The Workplace Will Move First

On 30 September, Anthropic published an unusual attempt to measure what robots can actually do at work today. Its researchers examined thousands of tasks across the US economy, from moving freight and dispensing medication to digging trenches and stocking shelves. Their headline finding was striking: robots can already perform roughly three-quarters of physical work tasks in at least some circumstances. Those tasks account for around a third of all working time.

It sounds like the beginning of a familiar automation story. Machines can already perform an enormous amount of human work; therefore, large numbers of jobs must be close to disappearing.

Except another number in the same research makes that interpretation much harder to sustain.

Anthropic estimates that robots are currently cost-competitive with human workers for just 0.3 per cent of job tasks.

The distance between those two numbers - 74 per cent technically possible and 0.3 per cent economically competitive is where the interesting story begins.

A robot being capable of doing something is not the same as being able to do it usefully, cheaply and reliably in the place where the work actually happens. The world humans work in is inconveniently complicated. Objects are left in the wrong places. Floors are uneven. Packages arrive damaged. Customers change their minds. Wires become tangled. Someone parks where they should not. A patient reacts unexpectedly. A screw has been overtightened by whoever was there last Tuesday.

People navigate this disorder almost without noticing it.

Machines generally do not.

Anthropic's researchers therefore classified robotic capability partly according to the environment a machine requires. At one end are tasks robots cannot currently perform. Then come tasks they can perform inside purpose-built robotic environments, followed by tasks possible in structured human workplaces such as warehouses. Only the highest category covers machines capable of doing the work in relatively uncontrolled environments such as public roads.

That distinction changes the automation question.

We normally imagine progress coming from making robots progressively better at operating in our world. But there is another option. We can make our world progressively easier for robots to operate in.

The workplace can move first.

A factory provides the clearest version of this idea. Industrial robots became useful long before machines could understand a factory in anything resembling the way a human worker could. Instead, production lines made their environment predictable. Components arrived in known positions. Movements were repeated. Physical barriers kept people away from machinery. Tasks were separated into sequences that machines could perform reliably.

The intelligence was partly in the robot. But some of it was effectively built into the environment.

Warehouses are becoming a more sophisticated version of the same principle. Amazon describes autonomous robots that navigate fulfilment centres, systems coordinating thousands of mobile units, and its Vulcan robot, which combines computer vision with force sensing so that it can manipulate objects while detecting physical contact. These are impressive technical advances. Yet the warehouse itself remains part of the technology. Shelving, inventory systems, barcodes, package dimensions, routes, software and working practices all contribute to making the physical world legible enough for automation.

That matters because redesigning an environment can sometimes be easier than solving intelligence.

Consider a robot that struggles to find randomly shaped goods on cluttered shelves. One approach is to spend years improving its vision and dexterity until it can cope with whatever humans place in front of it. Another is to standardise the containers, control where goods appear and change the shelving.

From the robot's perspective, the problem has become easier. From the organisation's perspective, the work has changed.

We already do this constantly without describing it as automation. Self-service checkouts ask customers to perform part of the retailer's workflow. Airport passengers scan their own boarding passes. Warehouses determine where workers should walk and what they should pick. Restaurants redesign menus around production systems. Digital forms constrain messy human circumstances into fields that software can process.

Sometimes the machine becomes more capable. Sometimes the surrounding system becomes less ambiguous.

Robotics makes this trade particularly visible because physical reality is stubborn. Anthropic found that only about 2 per cent of physical work can currently be performed by robots in unstructured environments. Much more becomes possible once the environment is constrained.

This suggests that one of the most important questions about automation is not simply what machines will learn to do. It is what organisations will change so that machines do not have to learn as much.

That distinction matters for workers because jobs rarely disappear as indivisible objects. They are collections of tasks, relationships, exceptions and responsibilities.

A warehouse employee might spend part of a shift transporting goods, part resolving inventory discrepancies, part handling damaged packages and part responding when the process stops behaving as expected. A machine may be excellent at the first activity and poor at the last three. Automation therefore does not necessarily remove the worker. It can remove or reorganise the predictable part of the job and leave the person responsible for everything that does not fit.

Economists have been studying this task-based character of automation for years. David Autor argued in 2015 that technology both substitutes for and complements human labour; occupations persist partly because jobs combine tasks that computers handle well with others where human adaptability remains valuable. Research by Daron Acemoglu and Pascual Restrepo has meanwhile found substantial labour-market effects from industrial robot adoption, including lower employment and wages in more exposed US labour markets.

The important point is that neither technical exposure nor demonstrated capability translates mechanically into unemployment.

Current US labour projections illustrate the problem nicely. Anthropic identifies warehouse-related work as among the areas with substantial robotic exposure. Yet the US Bureau of Labor Statistics expects employment of stockers and order fillers to grow by about 9 per cent between 2025 and 2035, while employment of hand packers and packagers is projected to fall by about 5 per cent. Both occupations operate in environments being transformed by automation. Their employment trajectories are nevertheless different because demand, technology, task composition and the economics of adoption differ.

