I spend a lot of time at conferences and in conversations about AI and the future of work. Whenever the topic of job displacement comes up, someone raises the ATM story: when ATMs arrived, everyone predicted bank tellers would disappear, yet teller employment grew. Therefore, the argument goes, we should relax about AI.¹

I had always found this story a little too convenient, so I pulled the data myself. The results do not support the conclusion that is usually drawn from them.

What the Teller Employment Numbers Actually Show

Let’s start with the raw figures. Microdata from IPUMS (Integrated Public Use Microdata Series), a harmonized archive of U.S. Census and survey records hosted at the University of Minnesota, put U.S. bank teller employment at roughly 193,000 in 1960, 365,000 in 1970, and 520,000 in 1980.² The first American ATM went live at a Chemical Bank branch in Rockville Centre, New York, on September 2, 1969.³ By 1980, about 18,500 ATMs were running. By 2000, about 273,000.⁴

On the surface, the story holds: ATMs multiplied, and so did tellers.

Figure 1. Bank tellers, ATMs, and branches all grew between 1960 and 2000. The aggregate picture hides the mechanism.

But this aggregate picture hides a more useful decomposition, and it is one that Bessen himself identifies. Teller employment grew because bank branches grew. Branch counts nearly quadrupled, from 10,556 in 1960 to 38,738 in 1980.⁵ More branches meant more tellers, mechanically. This is exactly the relationship Bessen highlights in his analysis. Where I part company with the standard telling is on the next question: what drove that branch expansion? Bessen attributes it in large part to ATM-driven cost reductions. The data, as we will see, point elsewhere.

The Branch Boom and the Question of Causation

The forces behind the 1960-1980 branch expansion were structural, regulatory, and demographic. Each is worth examining, because together they leave little room for the ATM as a causal factor.

Figure 2. Branches were already growing rapidly before the first ATM was installed in 1969, driven by suburbanization and deregulation.

The largest force was postwar suburbanization. Millions of households moved to communities with no bank presence at all. Banks followed them, opening branches in shopping centers, strip malls, and new residential developments along the expanding suburban fringe. Banks followed their customers.

Then came branching deregulation, a rolling transformation of state banking law that unfolded across three decades. Before the 1960s, twelve states prohibited branching outright (so-called “unit banking” states), and 34 more imposed tight geographic limits. These restrictions fell piecemeal: New Jersey opened statewide branching in 1973, New York in 1975, Florida shifted from unit banking to countywide in 1977, then statewide in 1980. By 1992, all but four states allowed some form of statewide branching.⁶ Each time a state relaxed its rules, it opened a new geography for competitive branch-building.

The Bank Holding Company Act of 1956 reinforced this dynamic from the federal side. By blocking interstate expansion, the law forced banks to compete within their home states. If you could not grow outward, you grew denser. Rising household banking participation, the spread of payroll direct deposit, the popularity of drive-through windows, and cheaper standardized construction all made it possible to run a profitable branch on a thinner deposit base than ever before.

Here is the correlation-versus-causation problem at the heart of the ATM parable. The branch expansion trend started a full decade before the first ATM was installed and continued through a period when ATM adoption was still in the low thousands. It is difficult to attribute a trend to a technology that was not yet present when the trend began and was still negligible during the trend’s fastest growth phase. The branch boom is better explained by the structural forces just described. That does not mean ATMs had zero influence on branch economics. It means the case for ATMs as the primary driver of branch expansion has a serious timing problem.

What Bessen Actually Argues, and Where It Gets Complicated

The standard account of the ATM paradox comes primarily from James Bessen’s Learning by Doing: The Real Connection between Innovation, Wages, and Wealth (Yale University Press, 2015), summarized in his article “Toil and Technology” for the IMF’s (International Monetary Fund) Finance & Development in March 2015.⁷ His argument deserves a fair hearing in its strongest form before I complicate it.

Bessen shows that ATMs cut the number of tellers needed per branch from roughly 20 to 13 in a typical urban branch between 1988 and 2004. That is a 35% reduction in labor intensity per operating unit. But, he argues, this reduction in staffing requirements lowered the cost of operating a branch, which made it economical for banks to open more branches, which required more total tellers even with fewer per site. Economists call this the “complementarity” channel: cost-reducing automation expands the scale of the activity enough to offset the per-unit labor savings. Bessen also makes a second, related argument about task recomposition: as cash-handling migrated to machines, the teller role shifted toward relationship banking, customer service, and cross-selling of financial products.

