Artificial intelligence is already raising productivity in narrow white-collar tasks, yet Stanford economist Charles I. Jones shows that even infinite automation of half the economy lifts total output by only about 19 percent because physical tasks remain the binding constraint. Growth per worker has hovered near 2 percent for 150 years; the best near-term guess is a rise toward 3 percent this decade before robots unlock sharper gains.
Bill Conerly, writing in Forbes, frames the practical question for firms: prepare for stronger demand, or for obsolescence. The Jones model supplies a clearer timetable than pure hype or pure skepticism.
The timetable matters because the two paths diverge slowly at first and then sharply later. Early gains arrive through software harnesses that already exist. Larger gains wait on machines that can close the physical half of the task chain. That sequence, not a single leap, is what the model and the recent data both describe.
What the Jones weak-link model actually predicts
In his NBER working paper “A.I. and Our Economic Future,” Jones treats production as a chain of complementary tasks. Easy tasks (cognitive work done on computers) can be automated by current large language models and agents. Hard tasks (anything requiring physical presence, dexterity or real-world judgment) stay with humans for longer.
When the elasticity of substitution is low, output behaves like a harmonic mean: the weakest link dominates. Automate the easy half of GDP to infinity and the formula yields a 19 percent output gain from half automation. Double living standards still requires infinite automation of roughly 94 percent of tasks.
| Share of GDP automated to infinity | Approximate GDP multiplier (σ=0.2) |
|---|---|
| 2% (roughly today’s software) | ~1.02 |
| 50% (all computer cognitive work) | 1.19 |
| 94% | ~2.0 |
| Near 100% | Explosive |
Jones notes the familiar desktop computer: users now control roughly 100 million times more transistors than in the early 1970s, yet economists are not 100 million times more productive. Someone still has to decide which matrix to invert.
The harmonic-mean logic is unforgiving at low substitution. Each unfinished physical task pulls the whole chain back toward the slower pace of human hands. Cognitive abundance therefore raises the value of the remaining scarce steps rather than erasing them. That is why the jump from 50 percent to 94 percent automation matters far more than the jump from today’s software slice to the full cognitive half.
In short, the model separates two claims that often get mixed together. Infinite progress on easy tasks is already priced into a modest multiplier. Explosive growth requires nearly complete coverage of the hard tasks as well.
White-collar harnesses run ahead of blue-collar hands
Conerly’s split matches the data so far. Narrow “harnesses” wrap foundation models around insurance claims, warehouse estimating or code generation. Software engineering timelines measured by METR have lengthened from seconds to hours in a few years. Entry-level hiring in AI-exposed digital occupations has already softened in some datasets.
- Easy today: cost estimating, document drafting, claims triage, spreadsheet modeling, basic coding agents
- Still hard: running conduit in a building, lifting irregular packages, roofing, cleaning restrooms, nursing hands-on care, fixing a leaky sink
- Transition zone: warehouse sorting and inspection, where mobile robots and arms are gaining but still need human oversight
Physical work has improved, yet the gap remains wide. A robot that lifts a ten-pound box is not yet ready for a ten-pound baby or a complex construction site. Crowd conversation on X tracks the same asymmetry: white-collar knowledge work faces restructuring first while many trades look temporarily sheltered.
The harness pattern also explains why measured productivity can rise in pockets without yet rewriting the aggregate trend. Claims triage and basic coding agents compress hours inside specific workflows. They do not remove the need for a person on a roof or at a bedside. Until those complementary steps move, the easy gains stay local.
The transition zone is where the next evidence will show up first. Warehouse sorting and inspection already mix mobile robots with human oversight. Progress there narrows the physical gap without requiring a full general-purpose humanoid. It is incremental, visible, and still short of the coverage the Jones curve needs for explosive rates.
The decade ahead looks faster, not explosive
Conerly’s working forecast starts from the long-run 2 percent per-worker trend. He expects that to climb toward 3 percent this decade as AI-easy tasks spread. Ten years at 2 percent leaves the economy 22 percent larger; at 3 percent it is 34 percent larger. Useful compounding, far short of the 5-plus percent rates Jones says become possible once weak links mostly vanish.
