When Will Artificial Intelligence Make Us Richer? The Productivity Puzzle
By Matthew Parish
Sunday 13 September 2026
There is something peculiar about the artificial intelligence revolution. Almost everybody who uses the most sophisticated forms of artificial intelligence can see that the technology is extraordinary. A competent large language model can draft documents, analyse data, write computer code, translate languages, summarise enormous quantities of material, undertake research and perform intellectual tasks in minutes that might previously have occupied a human being for hours or even days. Yet if artificial intelligence is already capable of performing all these things, one might reasonably ask why the developed world’s economic statistics do not reveal an equally extraordinary productivity revolution. The answer may be that we are asking the question too soon. The history of technological revolutions suggests that the invention of a transformative technology and the appearance of its full consequences in national productivity statistics are two very different events.
The apparent contradiction is an old one. In 1987 the economist Robert Solow famously observed that one could see the computer age everywhere except in the productivity statistics. Computers were proliferating across offices and factories, enormous sums were being spent upon information technology and businesses insisted that computerisation was transforming the way they worked. Nevertheless aggregate productivity growth remained disappointing. Eventually the paradox substantially resolved itself. Computers became connected to networks, databases became standardised, businesses reorganised themselves around electronic information and, above all, the internet transformed the computer from an isolated calculating device into a communications infrastructure. Productivity improvements followed. Artificial intelligence may now be passing through an analogous period in which the technology is conspicuous but its macroeconomic consequences remain frustratingly difficult to measure.
Indeed the history goes back much further. The steam engine did not instantaneously create the Industrial Revolution. Early steam engines were expensive, inefficient and useful principally for particular applications such as pumping water from mines. It took decades of engineering improvements, capital investment and complementary innovations before steam power transformed manufacturing and transport. Electricity provides an even more striking example. Installing an electric motor in a nineteenth-century factory designed around steam power did not necessarily produce spectacular productivity improvements. The factory itself had to be redesigned. Steam-powered factories tended to arrange machinery around central shafts and belts because mechanical energy had to be transmitted physically throughout the building. Electric motors eventually permitted machinery to be distributed according to the logic of production rather than the geometry of power transmission. Factories became lighter, cleaner, safer and radically more efficient. But these benefits depended upon organisational innovation as much as upon electricity itself.
This historical phenomenon is sometimes described as the productivity J-curve. Businesses initially spend money adopting a new technology while simultaneously incurring the costs of reorganising themselves around it. Measured productivity may therefore stagnate or even decline before subsequently accelerating. Research into industrial artificial intelligence in American manufacturing has already found evidence consistent with precisely such a pattern: adoption can impose short-term costs while earlier adopters subsequently demonstrate stronger growth. The Federal Reserve has similarly emphasised that earlier general-purpose technologies required changes in business practices, management and worker skills before their full benefits became apparent.
Artificial intelligence fits unusually well into this historical category of what economists call a “general-purpose technology”: an innovation capable of improving repeatedly, spreading throughout much of the economy and generating further innovations around itself. Steam power, electricity and information technology were general-purpose technologies. Artificial intelligence may be an unusually powerful example because its principal economic function is not merely to perform a particular physical operation more cheaply. It potentially reduces the cost of cognition itself. If steam multiplied human muscle and electricity allowed energy to be transported efficiently, artificial intelligence may multiply some categories of human thought. That distinction matters because thinking, analysing, designing, calculating, communicating and organising are inputs into virtually every sophisticated economic activity. Historical research on general-purpose technologies consequently suggests that AI’s most important productivity effects may still lie ahead.
There is already persuasive evidence of substantial productivity improvements at the level of particular tasks. One influential study of more than 5,000 customer-support workers found that access to a generative AI assistant increased productivity by approximately 14 per cent on average. Interestingly, the improvements were substantially greater among less experienced and lower-performing workers, suggesting that artificial intelligence may operate partly by distributing the accumulated knowledge and techniques of the most competent employees across an organisation. Another large field experiment involving more than 7,000 knowledge workers across 66 firms found that workers using generative AI spent approximately two fewer hours each week on email. These are not hypothetical predictions about some distant superintelligence. They are measurable effects of technologies that already exist.
Nevertheless there is a formidable distance between making individual tasks faster and making an economy substantially more productive. Recent research into AI-assisted software development illustrates the distinction particularly well. AI coding tools can produce spectacular increases in coding activity, but those gains diminish as one moves along the chain from writing code towards producing finished software that people actually use. A 2026 study of more than 500,000 developers found large increases in commits associated with increasingly capable AI coding systems but dramatically smaller increases in projects and actual releases. The bottleneck had simply moved. Computers could generate code more rapidly but human beings still had to supervise, integrate, test, approve, market and deploy the resulting products.
