TOP

LOGIN
60, 30, 15, … 7.5?
What History Suggests About Generative AI
LH General Industry, PC General Industry
December 02, 2025 Frank Schmid
English Deutsch

General-purpose technologies (GPTs) typically don’t move measured productivity immediately on arrival. The usual sequence is invention → diffusion → complement-building (data, processes, skills, governance) → reorganization of production that finally yields measurable gains. During those middle stages, companies spend on intangibles that accounting treats as expenses rather than capital, so the statistics lag the underlying productive capacity. Economic models describe a J‑curve in measured productivity: a dip while complements are built, followed by an upswing once redesigned processes begin to compound.1

The influential study by Bresnahan and Trajtenberg (1992) puts GPTs at the center of the history of economic development. They argue that whole eras of technical progress and economic growth appear to be driven by GPTs, which they define as technologies that are pervasive (used as inputs by many downstream sectors), have inherent potential for technical improvement, and exhibit “innovational complementarities,” meaning that the productivity of Research & Development in downstream sectors rises as the GPT improves. As GPTs diffuse and advance, they spread through the economy and bring about generalized productivity gains.2 This lens captures both the eventual power of GPTs and the long, complement-intensive road to realizing broad-based productivity gains.

Steam, electricity, and digital computing all fit this pattern.3 The familiar shorthand – steam 60 years, electricity 30 years, computing 15 – summarizes the time between core invention and the start of clear macro-level payoff. The question for generative AI is whether that lag can shrink again, perhaps toward seven or eight years, and what it would take for that to happen.

Steam – Invention First, Productivity Much Later

Dating the steam story to James Watt’s separate condenser (1769), the economy-level payoff arrived much later. For decades steam was chiefly stationary power (pumping; then mill drive via rotative engines) until high-pressure designs broadened its scope. Growth-accounting evidence shows steam’s measured contribution to British productivity is minimal before the 1830s; larger effects appear only in the mid- to late-19th century, as steam spreads across industry and transport – especially with the diffusion of locomotives and railways and the rise of large steam-driven mills and weaving sheds – supported by complements such as tracks and depots, signaling and timetables, maintenance shops, and new skills. In short, the engine preceded the economy by roughly six decades – a textbook GPT productivity lag.4

Electricity – A Motor in the 1890s, a Payoff After Factory Redesign

By the 1890s, core electrical technologies – generators, distribution, and the alternating-current induction motor – were established. Yet the big productivity response does not appear until the 1920s–1930s, once factories abandon the steam-era, line-shaft layout. Simply replacing a central steam engine with a single large electric motor did little; the gains arrived when plants reorganized around unit drives (small motors at each machine), enabling one-story factories, flexible flows, and tighter control. Electrification’s payoff was governed by organizational redesign and took decades.5

Computing – A Mid‑1990s Break After Process Replatforming

The personal computer of 1981 is an emblem, but the decisive break in productivity occurs after 1995. Industry-level and growth-accounting studies attribute much of that acceleration to both the production and the use of information and communication technology, with especially strong gains in industries that replatformed workflows, supply chains, and retail formats around digital systems. Buying hardware was not enough; use mattered – meaning redesigned processes and skills to match.6

The Importance of Organizational Capital – Firm-level Evidence from the Computerization Era

In a study of 1,216 publicly traded U.S. firms over the period 1987‑1997, Brynjolfsson, Hitt, and Yang (2002) set out to quantify the organizational capital that complemented computer assets. They treated organizational capital as an intangible, proxied by the firm’s ratio of market value to book value (Tobin’s q), which reflects investments in human capital and organizational traits. Firms that co‑invested in organizational innovation – hiring, training, new processes – alongside computer assets were estimated to generate on the order of $10 in organizational capital for every dollar of computer assets on the balance sheet. In other words, tech alone did not explain the gains; aligning organizational practices with IT investments was closely associated with higher firm value.7

This is exactly what the GPT framework would predict: digital technologies acted as a pervasive, continually improving platform, and firms that built the right complements saw the largest payoff.

Why the Lag Shortened – and Why Mainly in the Digital Era

The shorthand “steam 60 years, electricity 30, computing 15” captures a real pattern, but the causes differ. The gap between steam and electricity reflects context and dating conventions more than a structural change: both waves were constrained by the slow work of organizational redesign and infrastructure build-out; electricity did not eliminate those frictions.8

The mechanism-driven shortening appears with ICT (Information and Communication Technology). Here the Bresnahan–Trajtenberg characteristics of a GPT become especially visible:

  • Pervasiveness and modularity. Standardized software stacks, interoperable components, and network protocols made digital technology an input into a wide range of downstream activities.
  • Inherent potential for improvement. Rapid, sustained performance gains (and price declines) in computing hardware and software kept opening new application areas.
  • Innovational complementarities. As ICT improved, it raised the productivity of Research & Development and process innovation in user industries, especially where organizations experimented with new ways of working.