Recent baseline: nonfarm business labor productivity nonfarm productivity rose 1.4 percent at an annualized rate in the second quarter of 2026, and 2.2 percent from a year earlier. Over the current business cycle the annualized rate sits at 2.1 percent, matching the long-term post-1947 average.
- 2% trend: economy ~22% larger after a decade
- 3% path: economy ~34% larger after a decade
- Later acceleration: once robots clear most physical tasks, models allow rates above 5%
That intermediate window is the planning horizon for most capital budgets and hiring plans.
The recent baseline sits right on the long-term average, which keeps the near-term story disciplined. A climb toward 3 percent would be a clear step up from 2.1 percent cycle-to-date without requiring the full robot transition. The gap between 34 percent larger and the steeper path above 5 percent is exactly the physical-task gap the model highlights.
| Path | Decade outcome | What has to be true |
|---|---|---|
| 2% trend | ~22% larger | Long-run average continues |
| 3% path | ~34% larger | AI-easy tasks spread widely |
| Above 5% | Much steeper | Weak links mostly cleared by robots |
Past general-purpose technologies never broke the 2 percent line
Electricity, the internal combustion engine, semiconductors and the internet each transformed daily life. Average real income per person in the United States still tracked a remarkably straight 2 percent line for 150 years. Jones’s interpretation is that each new general-purpose technology offset the natural tendency for ideas to get harder to find inside the previous paradigm. Without the next wave, growth would have slowed; with it, the line held.
AI could be different because it automates intelligence itself, including the discovery of new ideas and the design of better robots. Yet the same weak-link logic that kept prior gains from exploding still applies until nearly every complementary task is also automated. In a Stanford GSB conversation, Jones describes the two extreme scenarios: business-as-usual continuation of the 2 percent path, or the Silicon Valley flywheel that eventually produces “a country of geniuses in a data center” and robots that match human physical capability.
The historical record therefore cuts both ways. Prior general-purpose technologies prove that transformation of daily life need not bend the per-person growth line. They also show why a technology that designs better robots could eventually do what earlier waves did not: close the complementary tasks that held the harmonic mean in place. The difference is not slogans about intelligence. It is whether the physical half of the chain finally moves.
- Prior waves: electricity, engines, semiconductors, internet kept the 2 percent line intact for 150 years.
- Jones reading: each wave offset harder-to-find ideas inside the old paradigm.
- AI twist: intelligence and robot design can be automated, opening a path the earlier waves lacked.
- Still binding: explosive rates wait until complementary physical tasks are automated too.
Robots remain the scarce piece of the flywheel
Industrial robot installations illustrate both progress and distance. The International Federation of Robotics recorded 542,000 industrial robots installed globally in 2024, more than double the level of a decade earlier, with operational stock at 4.66 million units. The United States installed 34,200 that year. Preliminary 2025 figures show a U.S. rebound toward 38,000 and global growth continuing. Asia still accounts for the large majority of deployments.
| Region / country | 2024 installations (approx.) | Share or note |
|---|---|---|
| Global | 542,000 | Second-highest year on record |
| China | 295,000 | 54% of world |
| United States | 34,200 | Largest in Americas |
| Europe | 85,000 | Down 8% from prior year |
These machines excel at repetitive factory and warehouse moves. General-purpose humanoids that can roof a house, clean a restroom or navigate an unstructured job site are still early. Data quality for physical trajectories, not just model size, is frequently cited as the current choke point. Until that layer matures, the 19 percent ceiling on cognitive-only automation stays relevant.
The geography of installation reinforces the same point. China alone took 54 percent of world installations in 2024, while the United States remained the largest market in the Americas at 34,200 units and Europe slipped 8 percent to about 85,000. Scale is rising, yet the stock of 4.66 million operational units is still concentrated on repetitive moves. That is progress inside factories and warehouses, not yet a substitute for unstructured trades.
