Managers of Machines

I  Thesis

AI repeats the pattern

The case for the view rests on three arguments. Machines take tasks, not whole jobs. Each wave creates a new layer of supervision. And the value moves to whoever redesigns the work around the machine.

01Machines take tasks, and people move up a level

The human computers are the closest precedent because their work was cognitive, skilled and trusted, as AI's targets are today. When NACA became NASA in 1958, Dorothy Vaughan's West Area Computing unit was dissolved. Its computers joined the new Analysis and Computation Division, where Vaughan became an expert FORTRAN programmer. The calculation was automated, but the need for people who understood the problems and checked the results remained. Before his 1962 orbital flight, John Glenn asked Katherine Johnson to run the IBM computer's equations by hand.

The power loom shows the same shape. A handloom weaver worked one loom. In Lancashire in the 1830s and 1840s, a power-loom weaver typically tended two, and four or more after the Lancashire loom arrived in 1842. The skill moved from throwing the shuttle to keeping machines running. James Bessen, studying the mills of Lowell, Massachusetts, argues that wages rose as workers built the hard-to-learn know-how the machines demanded.

02Each wave creates a supervisory layer

Automation removes work and also creates new tasks. Daron Acemoglu and Pascual Restrepo describe this as displacement balanced by reinstatement: new tasks in which humans have the advantage. David Autor and colleagues estimate that about 60% of US employment in 2018 was in types of work that did not exist in 1940. Most of those new jobs sit around machines: operating, maintaining, programming, auditing and managing them.

Bank tellers show the same effect at a smaller scale. Between 1988 and 2004, ATMs cut the tellers needed per urban branch from about 20 to 13. That made branches cheaper to run, and the number of urban branches rose 43%. Teller jobs did not fall, and the role moved from counting cash towards selling and service.

03Value goes to those who redesign the work

Paul David's study of electrification is the strongest argument that managers matter more after a technology shift, not less. The first central power stations opened in 1881, yet in 1899 electric motors supplied under 5% of US factory mechanical drive. Productivity gains came only in the early 1920s, four decades later. By then just over half of drive capacity was electric, and plants were being redesigned from shared line shafts to a motor on each machine. The technology was necessary. The redesign is what paid off.

AI fits this reading. An AI model bolted onto an old workflow gives modest gains. Rebuilding a reporting process, a campaign pipeline or a customer-service desk around agents gives large ones, and that rebuild is managerial work. It means deciding which tasks the machine takes, where humans check, and how quality is measured. The manager of the future defines the work, sets the standard and signs off the output. That is a manager of machines.

In every earlier wave the machine took over execution and people moved up to specification, supervision and verification. AI is a stronger machine, but the shape of the change is the same, and so is the lesson: the winners are the managers who learn the machine first and redesign the work around it.

II  Antithesis

AI breaks the pattern

The case against the view also rests on three arguments. The historical pattern was harsher than the hopeful story admits. The people displaced were rarely the people promoted. And AI differs from every earlier machine because it can do part of the managing.

01The pattern was real, but it took generations

The promotion story is true for the economy in aggregate and over decades. It is often false for the individuals who lived through it. Britain had about 240,000 cotton handloom weavers in 1820 and about 43,000 by 1850. Their weekly earnings had already fallen from 240 pence in 1806 to 75 pence in 1830. Robert Allen calls the period from 1780 to 1840 Engels' pause: output per worker rose 46% while the real wage rose only 12%.

Electrification makes the same point from the other side. Its payoff took four decades, and David ties it to the slow switch from factories built around steam and shared line shafts to factories designed for electric drive. Until that redesign, the new technology sat inside the old way of working.

02The displaced were rarely the ones promoted

The handloom weavers were not promoted into the mills that replaced them. Their trade shrank to a fraction of its size. When AT&T mechanised telephone switching between 1920 and 1940, James Feigenbaum and Daniel Gross found that later cohorts of young women were not worse off overall. They moved into clerical work, such as secretarial jobs, and into service jobs. The incumbent operators fared worse: a decade later they were more likely to be in lower-paid occupations or no longer working. The new jobs went to new people.

