Shape of the labor market

The next decade will be defined by where human labor remains scarce.

Jobs are bundles of tasks. Technologies first automate individual tasks, then workflows, then — if enough of the workflow becomes cheaper and more reliable to automate — the number of humans required to produce the same output falls. An occupation can remain intact while employing half as many people. Another can become far more productive without losing employment because lower costs create more demand.

Two automation curves are now advancing at once. Artificial intelligence is making cognitive work cheaper inside digital environments. Advanced robotics is beginning to extend that intelligence into the physical world.

The most useful question is Where will human labor remain scarce as intelligence and physical autonomy become cheaper?

This matters because two automation curves are now advancing at once. Artificial intelligence is increasingly capable of performing cognitive work inside digital environments. Advanced robotics is beginning to extend that intelligence into the physical world. Their progress will be uneven, but together they cover a much larger share of human work than either technology does alone.

The workforce today

The labor market is much less white-collar than discussions about AI would suggest. Office and administrative support, food service, transportation, sales, healthcare, management, construction, maintenance, and production still account for enormous shares of employment.

Four types of work

The workforce can be simplified along two dimensions: whether the work is primarily cognitive or physical, and whether it is routine or non-routine.

This creates four different automation problems.

Routine knowledge work is the natural territory of AI.

Routine physical work is the natural territory of robotics.

Non-routine knowledge work is more likely to be compressed and augmented than eliminated.

Non-routine physical work has the strongest near-term defenses because machines must combine intelligence with dexterity inside unpredictable environments.

The boundaries are not absolute. A nurse performs both routine and non-routine tasks. A software engineer alternates between commodity coding and architectural judgment. The question is which tasks constitute the economic reason the human is employed.

Two automation curves

Artificial intelligence operates most naturally in digital environments. It can already write, retrieve information, classify documents, analyze data, generate software, communicate with customers, operate software, and execute increasingly long sequences of cognitive work.

This makes digital, standardized work unusually exposed. Clerical occupations have the highest generative-AI exposure, with exposure also rising in highly digitized professional occupations. Not every exposed occupation will disappear, but a large portion of existing work will most likely be transformed.

Embodied autonomy faces a harder problem. A robot must perceive a changing physical environment, manipulate imperfect objects, avoid harming people, survive hardware failures, and do all of this cheaply enough to outperform human labor.

But the curve is moving. More than half of professional service robots sold are designed for transportation and logistics, while U.S. industrial robot installations rose 11% in 2025. Structured physical environments—factories, warehouses, and commercial facilities—are already the easiest places to deploy them.

The two curves are converging. Intelligence can increasingly perceive, reason, decide, and act. As the software improves, the bottleneck shifts toward hardware cost, reliability, regulation, and the complexity of the physical environment.

What makes a career durable?

A machine replacing a task requires the combination of capability, reliability, and economics.

Productivity does not mechanically destroy employment. If AI makes software, healthcare, legal work, or design dramatically cheaper, lower prices may create enough additional demand to offset some of the labor savings.

The occupational question is therefore not simply Can this be automated?
The better question is What remains scarce after the easy parts are automated?

Routine knowledge work

Routine knowledge work is the most exposed category over the near term. The work already occurs inside computers, often follows defined procedures, and usually produces outputs that can be evaluated digitally.

Basic financial analysis, commodity research, low-complexity legal work, generic copywriting, standardized design production, and portions of software development share the same underlying structure: receive digital information, manipulate it according to learned patterns, and produce another digital artifact.

I would be reluctant to begin a career whose economic value is primarily moving, retrieving, summarizing, formatting, or transforming information.

The better direction within knowledge work is toward jobs that own something downstream of the information: a decision, a client, revenue, a system, or an outcome. Complex sales, domain-specific operations, product ownership, high-stakes advisory work, and managing automated systems are all better positioned than pure information processing.

The work will not disappear.
The more plausible outcome is that fewer people supervise much more output.

Non-routine knowledge work

Non-routine knowledge work has a different problem. The occupation may remain valuable while the organization around it becomes much smaller.