Watch the composition as closely as the totals. A rebound toward 38,000 U.S. installations in preliminary 2025 figures signals demand, not a sudden arrival of humanoids that can handle a leaky sink or a complex site. The flywheel Jones describes needs both the data-center geniuses and the physical match to human capability. The second piece is the scarce one.
Who feels the mixed path first
Software and professional services capture the early productivity lift and the early labor-market pressure. Manufacturing and logistics see incremental robot density gains. Construction, hospitality, personal care and many trades keep human bottlenecks longer. In Europe, services already acting as Europe’s weak link shows how a slow sector can drag overall growth even when other parts accelerate.
Jones himself is relatively sanguine on the long-run labor market once the transition is complete, while stressing the double-edged nature of the technology: the same capabilities that raise living standards also raise catastrophic risks if misused. Firms that treat the next several years as a period of solid but not revolutionary demand growth, while monitoring robot cost curves and deployment rates, will be better positioned than those betting on either stagnation or immediate explosion.
- First movers: software and professional services, with early lift and early hiring pressure
- Incremental gains: manufacturing and logistics through higher robot density
- Longer bottlenecks: construction, hospitality, personal care, and many trades
- Aggregate drag risk: a slow services layer can hold back overall growth even when other sectors speed up
Sector timing therefore lines up with the weak-link math. The occupations already inside digital harnesses feel both the productivity and the labor-market effects first. The occupations still defined by dexterity and presence set the pace for the whole economy until robot capability catches up. Europe’s services drag is a live illustration of that arithmetic at national scale.
Complementary Tasks Hold the Multiplier Down
The 19 percent ceiling is not a forecast about model quality. It is a statement about task shares and low substitutability. When half of GDP can be automated to infinity and the other half cannot, the harmonic mean keeps total output close to the slower half. Doubling living standards still waits on infinite automation of roughly 94 percent of tasks, not on another leap in language-model benchmarks.
That is the same structure as the desktop-computer example. Transistors multiplied by roughly 100 million from the early 1970s, yet output per economist did not. The scarce complementary step, deciding which matrix to invert, capped the gain. Cognitive AI multiplies the easy steps in the same way. Physical presence, dexterity, and real-world judgment remain the steps that set the average.
Conerly’s firm-level question follows directly. Stronger demand is the right base case while easy tasks diffuse and the per-worker trend inches from about 2 percent toward 3 percent. Obsolescence across broad physical occupations is the wrong base case until installation numbers and autonomous task length show the hard half is finally moving.
Capital Plans Track the Robot Gap Closely
Capital budgets and hiring plans live inside the intermediate window, not inside the extreme scenarios. Ten years at 3 percent produces an economy about 34 percent larger, against about 22 percent at the old 2 percent trend. That gap is large enough to matter for capacity and headcount. It is still far short of the rates above 5 percent that appear only after weak links mostly vanish.
The practical monitors are already in the public numbers. Nonfarm productivity at 2.1 percent over the current cycle matches the post-1947 average and leaves room for a climb without proving a break. Global robot installations at 542,000 in 2024, a U.S. figure of 34,200, and a preliminary rebound toward 38,000 give a baseline for physical diffusion. METR timelines stretching from seconds to hours show how far cognitive harnesses have already come.
When autonomous physical tasks lengthen the way software tasks have, the steeper region of the Jones curve comes into view. Until then, the mixed path remains the coherent planning case: solid demand growth, early restructuring in digital occupations, and a binding constraint that still runs through human hands on job sites, in homes, and in care settings.
While we each have access to 100 million times more transistors on our desktop computer than people in the 1970s, we are not 100 million times more productive.
Jones has used that line to illustrate the weak-link constraint in talks and writing. The same logic applies to the broader economy until robots close the physical gap.
For capital allocation and hiring the implication is straightforward. Expect a few extra tenths of a percentage point on productivity and demand for several years. Keep watching the robot installation numbers and the length of autonomous physical tasks. When those move decisively, the steeper part of the Jones curve comes into view.