This is perhaps the central economic problem confronting artificial intelligence in 2026. A lawyer who can draft a memorandum four times faster does not necessarily cause a law firm to produce four times as much economically valuable legal work. The memorandum must still be checked, strategic decisions must still be made, clients must still be obtained and judges and counterparties remain stubbornly human. A doctor who can analyse medical records more rapidly cannot necessarily treat four times as many patients because hospitals, operating theatres, nurses, medicines and regulatory procedures remain constrained. An architect who can produce twenty designs in the time formerly required to produce one does not cause buildings to be constructed twenty times faster. The productivity of one component of a production process may rise enormously while the productivity of the entire process rises only modestly.
Artificial intelligence therefore creates what might be called migrating bottlenecks. If drafting becomes virtually instantaneous, checking becomes the bottleneck. If computer programming becomes virtually instantaneous, product design and testing become bottlenecks. If scientific literature reviews become instantaneous, laboratory experiments remain slow. If administrative analysis becomes instantaneous, managerial decision-making may become the constraint. This phenomenon explains how extraordinary improvements in the productivity of individual intellectual tasks can coexist with relatively ordinary improvements in measured national productivity. The machine accelerates one section of an economic pipeline until it collides with the next section that has not yet been accelerated.
There is another difficulty. Businesses have scarcely begun reorganising themselves around artificial intelligence. Most organisations currently use AI as an appendage to institutions designed for a world without it. Employees use language models to draft emails, summarise meetings, prepare presentations or write fragments of computer code, but the underlying organisational structures remain largely intact. Companies continue to have approximately the same departments, management hierarchies, approval procedures and administrative processes. In this sense contemporary businesses may resemble factories that replaced a steam engine with an electric motor while leaving all the old belts and shafts in place. The transformative productivity effects will arise only when businesses cease asking how AI can assist existing jobs and begin asking how organisations ought to be designed if inexpensive machine intelligence is assumed to exist from the outset.
This process may already be beginning, but adoption remains surprisingly shallow. An August 2026 study based upon a nationally representative American survey found that generative AI was being used across a wide range of occupations and tasks, yet within most of those occupations fewer than half of workers were actually using it. Exposure to AI and actual adoption are therefore quite different things. This matters enormously. A technology cannot transform national productivity merely because it exists. It must become embedded within ordinary economic behaviour.
There are nevertheless tentative indications that something may be changing. The OECD’s 2026 productivity survey observed that recent American productivity performance, together with multifactor productivity improvements in some service industries, might constitute early signals consistent with AI beginning to have a measurable effect, although the evidence remains far from conclusive. American non-farm labour productivity was 2.2 per cent higher in the second quarter of 2026 than a year earlier. It would be premature to attribute this performance principally to artificial intelligence, but it is becoming increasingly difficult to maintain the stronger proposition that AI cannot materially affect productivity at all.
The more interesting question is therefore one of timing. Here the historical evidence suggests neither instantaneous transformation nor the need to wait half a century. Steam power and electricity required decades before their economic potential became fully apparent. Information technology moved faster. Artificial intelligence should probably move faster again because the infrastructure necessary for its diffusion already exists. Billions of people possess internet-connected computers or smartphones. Software can be distributed globally almost instantaneously. Cloud computing allows businesses to adopt capabilities without constructing their own computing centres and employees already possess decades of experience adapting to digital technologies. The complementary infrastructure through which AI can spread is therefore dramatically more mature than the infrastructure available during previous technological revolutions.
The OECD’s modelling provides a useful indication of the plausible scale. Its estimates suggest that artificial intelligence might add roughly 0.4 to 1.3 percentage points to annual labour-productivity growth over the coming decade in highly exposed economies such as the United States and United Kingdom, with smaller gains elsewhere depending upon economic structure and adoption rates. Other studies generate wider ranges, which itself demonstrates the extraordinary uncertainty surrounding the subject. An additional percentage point of productivity growth may sound unimpressive beside claims that artificial intelligence will transform civilisation, but compounded across an economy it is enormous. An economy growing at three per cent rather than two per cent doubles approximately a generation earlier.
A reasonable central scenario might therefore divide the AI productivity revolution into three phases. The first, extending roughly from the emergence of mass-market generative AI in 2022 until the middle or latter part of this decade, is principally a period of experimentation. Individuals discover spectacular applications, particular professions experience substantial task-level productivity improvements and corporations spend enormous sums learning what works. Yet aggregate productivity effects remain comparatively modest because organisations have not been redesigned and bottlenecks elsewhere in production constrain the value of accelerated intellectual work.