His per-branch observation is solid. The decline from 20 to 13 tellers per urban branch is well-documented and consistent with both the FDIC (Federal Deposit Insurance Corporation) branch data and the CPS (Current Population Survey) employment series. His task recomposition argument holds up, as we will see when we look at wages. Where the story gets complicated is in the causal mechanism connecting ATMs to branch expansion.

That mechanism requires that ATMs made branches cheaper to operate in total, creating the economic incentive for banks to open more of them. This is where David Humphrey’s contemporaneous research inside the Federal Reserve system introduces an important complication.

Humphrey, a Federal Reserve Bank of Richmond economist who studied bank cost structures across the entire ATM rollout period, found something surprising. His 1994 paper in the Richmond Fed’s Economic Quarterly, “Delivering Deposit Services: ATMs Versus Branches,” showed that while a single ATM transaction cost about half as much as the equivalent teller transaction at a branch, customers used ATMs roughly twice as often as they had used tellers.⁸ The per-transaction savings were real. The aggregate savings were not. Banks did not spend less on deposit delivery after installing ATMs; they delivered more transactions at the same total cost.

A careful reader will notice that Bessen and Humphrey are measuring different things. Bessen points to the reduction in labor costs per branch (fewer tellers on payroll). Humphrey measures the total cost of deposit delivery (which includes ATM hardware, software, network fees, and maintenance alongside teller labor). Labor costs per branch may have fallen while total deposit delivery costs held steady, because the capital expense of ATMs replaced the labor savings.⁹ Consider what that means for Bessen’s causal chain. If the total cost of operating a branch did not fall, the economic incentive to open new branches on the basis of ATM-driven savings is weaker than the standard telling implies. A bank deciding whether to open a new branch cares about total operating cost, not just the teller line on the budget.

The timing problem is perhaps the most straightforward test of the causation question. Bank branches were growing at 6-7% per year compounded well before ATMs existed. The branch count rose from 10,556 in 1960 to 21,839 in 1970, a period during which there were zero ATMs in the country. After ATMs arrived at scale, the growth rate did not accelerate. It slowed. Over the decade from 1970 to 1980, when ATM penetration was still minimal, branches grew 77%. Over the two decades from 1980 to 2000, the period of mass ATM deployment, branches grew 70% total, which works out to a lower annualized rate over a longer span.¹⁰ If ATMs were making branches cheaper and thereby driving expansion, we would expect the growth rate to accelerate after ATMs arrived. The opposite happened.

This is a classic correlation-versus-causation puzzle. ATMs and branches both increased during the same period. The standard telling treats this correlation as causal. The timeline suggests the causation runs the other way, or does not run at all: branches expanded because of suburbanization and deregulation, and ATMs were deployed into those branches.

What Was Actually Driving Bank Costs

If ATMs were not the primary driver of branch expansion, what was reshaping the cost structure of banking during this period? Humphrey’s earlier and more comprehensive 1993 study in the Journal of Productivity Analysis, “Cost and Technical Change: Effects from Bank Deregulation,” answers this question with rigor, drawing on a panel of 683 large U.S. banks from 1977 to 1988.¹¹

The regulatory context is essential. Before 1980, Regulation Q prohibited banks from paying market interest rates on deposits. Unable to compete on price, banks competed on convenience: more branches, free deposit services, longer hours. This produced an oversupply of branches relative to what would have been economically optimal at market interest rates.¹² When the Depository Institutions Deregulation and Monetary Control Act of 1980 removed interest rate ceilings, banks had to pay market rates. Deposit interest costs per dollar of assets rose 75% between 1977-78 and 1987-88. Banks found themselves overbranched for a world in which the branch was no longer the primary competitive instrument.

The result was an industry under severe cost pressure. Humphrey’s data show employees per branch falling from 44.8 in 1980 to 35.2 in 1988, a 21% decline. Deposits per branch rose 27% over 1982-88 as banks squeezed more output through each location. A Booz-Allen & Hamilton study for the American Bankers Association concluded that roughly half of all U.S. bank branches were unprofitable at this time, kept open for market share and Community Reinvestment Act (CRA) compliance rather than standalone economics.¹³

The finding that deserves more attention than it has received: Humphrey measured net technical change in banking as negative over the entire 1977-88 period, averaging negative 0.8% to negative 1.4% per year. Despite an estimated $20 billion in annual information technology spending by 1989, rising interest costs outweighed any operating efficiency gains.¹⁴ Banks were spending more on technology and becoming measurably less productive throughout the entire ATM rollout period.