“Can a robot do this?” therefore turns out to be a relatively poor question on its own.

A better question is: what else must change before it makes sense for a robot to do this?

Sometimes the answer will be price. Anthropic estimates that a suite of equipment capable of performing much of the work of packers and packagers can cost more than $2 million to purchase and install, although those machines may replace the annual work of multiple employees. For many other occupations in the study, robotic alternatives remain several times more expensive than people.

But economics is not the only constraint. Dexterity matters. Regulation matters. Trust matters. Safety matters. Human preference matters. Anthropic estimates that limitations in robotic capabilities still obstruct automation across around 70 per cent of physical tasks, with manipulation, the apparently mundane business of handling physical objects particularly important. Regulation constrains another set of tasks, especially in areas such as healthcare, education and protective services.

And then there is the environment.

Once organisations begin altering environments for machines, automation can affect work before it removes work.

A job may become more prescribed because variation creates problems for the machine. Objects may need to arrive in particular places. Processes may need to happen in a fixed sequence. Exceptions may be routed to designated human workers. Performance may become easier to measure because the system needs precise information about what is happening. Some forms of skill may become less important while the ability to diagnose exceptions becomes more valuable.

The human role can gradually migrate towards the edges of the system.

This has an uncomfortable implication. The jobs most resistant to automation may not necessarily be those involving the most expertise. They may be those containing the most troublesome combination of ambiguity, physical variation, social interaction and exception handling.

A nurse does not merely transport medication. A maintenance engineer does not simply tighten bolts. A care worker does not perform a fixed sequence of physical actions. Much of the value of these jobs exists in noticing that today's situation is slightly different from yesterday's and changing behaviour accordingly.

Anthropic's own results point in this direction. Nursing and general repair remain comparatively difficult to automate, while driving and warehouse work are considerably more exposed. Work combining interpersonal contact with difficult physical manipulation is among the least accessible to today's combination of robots and language models.

That should make us cautious about describing occupations as simply “safe” or “at risk”. The boundary itself can move.

An organisation unable to automate a complex job may decompose it. The predictable tasks can be routed towards machines, the uncertain tasks towards people, and the entire process reorganised so fewer moments require judgement. Customers may also be asked to absorb some of the complexity: entering information themselves, packaging something differently, arriving within narrower windows or following procedures designed partly around what the system can process.

Costs do not always disappear when a process is automated. Sometimes they move.

Nor is redesigning work around machines necessarily undesirable. Dangerous, repetitive and physically punishing work is an obvious candidate for automation. The Bureau of Labor Statistics notes that many robot-exposed occupations involve lifting, hazardous conditions or other significant physical demands. Anthropic similarly finds that workers in highly robot-exposed jobs are more likely to be lower-paid and to work under physically demanding conditions.

A machine taking over those tasks can be an improvement.

But the design decision matters. There is a difference between using machines to remove the worst parts of a job and redesigning the job so that the person becomes the remaining flexible component in an otherwise automated system.

One expands human capability. The other can leave humans handling precisely the awkward, unpredictable and emotionally demanding work machines cannot manage.

The history of automation contains both.

This is why the 0.3 per cent figure may ultimately be more interesting than the 74 per cent one.

The large number tells us something about the technological frontier. The small number reminds us how much infrastructure, capital, organisational change and economic justification must sit between a demonstration and widespread adoption.

Anthropic's own projections make the uncertainty clear. Using historical rates of improvement and falling robot costs, its researchers estimate that reaching cost competitiveness across even 10 per cent of work could take decades. Under a much faster scenario, widespread automation arrives considerably earlier. These are scenarios rather than forecasts, and the researchers explicitly acknowledge that advances in AI-powered robotics could break with historical patterns.

But organisations do not have to sit still while waiting.

They can simplify environments. Standardise inputs. Separate tasks. Change buildings. Introduce sensors. Alter workflows. Shift responsibilities to customers. Remove exceptions where possible and create procedures for the ones that remain.

Gradually, a place designed for people can become a place designed for people and machines.

And when that happens, looking only at the robot misses half the technology.

The interesting thing about today's robots is therefore not simply how close they are getting to behaving like us. Many remain nowhere near it.

It is how much of the world around them we may be willing to change so that they do not have to.

References

Anthropic, Russell Legate-Yang and Maxim Massenkoff, What work can robots do?, 30 September 2026.

US Bureau of Labor Statistics, Hand Laborers and Material Movers, 2026 Occupational Outlook Handbook.

US Bureau of Labor Statistics, Occupational Projections and Worker Characteristics, 2025–35.

David H. Autor, Why Are There Still So Many Jobs? The History and Future of Workplace Automation, Journal of Economic Perspectives, 2015.

Daron Acemoglu and Pascual Restrepo, Robots and Jobs: Evidence from US Labor Markets, Journal of Political Economy / NBER.

Amazon Web Services, Learn from Amazon Stores on AWS — Robotics.