Improved statistical methods for capturing quality change and rapid price declines in digital goods also helped the macro data register productivity gains sooner.9

Generative AI Through the Historical Lens

Generative AI fits this template: rapid invention and diffusion, but complements remain the bottleneck. The modern equivalents of tracks and unit drives are data quality and lineage, retrieval and context design, evaluation and monitoring, safety and governance, and a workforce trained to integrate the capability into end-to-end flows.10 Historical perspectives similarly frame AI as a GPT with complement-intensive diffusion and nontrivial productivity lags.11

Early evidence is compelling at the task level. In a preregistered study with mid-career professionals, access to a generative assistant reduced time to complete writing tasks by about 40% and improved quality by about 18%.12 In a large, real-world rollout to more than 5,000 customer-support agents, a conversational assistant raised issues resolved per hour by about 14% on average – especially for less experienced workers – consistent with codification and diffusion of best practices.13 These are micro-level gains, not macro outcomes, but they are the raw material from which productivity surges are built.14

Could the lag shrink to seven or eight years? Possibly – but only if firms treat complement-building as the main event: substantial up‑front investment, rapid process rewiring, and careful measurement tying AI system performance to business outcomes. The lesson from the 1990s is clear: productivity moved when organizations replatformed entire processes, not when they merely accumulated tools.15

General-Purpose Productivity Gains Take Time to Arrive16

GENERAL-PURPOSE TECHNOLOGY

CORE TECHNOLOGY ESTABLISHED

PRODUCTIVITY PAYOFF

TIME LAG (YEARS)

WHY THE TIME LAG?

Steam

1769 (Watt’s condenser)

1830s+

~60

Needed factory redesign, rail infrastructure, and new skills

Electricity

1890s (alternating-current induction motor, electric grids)

1920s–1930s

~30

Gains came after unit drives, one-story factory layouts, and more flexible flows

Computing

1981 (personal computer as emblem of ICT diffusion)

Post-1995

~15

Required digital workflows, reconfigured supply chains, and complementary skills

Generative AI

Nov. 2022 (release of ChatGPT)

2030?

~7.5

Requires data foundation, workflow redesign, and organizational transformation

Closing

The long road to productivity is not a defect; it is the price of transformation. Steam, electricity, and digital computing teach the same lesson that Bresnahan and Trajtenberg articulated for GPTs more broadly: pervasive, ever-improving technologies with strong innovational complementarities reshape entire eras of growth – but only once organizations learn how to build and use the right complements. Generative AI will be no exception. The challenge – and the opportunity – is to bend the curve faster than history suggests.

  1. See Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J‑Curve,” American Economic Journal: Macroeconomics Vol. 13, No. 1, 333–372, 2021. Working paper (October 2018, revised January 2020) at https://www.nber.org/system/files/working_papers/w25148/w25148.pdf.
  2. See Timothy F. Bresnahan and Manuel Trajtenberg, “General Purpose Technologies: ‘Engines of Growth’?,” NBER Working Paper No. 4148, August 1992, https://www.nber.org/system/files/working_papers/w4148/w4148.pdf.
  3. Bresnahan and Trajtenberg, op. cit., identify as general-purpose technologies the steam engine, the electric motor, and semiconductors & computers.
  4. See Nicholas Crafts, “Steam as a General Purpose Technology: A Growth Accounting Perspective,” The Economic Journal Vol. 114, No. 495, 338–351, 2004.
  5. See Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review Vol. 80, No. 2, 355–361, 1990.
  6. See Dale W. Jorgenson and Kevin J. Stiroh, “Information Technology and Growth,” American Economic Review Vol. 89, No. 2, 109–115, 1999; Stephen D. Oliner and Daniel E. Sichel, “Information Technology and Productivity: Where Are We Now and Where Are We Going?” Journal of Policy Modeling Vol. 25, Issue 5, 477–503, 2003, working paper (May 2002) at https://www.federalreserve.gov/pubs/feds/2002/200229/200229pap.pdf; Erik Brynjolfsson and Lorin M. Hitt, “Beyond Computation: Information Technology, Organizational Transformation and Business Performance,” Journal of Economic Perspectives Vol. 14, No. 4, 23–48, 2000; and Erik Brynjolfsson, Lorin M. Hitt, and Shinkyu Yang, “Intangible Assets: Computers and Organizational Capital,” Brookings Papers on Economic Activity 2002(1), 137–199.
  7. See Erik Brynjolfsson, Lorin M. Hitt, and Shinkyu Yang, “Intangible Assets: Computer and Organizational Capital,” Brookings Papers on Economic Activity 2002(1), 137–181, https://www.brookings.edu/wp-content/uploads/2002/01/2002a_bpea_brynjolfsson.pdf.
  8. See Paul A. David, op. cit., and Nicholas Crafts, op. cit.
  9. See Dale W. Jorgenson and Kevin J. Stiroh, op. cit., and Stephen D. Oliner and Daniel E. Sichel, op. cit.
  10. See Erik Brynjolfsson, Daniel Rock, and Chad Syverson, op. cit., and Erik Brynjolfsson and Lorin M. Hitt, op. cit.
  11. See Nicholas Crafts, “Artificial Intelligence as a General-Purpose Technology: An Historical Perspective,” Oxford Review of Economic Policy Vol. 37, No. 3, 521–536, 2021.
  12. See Shakked Noy and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science Vol. 381, Issue 6654, 187–192, 2023.
  13. See Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” NBER Working Paper 31161 (April 2023, revised November 2023), https://www.nber.org/papers/w31161.
  14. See Dale W. Jorgenson and Kevin J. Stiroh, op. cit., and Stephen D. Oliner and Daniel E. Sichel, op. cit.
  15. Ibid.
  16. See Paul A. David, op. cit.; Dale W. Jorgenson and Kevin J. Stiroh, op. cit.; Stephen D. Oliner and Daniel E. Sichel, op. cit.; and Erik Brynjolfsson and Lorin M. Hitt, op. cit.

Frank Schmid

Chief Technology Officer

See All Articles