This does not mean ATMs were useless. They reduced per-branch labor requirements, as Bessen documents. But ATMs were one cost-cutting tool among several in an industry restructuring under pressure from interest rate deregulation. The story was not “ATMs made branches cheaper, so banks opened more.” It was closer to “deregulation created intense cost pressure, banks deployed ATMs as part of a broader efficiency response, and branches kept expanding for reasons (suburbanization, state-level deregulation of branching restrictions) that predated and were independent of the ATM.”

ATMs were an instrument of a restructuring that was already underway. They did not set it in motion.

The Per-Branch Numbers Tell the Real Story

Figure 3. Divide total tellers by total branches and the labor-saving effect of ATMs appears. The ratio peaked in the early 1970s and fell steadily after.

Consider what the per-branch decline means at the level of an individual worker. A branch that once employed 20 tellers now employs 13. A new branch opens across town. The displaced worker may or may not be hired there. Aggregate statistics wash this experience out, but it was real.

What the optimists leave out is the mechanism. The complementarity between ATMs and tellers depended on an expanding frontier of branch-opening opportunities. Demographic shifts and branching deregulation created that frontier. ATMs did not. Confusing the correlation (ATMs and branches both grew) with causation (ATMs caused branches to grow) is the central error in the standard telling of this story.

The Counterfactual That Nobody Mentions

There is another way to read the employment data, one the standard telling quietly avoids.

From 1960 to 1981, teller employment grew by 409,000, from 193,000 to a peak of 602,541.¹⁵ This was a period of rapid banking expansion, rising household participation, and the opening of thousands of new branches. Then the growth stopped. Employment after 1981 never again reached 602,000. It moved within a band near 500,000 for two decades before falling to 430,000 by 2000.

The question the standard parable never asks is: what would teller employment have looked like without ATMs? Banking output grew between 1981 and 2001. The branch network expanded from 38,738 to 66,000. If the pre-1981 relationship between banking growth and teller employment had continued at even a modest rate, you would expect more than 515,000 tellers by 2001. The leveling-off coincides with the period of rapid ATM deployment.

This is the destruction of potential jobs rather than existing ones. It is economically equivalent from the worker’s perspective, but it is politically invisible.¹⁶ No one organized a protest over teller positions that were never created. The industry avoided mass layoffs; it stopped hiring at the rate that banking growth would otherwise have demanded. This is displacement by absorption, and it is the hardest form of technological unemployment to see in aggregate data.

What Happened to Wages

The employment numbers tell only half the story. If we want to know what automation did to bank tellers as workers, we have to look at what they were paid.

Figure 7. Real wages fell through the 1980s, stagnated into the mid-1990s, spiked in 1998, then returned to the starting level. Two decades of automation left the median teller’s purchasing power essentially unchanged.

I pulled weighted median wages from the same IPUMS CPS ASEC (Annual Social and Economic Supplement) microdata I used for the employment series and deflated them to 1980 dollars using CPI-U (Consumer Price Index for All Urban Consumers) annual averages.¹⁷ The pattern is striking and underappreciated.

In 1980, the median bank teller earned $7,200 in real terms. By 1989, with ATM installations having quadrupled from 18,500 to 76,000, that figure had fallen to $6,645. That is an 8% decline in real purchasing power across a decade in which nominal wages rose from $7,200 to $10,000. Inflation ate the raise and then some.

The most straightforward reading of this period is deskilling. As cash-handling migrated to machines, the residual teller task bundle shifted downward in the labor market. Banks could hire and retain tellers at lower real cost because the job required less specialized skill and carried less scarcity value. The productivity gains from ATMs flowed to bank shareholders through lower branch operating costs. They did not flow to workers through higher pay.¹⁸

From 1990 through 1997, real wages moved within a narrow band between roughly $6,300 and $7,100 while ATM counts more than doubled from 80,000 to 165,000. Headcount held roughly stable. The paradox looked intact at the level of employment. Workers, however, saw no wage benefit from the ongoing automation. The teller role was in a transitional state: automation had stripped out the cash-handling component, and the shift toward relationship banking and product cross-selling had not yet shown up as a measurable wage premium at the median.

