What the History of Automation Actually Tells Us About AI
The optimists cite history to argue we shouldn't worry. The pessimists cite history to argue this time is catastrophically different. Both are misreading it. The history of automation is more complicated — and more honest — than either side is willing to be.
Founder, Majhi Group & Majhi OS
I grew up in Kalahandi in the 1990s. The economic history of that decade, described from the aggregate India perspective, is a story of liberalisation, growth, and opportunity — the IT sector in Bangalore was building the India that the world would later notice. The economic history of that decade, described from where I was watching it, is a different story. The gains were real. They were elsewhere.
This is what the history of automation actually looks like, if you look at it carefully rather than at the aggregate statistics that make the story tidy.
The optimists cite history to argue we shouldn't worry about AI: technology always creates more jobs than it destroys, living standards always rise, the doomsayers of every previous wave were wrong. The pessimists cite history to argue this time is catastrophically different: AI is broader than any previous automation, the reinstatement won't happen, the end of work is near.
Both sides are misreading the evidence. The history is more honest — and more complicated — than either is willing to be.
The agricultural transition, accurately described
Agricultural mechanization is the example most commonly cited to show that automation creates rather than destroys in the long run. The numbers support the claim: agriculture employed 84% of the US workforce in 1810, 40% by 1900, and under 2% today. Over the same period, average living standards rose dramatically. The workers who left farming found jobs in manufacturing and services. Aggregate employment held.
This is all true. Here is what the summary leaves out.
The transition took 150 years. The workers who were displaced by the mechanical reaper in the 1860s did not retrain as factory workers. They struggled, often in poverty, in communities that were hollowing out. Their children and grandchildren eventually found their way into manufacturing, in different cities, in different industries. The aggregate outcome was good. The experience of displacement was not.
The optimist account of this transition cites the outcome — aggregate employment held, living standards rose — without examining the mechanism that produced it. The reason reinstatement happened is that agricultural productivity released labour at the same time that industrial demand for labour was growing fast enough to absorb it. That was not an automatic consequence of the technology. It was the result of a specific historical conjunction of industrial demand, cheap energy, population growth, and eventually — deliberately — policy.
The computerization wave, accurately described
A second case is the computerisation of office work. Computers were introduced into American offices in the early 1980s. The standard prediction was that clerical employment would fall as machines replaced secretaries, typists, and data entry workers.
The actual sequence was more interesting. The share of secretaries using a computer at work rose from 46% in 1984 to 77% by the end of the decade — rapid and extensive adoption. And yet clerical employment kept growing. The jobs were not immediately eliminated. They were transformed: the same workers were now producing more output using the new tools, and organisations expanded to absorb the increased productivity.
Then, with a lag, the collapse came. Secretaries and administrative assistants had grown to nearly 17% of US employment by 1980. By 2016, the share had fallen back to levels last seen in 1960. The technology arrived in the early 1980s. The employment impact arrived a decade or more later, as organisations restructured around what computers could now handle.
This lag is not a historical accident. It is a structural feature of automation transitions. Technology is adopted into existing organisational forms. The restructuring that reduces headcount comes later, after the technology is embedded and the business model has been redesigned around it. The gap between "AI adoption" and "AI employment impact" is likely to follow the same pattern — which means the disruption we are now seeing in AI adoption is not the disruption we will see in a decade.
The three things the history actually shows
First: the aggregate story and the local story are always different. The numbers that describe overall employment in the long run are not the numbers that describe what happened to the people who were displaced. The transition is always good in aggregate and often brutal in the specific. Every community, industry, and cohort that sits at the intersection of what the new technology displaces faces a real and concentrated cost, while the benefits of the technology distribute diffusely across time and geography.