Typing shows the same thing for a whole occupation. In the US, "word processors and typists" fell from 271,310 jobs in 1999 to 53,130 in 2018, a drop of 80%. The Bureau of Labor Statistics links it to the microcomputer revolution that began in the early 1980s. The occupation shrank rather than moving up a level. Vaughan is remembered because she is the exception: a capable manager who moved into the new division and mastered its tools.

03AI can manage, not just execute

This is the argument that breaks the analogy. Every earlier machine did a narrow task and needed a human to direct it. A loom could not decide what to weave, and an early computer could not decide what to compute. The supervisory layer survived because the machine could not supervise.

AI is the first technology that targets the coordinating layer itself: planning, assigning, reviewing, summarising and reporting. Corporate flattening gave a preview. Raghuram Rajan and Julie Wulf studied more than 300 large US firms from 1986 to the late 1990s. The number of layers between division heads and the CEO fell by 25%, and the CEO's direct reports rose from about 4 to about 7. They attribute this to technological and environmental change. AI extends that pressure to judgement-adjacent work. An agent can draft the plan, split it into tasks, send those tasks to other agents and check the results against a rubric. Even the new literacy repeats the problem: in 1960 programming was a skill only people had, and today AI writes code too.

Leontief's warning from 1983 applies here. He compared human labour to horses, which tractors "first reduced and then completely eliminated". Horses were not promoted to manage tractors. If AI matches humans on enough cognitive tasks, including coordination, the manager of machines may itself become a shrinking role.

04Speed removes the cushion

Earlier transitions were slow because they required physical capital: mills, power grids, tractors. That slowness let adjustment happen through generational turnover: fewer young people entered the old trade, and the old workers retired. AI spreads as software, in months rather than decades. The cushion that made earlier transitions survivable, time, is much thinner.

The promotion story describes the economy, not the people. Displaced workers and their managers were usually replaced by a new cohort rather than promoted. AI is faster than any earlier technology and reaches into the managerial layer itself, so the ladder may have fewer rungs than before.

III  Conclusion

The role survives. Most of the people who hold it do not.

The view holds for the role and fails for most of the people who hold it. Manager of machines is a real job that every wave has created, and AI will create it again. History also shows that the job goes to whoever takes it early. That is rarely the incumbent, and AI shrinks the window for taking it.

What survives from the thesis

The mechanism is sound. Machines take tasks, new supervisory tasks appear, and the large gains come from redesigning the work, as electrification showed. For AI, the redesign still needs people who can define the work, set the quality bar, check the output and carry accountability. An agent can check against a rubric. It cannot be held responsible for which rubric was chosen, or answer to a regulator, a client or a board. That accountable layer is the most durable form of managing machines.

What survives from the antithesis

Three corrections are needed. First, the promotion is not automatic. It went to people like Vaughan, who moved into the new division and mastered its tools. It did not go to the incumbent telephone operators, who ended up in lower-paid work. Second, the layer shrinks. Where a supervisor once managed ten people, a manager of agents may oversee the output of a hundred, so fewer managers are needed. Third, the time to adjust is shorter. What took decades with looms and electricity may take a few years with software.

Synthesis

The best historical equivalent is not a single episode. It is Vaughan inside electrification: an individual who retrained early, inside an economy whose gains depended on redesigning the work rather than adding a tool to it. For a manager today, three practical rules follow.

1. Move early. Vaughan moved into the new computing division in 1958 and became an expert FORTRAN programmer. The equivalent now is to build working fluency with AI agents while the old workflow still runs.
2. Redesign, don't bolt on. The value is in rebuilding processes around agents: deciding what the machine does, where humans check and how quality is measured. This is the part that belongs to management.
3. Own accountability, not execution. The durable role is to define the standard, verify the result and answer for it. The layers most exposed are those that only carry information: relaying, compiling and summarising.

So AI will turn some managers of people into managers of machines, as every earlier wave did. It will do so faster, with fewer seats, and only for those who choose the role before it is chosen for them.