The second phase, plausibly running from approximately 2027 or 2028 through the early 2030s, may be considerably more important. AI agents are likely to become increasingly capable of completing sequences of tasks rather than merely producing individual answers. Corporate databases will increasingly be structured for machine use. Business software will be redesigned around AI rather than merely incorporating AI buttons into pre-existing applications. Managers will acquire experience in supervising combinations of human and machine labour. Entire administrative processes may be automated rather than individual components of them. If a substantial macroeconomic productivity acceleration is going to emerge from the current generation of artificial intelligence, this is the period in which one might reasonably expect it to become unmistakable.
The third phase is more speculative but potentially far more consequential. Artificial intelligence differs from electricity because AI may accelerate the process through which further technologies are invented. If machine intelligence materially improves scientific research, pharmaceutical development, engineering, mathematics, materials science and the design of subsequent generations of artificial intelligence itself, then AI ceases merely to be another productivity-enhancing technology. It becomes a technology for increasing the productivity of technological progress. Federal Reserve researchers have identified precisely this question — whether generative AI ultimately resembles a temporary productivity-enhancing invention or a general-purpose technology capable of generating continuing waves of innovation — as central to understanding its long-run economic significance.
Were this second-order effect to become substantial during the 2030s, historical comparisons might begin to break down. Electricity made factories more productive, but electricity did not itself conduct experiments intended to discover better forms of electricity. A sufficiently sophisticated artificial intelligence system might assist engineers designing better processors, scientists developing new energy technologies and computer scientists constructing better artificial intelligence systems. The productivity consequences would then become recursive. More capable AI would accelerate research; accelerated research would create more capable technology; better technology would increase productivity and provide greater resources for further research. This possibility explains why apparently modest estimates of AI’s effects over the next decade should not necessarily be interpreted as estimates of its ultimate economic importance.
There are powerful reasons for caution. Some tasks are intrinsically physical. Human beings still have to build houses, repair roads, nurse elderly people, harvest crops, manufacture machinery and transport goods. Regulation may slow adoption in medicine, law, finance and government. Artificial intelligence may generate convincing errors that impose substantial verification costs. Cheap intellectual production may create enormous quantities of low-quality material through which human beings must sift. Organisations may respond to increased productive capacity not by producing more but by multiplying meetings, reports and bureaucracy. Jevons-like effects are also possible: when producing analysis becomes cheaper, organisations may simply demand vastly more analysis. Productivity improvements measured at the task level can consequently disappear into expanded expectations.
Distribution presents another complication. Productivity and prosperity are not synonymous. A technology might increase national output substantially while concentrating its gains among the owners of capital, computing infrastructure and intellectual property. Previous industrial revolutions eventually produced enormous increases in living standards but their transitional periods could be economically brutal. Artificial intelligence may simultaneously increase aggregate productivity, destroy particular categories of employment and depress the market value of skills that required decades to acquire. An economy can become richer while substantial numbers of its citizens temporarily become poorer.
Yet the historical lesson should ultimately incline us towards qualified optimism about productivity itself. Every major general-purpose technology has initially been misunderstood because observers naturally imagine the new technology operating inside the old economy. The more profound effects emerge when the economy reorganises itself around the technology. Railways did not merely make horse journeys faster; they reorganised cities, markets and industrial geography. Electricity did not merely replace steam engines; it transformed the architecture of factories and eventually households. Computers did not merely replace typewriters; they created networks, digital markets and entirely new industries.
Artificial intelligence will probably follow the same pattern. At present we are mostly using machines to perform existing intellectual tasks somewhat faster. That is useful but it is unlikely to represent the principal economic consequence of machine intelligence. The larger productivity revolution will begin when businesses, governments and ultimately societies redesign their institutions upon the assumption that many forms of cognition are abundant and inexpensive.
If history is a useful guide, therefore, the answer to the productivity puzzle is that artificial intelligence is probably already raising productivity, but not yet by an amount remotely commensurate with the technological spectacle surrounding it. The strongest evidence currently appears at the level of particular workers and tasks rather than whole economies. The transition from one to the other requires complementary investment, institutional reconstruction, worker adaptation and the elimination of bottlenecks. These things take years rather than months.
Nevertheless the waiting period may be shorter than it was for steam or electricity. A plausible expectation is that measurable AI-related productivity gains will continue accumulating through the remainder of the 2020s, that a significant macroeconomic effect may become apparent around the end of this decade and that the first half of the 2030s will reveal whether artificial intelligence truly belongs alongside electricity and the computer as one of the great general-purpose technologies. The OECD’s estimates of potentially substantial additional annual productivity growth over the next decade are entirely compatible with such a timetable.
Beyond approximately 2035 prediction becomes hazardous because the central question changes. We will no longer merely be asking how much more productive artificial intelligence makes today’s workers. We may instead be asking how much more productive artificial intelligence makes the process of invention itself. If the answer is “substantially”, then the greatest economic effect of artificial intelligence will not be that machines perform our existing work more quickly. It will be that the interval between one technological revolution and the next begins to shrink.
That would be something genuinely new in economic history.
