Ten to twenty years: why technology's payoff always comes late

Computers took about twenty years to show up in the productivity statistics; in 1987 the economist Robert Solow could still joke that the computer age was visible everywhere except there. Electricity was slower still: about forty years passed between the first power stations and the point where half of American factory power was electric, and the big productivity jump came only in the 1920s. In 2026, firms have so far blamed AI for 121,401 US job cuts this year. History suggests they are cutting before the payoff, not because of it.

Contents · 10 sections

TL;DR #

• New general-purpose technologies pay off late: electricity took decades to reshape factories, and computers took about twenty years to show up in productivity statistics.
• The reason is reorganisation. The gains come only after firms redesign work, processes and skills around the new tool.
• Economists call the pattern the productivity J-curve: measured productivity dips while firms invest in things national accounts barely count, then rises.
• Measured AI displacement in 2026 is still small. Cutting the people who would redesign the work may delay the payoff.

This post is about that gap, why it exists, and what it means for the claim that AI is replacing workers now.

The dynamo that didn’t pay #

The best story about the lag is the electric factory. In my library, Charles R. Morris tells it well in A Rabble of Dead Money, his history of the 1920s boom and crash. American factories at the turn of the century were impressive, but:

“Productivity was drastically limited by their dependence on centralized coal-fired steam power connected to working machinery by belts and shafts.”

The first electric motors simply replaced the steam engine in the basement. The belts and shafts stayed. The building, the layout and the work stayed. So did the productivity. The gain came only when factories were redesigned around electricity. Morris again:

“The final step came mostly in the 1920s with the adoption of ‘unit-drive’ layouts, in which each machine had its own electrical power source.”

Single-storey plants, machines placed in the order of the work, and the assembly line followed from that. Morris adds that the car industry “provided the critical demonstration of the productivity gains from electric factories.”

The economic historian Paul David made this the classic case in a 1990 paper, “The Dynamo and the Computer”. His point was that the delay was mostly the slow pace of factory redesign. It did not pay to rip out plants that still worked. You waited until they wore out, or until a new competitor built from scratch. According to Morgan Stanley’s chief US economist, as quoted by Fortune this spring, electrification ended up doubling output per hour in US non-farm business between 1900 and 1929. But most of that came at the end of the period, not the start.

The same story, three more times #

This was not a one-off.

• Railways. In my library, Roger Bootle’s The AI Economy notes, almost in passing, that the economic impact of the railways “came decades later” than the engineering.
• Computers. Solow’s 1987 remark is quoted in the same book. US productivity growth finally picked up from the mid-1990s, when firms rebuilt their processes around networks and databases. That was some twenty years after mainframes became common in big offices.
• The internet. The Morgan Stanley team puts it at labour productivity growth rising from about 1.5% to nearly 3% a year by 2000, years after the web arrived.

Bootle also reminds us what the waiting years felt like for workers in the first industrial revolution: “during this period of ‘progress,’ large numbers of people underwent immense suffering.” Economic historians call that stretch, when output rose but wages stalled, “Engels’ pause”.

Why the curve dips first: the J-curve #

In 2021, Erik Brynjolfsson, Daniel Rock and Chad Syverson gave the lag a shape and a name: the productivity J-curve, published in the American Economic Journal: Macroeconomics.

The mechanism is simple. A general-purpose technology such as electricity, computing or AI only pays off after firms invest in things that are hard to see:

• new processes and workflows;
• new business models;
• new skills, which means people learning on the job;
• data and the organisation around it.

These are intangible investments. National accounts barely count them. So in the early years, measured productivity looks worse than reality, because people and money go into building something that doesn’t show up as output yet. Later, when those hidden assets start producing, measured productivity looks better than reality. Hence the J: a dip, then a rise.

Their numbers for the last wave are striking. Counting the intangible investment tied to computer hardware and software, they estimate that US total factor productivity was 15.9% higher than official figures showed by the end of 2017. The payoff was real. It was just late, and the costs came first.

What this means for “AI is replacing workers now” #

Put the J-curve next to 2026 and three things stand out.