Consider a law firm, investment firm, software company, or consulting practice. A significant share of junior labor exists to research, draft, model, code, prepare presentations, document decisions, and synthesize information for more senior people.

AI directly attacks that layer of the pyramid.

A future professional-services firm may not need eight junior employees beneath every senior employee. It may need two, each operating substantially more capable software.

This creates an important distinction between occupational survival and career opportunity. Lawyers may remain. The number of junior lawyers required per partner may decline. Software remains essential while fewer engineers may be required to ship a given amount of software. Management remains necessary while reporting and coordination layers become thinner.

If I were choosing within this quadrant, I would orient toward entrepreneurship, applied AI, senior engineering, technical infrastructure, research, complex sales, investing, product ownership, and positions with direct decision authority.

I would be more cautious about generic junior analyst work, commodity software production, low-complexity legal services, generic content production, and middle-management positions whose primary function is moving information between other people.

The durable premium is moving upstream. Knowing information matters less when knowledge is cheap. Selecting the right problem, making the decision, and being accountable for the result matter more.

Routine physical work

Routine physical work is less exposed to AI but more exposed to the next generation of robotics.

Warehouses, factories, highways, and commercial kitchens are attractive automation environments because they are comparatively structured. Objects can be standardized. Routes can be mapped. Humans can sometimes be separated from machines. A robot performing the same motion ten thousand times is far easier to engineer than one entering an unfamiliar home and repairing an unfamiliar plumbing problem.

I would therefore be cautious about starting a decades-long career centered on long-haul driving, warehouse picking, checkout, repetitive assembly, stocking, or highly standardized food preparation.

The better position is adjacent to the automation rather than underneath it: robotics maintenance, industrial automation, electrical systems, controls engineering, fleet operations, equipment repair, infrastructure, and technical field service.

The machine does not need general human capability.
It only needs enough capability for that environment.

Non-routine physical work

Non-routine physical work has the strongest near-term combination of defenses.

An electrician enters different buildings, diagnoses poorly documented systems, manipulates physical components, adapts to previous workmanship, interacts with customers, complies with codes, and bears real consequences for failure. Plumbing, HVAC, mechanical repair, and much of construction have similar properties.

Healthcare adds another layer: physical intervention, regulation, liability, trust, and demand.

These occupations are not immune to automation. A nurse will use AI for documentation, monitoring, and decision support. An electrician may use computer vision for diagnostics and software agents for quoting, scheduling, and procurement. A mechanic may diagnose faults with AI before touching the vehicle.

That is precisely why many of these careers look attractive: the technology automates peripheral work while leaving the human bottleneck intact.

If I were choosing a career from scratch, I would take skilled trades more seriously than the prestige hierarchy of the last few decades suggests. Electrical work, HVAC, plumbing, industrial maintenance, specialized construction, mechanical repair, nursing, dentistry, and intervention-heavy medical specialties all combine durable demand with relatively difficult automation.

The strongest version often adds ownership. An electrician using AI is valuable. An electrician who owns the electrical company, automates its back office, and coordinates a fleet of technicians has considerably more leverage.

Where human scarcity moves

In ten years, entirely new occupations will emerge, so the composition of the workforce cannot be predicted precisely — but the direction is clearer.

We should expect:

That does not imply the end of work. Healthcare demand should continue to grow. Physical infrastructure still needs to be built and maintained. And the systems doing the automating require enormous amounts of capital, energy, hardware, and technical labor. New work will emerge around those constraints.

The deeper change is in what remains scarce. Industrial machinery made physical force cheap. Computers made calculation cheap. The internet made information distribution cheap. AI is making cognition cheaper. Robotics will increasingly make physical execution cheaper. Human economic value does not disappear when something becomes abundant. It migrates.

Over the next decade, I expect the premium to move toward judgment, responsibility, trust, persuasion, domain expertise, dexterity in irregular environments, control over scarce resources, and ownership.

Ownership becomes particularly important as labor becomes more leveraged. A worker sells one person's time. An owner can deploy software, machines, capital, and other people's labor. As those inputs become more productive, control over decisions, customer relationships, distribution, equity, intellectual property, and productive assets becomes correspondingly more valuable.