The inflection came in 1998, when median nominal wages jumped from $12,300 to $15,000, a 22% increase in a single year, and real wages hit $7,583, the highest point in the entire 21-year series. This inflection lines up with the period when ATMs had crossed 187,000 units and routine cash transactions were unambiguously machine-dominated. The timing is consistent with Bessen’s upskilling hypothesis, and this is where his analysis is most persuasive: by the late 1990s, banks had reorganized the teller role around customer relationship management, cross-selling of financial products, and problem resolution. These tasks were harder to automate and commanded a premium. The remaining teller workforce was, on average, more skilled than its 1980 predecessor.¹⁹

The full arc, however, is less encouraging. It took roughly eighteen years of real wage erosion and stagnation before the “upskilling premium” appeared. And when it did appear, it brought real wages back to approximately where they had started in 1980. The 2001 real median of $6,979 is within 3% of the 1980 figure of $7,200. Two decades of automation, and the median teller’s purchasing power went nowhere.²⁰

A caveat on the data: CPS ASEC income variables are coarse (prior-year, topcoded, self-reported), and the year-to-year volatility in a single-month survey is real. The 1986 figure ($7,624 real) looks anomalously high relative to its neighbors and may reflect sampling noise. Directional trends across multi-year periods are more reliable than any individual year-over-year movement. But the broad shape of the series is clear, and it complicates the ATM parable in ways that matter.

The Plateau and the Decline

Figure 4. After 1980, teller employment stopped growing. It oscillated around 500,000 for two decades, then fell, while the branch network kept expanding.

The CPS ASEC annual data from 1980 forward show what happened to headcount once the structural tailwind weakened. Teller employment moved between roughly 450,000 and 590,000 through the 1980s and 1990s, with substantial year-to-year noise (a known feature of the CPS single-month survey design). It never sustainably exceeded the 1980 level of 520,075. By 2000, it had dropped to 430,115.

Meanwhile, the branch network kept growing, reaching about 66,000 by 2000. The divergence between these two curves is instructive. Branch counts rose. Teller headcounts went sideways and then fell. The labor-saving arithmetic of ATMs did what labor-saving arithmetic always does. For two decades, the addition of new branches generated enough new positions to absorb the positions lost per branch. When branch growth decelerated, the underlying displacement showed through.

Teller employment peaked around 2010 and has fallen steadily since, as online and mobile banking reduced branch foot traffic below the threshold at which new openings could compensate for per-branch staffing reductions. The occupation is now in secular decline by any measure.

Why This Analogy Cannot Bear the Weight Placed on It

Grant the ATM optimists their strongest version of events: a labor-saving technology’s displacement effects were masked for a generation by a coincident structural expansion. Even on these terms, the analogy to AI does not hold.

The first problem is scope of automation. The ATM automated a single, tightly bounded task: dispensing cash and accepting deposits. The teller role had always been broader. Tellers opened accounts, processed loan payments, answered questions, and handled the sort of face-to-face relationship banking that customers valued. When ATMs absorbed the cash-dispensing function, tellers shifted toward advisory, sales, and service work. This task recomposition worked because the automation was narrow enough to leave most of the role intact. Large language models (LLMs) and their successors are general-purpose cognitive technologies. They operate across the full range of symbolic and analytical tasks that define knowledge work. They do not automate one function within a job; they compress entire occupational profiles.²¹

The second is the absence of a structural offset. The branch expansion that absorbed displaced tellers was historically specific: a one-time deregulatory opening compounded by a one-time demographic shift. There is no obvious equivalent for the knowledge economy. When AI lowers the cost of producing legal analysis, financial modeling, software, or marketing strategy, the question becomes whether demand for these outputs is elastic enough that cheaper production generates sufficient new volume to offset the labor savings. For some narrow categories this may hold. For professional and analytical services broadly, the empirical literature on demand elasticity gives little reason for optimism.²²

The third is speed and friction. Each ATM was a physical device: manufactured, shipped, installed, wired into a network, and maintained. Deploying 273,000 of them across the United States took roughly three decades. AI has no equivalent deployment friction. A large language model is available to anyone with a browser the day it launches. The ATM transition played out across four decades; a full generation of tellers experienced a slowly tightening occupational labor market before the decline became unambiguous. AI adoption is measured in quarters. The range of affected occupations is orders of magnitude wider. A displacement cycle that took 40 years in retail banking may compress into five or ten years across dozens of white-collar fields simultaneously. The adjustment costs of a slow transition and a fast one are qualitatively different, even if the eventual equilibrium is the same.

The wage data adds a fourth consideration that the standard telling omits. Even during the period when the complementarity channel appeared to be working, when headcount was stable and the parable looked intact, workers were absorbing a real cost. Eighteen years of real wage erosion is not a minor footnote. If that pattern repeats at AI speed and AI scale across the knowledge economy, it would represent a significant transfer of value from labor to capital.

