Europe. 5 MENA Effie Awards.

More about Martin Alva → Connect on LinkedIn →