First, the measured displacement is small. In its August 2026 update, the Stanford Digital Economy Lab team behind the “canaries” paper wrote: “We do not see widespread, economy-wide job displacement associated with AI.” Goldman Sachs economists estimate AI has added about 0.1 percentage point to US unemployment so far, according to Fortune. And in October, according to Challenger, AI was cited for only 1,265 of 43,994 announced US cuts, under 3%.

Second, a lot of “AI layoffs” are cost cuts with a new label. Australian unions have accused Qantas and the big banks of “AI-washing”, and a parliamentary inquiry into AI and offshore job cuts is under way. When Qantas moves 850 head-office jobs to Accenture, mostly delivered from India, that is the 2000s offshoring template with an AI sticker on it.

Third, cutting people during the dip removes the people who would build the payoff. This is the part I worry about. The J-curve says the gains come from co-invention: staff who know the old process well enough to redesign it around the new tool. That knowledge sits in people’s heads. If firms cut those people in year two to show savings, they may be removing the very intangible capital that would produce the gain in year eight.

There is an old warning about this in my library. Wayne Cascio and Clifford Young, writing in Resizing the Organization, followed S&P 500 firms from 1982 to 2000 and concluded that “it was not possible for firms to ‘save’ or ‘shrink’ their way to prosperity.” The firms that grew by investing did better than both the cutters and the stable employers.

The counter-argument, fairly stated #

AI may move faster than electricity did. You do not need to rebuild a factory to use a language model; you rent it by the token. Adoption is far quicker than it was for dynamos or mainframes.

I accept that adoption is faster. But the J-curve is not about adoption. It is about reorganisation, and organisations change at the speed of people, contracts, regulation and habit. Electric motors were also available quickly; the redesign of the work took a generation. Software is easy to install. Changing who decides what, who checks what, and who is liable when it goes wrong is not.

Honestly, we don’t know yet whether AI’s dip will be five years or fifteen. Nobody does. Anyone who tells you the payoff is here now is reading company press releases, not the productivity statistics.

What to watch, 2027–2030 #

If you want to know when the payoff is real, ignore layoff announcements and watch these instead:

1. Productivity growth in official statistics, sustained over several years, not one good quarter.
2. Intangible investment: spending on software, training and process redesign, not only on data centres.
3. New products and processes: firms selling things they could not sell before, not the same things with fewer people.
4. Hiring of people who redesign work: process engineers, trainers, domain experts working alongside the tools.

When those move together, the curve will be turning up. Until then, I read most “AI replaced them” announcements as what the J-curve predicts in the dip: costs first, gains later, and a strong temptation to claim the gains early.

What I don’t know #

The 10-to-20-year range comes from a handful of historical cases, and history does not repeat exactly. The J-curve estimates rest on assumptions about how to value intangible assets. The 2026 figures are announcements and surveys, not audited accounts. And LSN News, my news engine, covers only the days since 7 November, so I can’t use it to measure a trend over years.

Frequently asked questions #

What is the productivity J-curve? #

It is the pattern, described by Brynjolfsson, Rock and Syverson, in which measured productivity first dips and then rises after a new general-purpose technology arrives. Firms spend money and staff time on new processes, skills and data, which national accounts barely count, so the gains show up years later.

How long did electricity take to raise factory productivity? #

Decades. Factories first swapped steam engines for electric motors and kept the old layout. The big gains came in the 1920s, when plants were redesigned so that each machine had its own motor.

Is AI replacing workers now? #

Not at a scale the statistics can see yet. The Stanford team behind the “canaries” study says it does not see widespread, economy-wide displacement, and AI was cited for under 3% of announced US cuts in October 2026. Specific groups, such as young workers in exposed jobs, are affected.

Won’t AI spread faster than electricity did? #

Adoption is faster, because you can rent a model rather than build a factory. But the payoff depends on reorganising work, and organisations change at the speed of people, contracts and regulation. How long AI’s dip will last, nobody knows yet.

The lesson #

The gains from a general-purpose technology arrive after the work is redesigned, not when the tool is installed. Firms that cut the people who know the old process, in order to show savings early, may be removing the very knowledge that would produce the gain later. When someone says AI is replacing workers now, ask to see the productivity statistics, not the press release.