Second: reinstatement is not automatic — it depends on conditions. MIT economists Daron Acemoglu and Pascual Restrepo, studying the pattern across modern automation, have found that unlike 19th-century mechanization, current automation may not generate an equivalent reinstatement of labour demand. The 19th-century transition worked because industrial demand was expanding fast enough to absorb displaced farm workers. That conjunction is not guaranteed. It is not created by the technology — it is created by the broader economic conditions, investment patterns, and policy choices that accompany it.
Third: the speed of the transition matters enormously. A transition that takes 150 years distributes the disruption across generations. A transition that takes 15 years concentrates it within careers. The speed of AI capability development is faster than any previous automation wave, which means the period for adjustment — the time between displacement and the availability of new roles — is shorter. Shorter adjustment periods produce more acute disruption even if the long-run outcome is equally positive.
The aggregate story and the local story are always different. The numbers that describe overall employment in the long run are not the numbers that describe what happened to the people who were displaced. The transition is always good in aggregate and often brutal in the specific.
What is actually different about AI
Previous automation waves were narrow. Agricultural mechanization automated specific physical tasks in farming. Industrial automation replaced specific manual operations in manufacturing. Computerization handled specific clerical functions. Each wave left large areas of work untouched and created time for workers and institutions to adjust.
AI is different in its breadth. The same underlying system can draft legal documents, generate code, analyse financial statements, synthesise research, and produce marketing copy. That breadth changes the adjustment problem. When the technology is domain-specific, workers can shift into adjacent domains while the new technology is absorbed. When the technology enters adjacent domains simultaneously, the lateral move is harder to find.
This does not make the pessimists right. It makes the adjustment harder and the policy requirement more urgent. The historical optimism is not wrong — aggregate outcomes from previous transitions have been positive. But that optimism is not a plan. It is a description of what happened when the conditions were right. Building those conditions is a deliberate act, not a default consequence of the technology.
The honest reading
The history of automation tells us the following, if we read it without the agenda of either reassurance or alarm.
Automation transitions are good for economic output in the long run. They are disruptive for specific workers, industries, and communities in the near term. The disruption is not evenly distributed — it concentrates in the places and roles where the technology is substituting directly, while the gains disperse widely and arrive later.
The reinstatement of displaced workers into new employment is not automatic. It depends on the expansion of demand in complementary areas, on the willingness of workers and institutions to adapt, and on policy choices that either accelerate or impede the transition.
AI is different from previous waves in its breadth, which means the historical record — drawn from narrow waves — is only partially informative about the specific adjustment challenge ahead.
I place senior leaders and build automation infrastructure for hiring systems. The disruption I watch in real time — the roles being eliminated, the skills becoming more valuable, the organisations that are adapting quickly and those that are not — follows the historical pattern in its broad shape. The aggregate outcome may well be positive. The specific people in the path of the transition are not helped by that knowledge. What helps them is a clear account of what is actually happening, not a reassurance drawn from statistics that smooth away exactly the thing they are experiencing.
The history is useful. It is not a substitute for honesty about what the transition requires.
Sources
Forbes — Agricultural employment historical data: 84% to 2% of US workforce
NBER — Computerization of White Collar Jobs (Dillender, 2022)
BLS — Growth Trends for Selected Occupations Considered at Risk from Automation
Did this land? Push back? Add something I missed?
Reply to Manas →Continue Reading
Related writing
Future of WorkWhat AI Agents Actually Are
The term has been diluted to the point of uselessness. Here's what agents actually are, why most production deployments fail, and what the infrastructure gap looks like from the inside.
Future of WorkWhy Human Judgment Still Matters in Hiring
Every few years, a new tool promises to take the human bias out of hiring and replace it with something more rigorous. The tool always underperforms on the thing that matters most: predicting whether a specific person will succeed in a specific role at a specific company. Here is why that prediction remains stubbornly human.
Future of WorkThe Global Hiring Floor
AI is not just changing who does the work. It is changing what the work is worth. As AI tools raise the output floor for every knowledge worker, the question of what human talent commands — and why — is being rewritten in real time.