AI as a General-Purpose Technology — What the Productivity J-Curve Reveals── Steam took 60 years, electricity 40 — AI productivity gains are still at the bottom of the J-curve

01AI investment has passed $1 trillion, but productivity statistics have barely moved

According to Goldman Sachs, global AI-related investment will exceed $1 trillion in 2026, with the United States accounting for $581 billion. Corporate AI adoption is accelerating: across OECD countries, the share of firms using AI rose from 8.7% in 2023 to 20.2% in 2025. In the United States, roughly 78% of firms, weighted by employment, had adopted AI by November 2025.

Yet the OECD's 2026 Compendium of Productivity Indicators describes AI's macroeconomic impact as showing only "early, tentative signals." Goldman Sachs projects that generative AI could eventually raise global GDP by 7%, roughly $7 trillion, and lift productivity growth by 1.5 percentage points annually over a decade. McKinsey estimates the annual value of generative AI alone at $2.6 to $4.4 trillion across 63 use cases. The projections are large. The evidence in the statistics is not.

In 1987, economist Robert Solow wrote: "You can see the computer age everywhere but in the productivity statistics." Forty years later, AI presents the same picture.

02General-purpose technologies reshape entire economies, but the process takes decades

Economics has a precise concept for technologies that transform the entire economy: the general-purpose technology, or GPT. Timothy Bresnahan and Manuel Trajtenberg formalized the idea in 1995. A technology qualifies as a GPT if it meets three conditions: it is used across a wide range of industries; it improves over time; and it triggers complementary innovations in technology and organization. Steam, electricity, and semiconductors meet all three. So does AI.

The defining feature of every GPT is a long lag between adoption and measurable productivity gains. The theoretical framework that explains this lag is the "productivity J-curve," published by Erik Brynjolfsson, Daniel Rock, and Chad Syverson in the American Economic Journal: Macroeconomics in 2021. When a new GPT arrives, firms invest heavily not in the technology itself but in the complementary changes needed to use it effectively. These investments do not register as output in GDP statistics. As a result, measured productivity actually declines in the early phase of adoption.

03Steam took 60 years and electricity 40 to show up in productivity data

James Watt patented his improved steam engine in 1769. But the effect of steam on British productivity statistics did not become clear until the 1830s — a lag of roughly 60 years. Early steam engines were confined to pumping water out of coal mines. Spreading to textile mills and railways required the development of high-pressure engines and the expiration of Watt's patent in 1800.

The pattern repeated with electricity. Thomas Edison opened the Pearl Street power station on September 4, 1882, initially serving 82 customers with 400 lamps. Yet the productivity effect of electrification in American manufacturing did not appear in the data until the 1920s. Paul David analyzed this roughly 40-year lag in his 1990 paper "The Dynamo and the Computer." Early factories simply replaced the central steam engine with a central electric motor, preserving the old layout. Productivity gains came only after factories adopted the "unit drive" approach — installing individual motors on each machine and redesigning the entire factory floor. Replacing the technology was not enough; the work itself had to be reorganized around it.

04The J-curve is driven by invisible complementary investment

Brynjolfsson and colleagues locate the cause of the lag in complementary intangible investment. When firms adopt a new GPT, they must also redesign business processes, restructure organizations, retrain workers, and build new data infrastructure. These investments are poorly captured in national accounts — they show up as costs, not assets.

In the early phase, measured productivity therefore falls: the downstroke of the J. Once complementary investments bear fruit and new processes take hold, productivity rises sharply: the upstroke. According to Brynjolfsson, Rock, and Syverson's estimates, adjusting for intangible assets related to computer hardware and software, U.S. total factor productivity was 15.9% higher than official figures by the end of 2017. Official statistics had substantially underestimated the productivity gains from the computer era. The same dynamic may be unfolding with AI.

The Acemoglu (2025) estimate reviewed by the IMF takes a conservative position: AI will produce total factor productivity gains of roughly 0.7% over ten years, about 0.07 percentage points per year. Aghion and Bunel (2024) arrive at far higher figures, noting that 60% of tasks in developed economies are exposed to AI. The gap between these two estimates reflects genuine uncertainty about how quickly complementary investments will convert potential into measured output.

05U.S. productivity data from 2024–2025 show early signs of change

If the J-curve theory is correct, the current early phase of AI adoption should correspond to the downstroke or the bottom of the curve. But U.S. labor productivity data tell a somewhat different story. According to the Bureau of Labor Statistics, nonfarm business labor productivity rose 3.0% in 2024, more than double the 2010–2019 average. In 2025, the annual figure was 2.1%, still well above the prior decade's trend.

Attributing this improvement to AI, however, would be premature. The OECD's 2026 report uses the phrase "early, tentative signals" and notes that a causal link to AI has not been established. Goldman Sachs projects that AI will begin to have a measurable effect on U.S. GDP in 2027. In Europe, labor productivity grew only 0.2% in 2024, and the correlation between AI adoption progress and productivity growth varies sharply across countries.

06Healthcare and drug development show AI productivity effects ahead of other sectors

The productivity effects of a GPT tend to appear first in industries where information processing is central and where existing processes carry large inefficiencies. Healthcare and drug development meet both criteria.

In drug development, the traditional process takes an average of 12 years, at an estimated cost of approximately $2.6 billion per approved compound. AI-driven discovery programs have shortened the discovery phase from an industry average of 42 months to 18 months in documented cases. As of 2026, more than 170 AI-originated drug candidates are in clinical development, with 15 to 20 expected to enter late-stage trials during the year.

Change in clinical practice is also rapid. Among U.S. physicians, 81% reported using AI professionally in 2026, more than double the 38% in 2023. Hospitals using AI for billing rose from 36% in 2023 to 61% in 2024. Task-level productivity studies reviewed by the IMF found efficiency gains of 20–60% in controlled experimental settings and 15–30% in real-world work environments.

However, task-level efficiency gains do not automatically translate into organization-wide or macroeconomic productivity. Here too the J-curve intervenes. Even if individual physicians speed up diagnoses with AI, reimbursement structures and hospital operating models must also change for the gains to register in aggregate statistics.

07Organizations at the bottom of the J-curve need complementary investment now

The practical lesson from steam and electricity is clear: investment in the technology alone does not produce productivity gains without complementary investment. Many organizations today stand at the same point as the factories of 1900 that installed electric motors but did not rearrange the factory floor. The OECD projects that AI-driven annual labor productivity gains could reach 0.4 to 1.3 percentage points in high-exposure economies such as the United States and the United Kingdom — but only if the complementary conditions are met.