Humphrey’s work introduces a fifth: the causal mechanism that the standard story depends on is, at best, weaker than commonly presented. If the total cost of deposit delivery did not fall despite ATMs processing transactions at half the per-unit cost of tellers, then we cannot straightforwardly attribute the branch expansion to ATM-driven cost savings. The branches expanded for other reasons. The teller headcount held up for other reasons. The ATM was one factor in a complex system, and the popular version of the story inflates its causal role to the point of distortion.²³

What the Data Say

Figure 5. As ATM density rose, the per-branch teller count fell. By 2000 there were roughly four ATMs for every branch.

Figure 6. Indexed to 1980 = 100. Branches kept climbing, tellers per branch kept falling. Total teller employment is the product of these opposing forces.

Between 1960 and 1980, U.S. bank branches grew at roughly 6.7% per year compounded, adding close to 28,000 locations. This structural expansion absorbed the per-branch labor reduction from about 18,500 ATMs. By 2000, with 273,000 ATMs running and 66,000 branches open, teller employment had fallen to 430,000 from the plateau of 500,000-plus that characterized the late 1980s. The median teller’s real wage in 2001 was almost exactly what it had been in 1980. And the employment peak of 602,000 in 1981 was never reached again, despite two decades of growth in banking output.

I want to be precise about what I am and am not claiming. I am not making a prediction about AI and employment. Prediction is hard, and the historical record offers no shortage of confident forecasts, in both directions, that turned out wrong.

What I am saying is narrower. The ATM-teller story, as commonly told, does not support the proposition that automation creates as many jobs as it destroys. It is a case study in which a narrow automation technology produced a steady per-unit labor reduction that an unrelated structural expansion temporarily masked, while workers bore the technology’s costs in the form of suppressed wages and a broken growth trajectory for nearly two decades. The causal mechanism most often cited to explain the paradox (ATMs lowered costs, enabling more branches) is, at best, weakly supported by the available cost data and contradicted by the timing of branch expansion.²⁴

None of this proves that AI will destroy jobs. It proves that a very popular piece of evidence offered to the contrary does not survive contact with the data it claims to describe. The people who want to have a serious, fact-based conversation about AI and labor markets deserve a better starting point than a parable that falls apart when you look at it closely.

We should retire this story and start from what we actually know.

---

Footnotes

1. The most influential version of this argument is probably the one from the American Enterprise Institute (AEI), drawing on James Bessen’s research: “What the Story of ATMs and Bank Tellers Reveals About the Rise of the Robots and Jobs,” AEI, 2016. It has since been repeated in countless LinkedIn posts, conference talks, and op-eds, almost always without reference to the underlying data. ↩

2. IPUMS USA decennial census microdata, 1% general-purpose samples, weighted by PERWT. OCC (occupation) code 301 for 1960 and 1970 (1960/70 census classification); OCC code 383 for 1980 onward (1980/90 classification). The 1980 figure here is from the CPS ASEC rather than the decennial census, as it provides annual granularity for the subsequent series. See IPUMS USA (usa.ipums.org). ↩

3. This is the conventionally cited date, though there is some ambiguity about whether earlier cash-dispensing devices in the UK and elsewhere count as “ATMs” in the modern sense. For our purposes it does not matter: the U.S. branch network was already growing rapidly for a decade before the first machine was installed. ↩

4. ATM counts are industry estimates compiled from ATM Marketplace, NCR Atleos historical commentary, and Bessen (2015). No official government series for ATM installations exists, which is a recurring frustration in the literature. The figures should be read as approximate. ↩

5. FDIC Table CB-1, “Number of Insured Commercial Banks, Branches and Total Offices at Year End, 1934-1996,” published in Statistics on Banking, FDIC Division of Research and Statistics. Archived at FRASER (Federal Reserve Archival System for Economic Research), Federal Reserve Bank of St. Louis. This is the authoritative source for branch counts through 1996. The 2000 figure of approximately 66,000 comes from the FDIC BankFind Suite and is consistent with Federal Reserve and Statista estimates. ↩

6. Federal Reserve Bank of Minneapolis, “Loosening the Strings of Regulation” (2004). The pace and sequencing of state-level branching deregulation is one of those topics that sounds dry until you realize it explains a massive share of the variation in branch counts. If you are interested in the institutional detail, this is the piece to read. ↩

7. James Bessen, Learning by Doing: The Real Connection between Innovation, Wages, and Wealth (Yale University Press, 2015). The ATM argument is summarized in his “Toil and Technology” article for the IMF’s Finance & Development, March 2015. Bessen’s research is serious work by a serious scholar. His observation about per-branch labor reduction is well-documented. His insight about task recomposition toward relationship banking is supported by the wage data. My quarrel is with the causal mechanism connecting ATMs to branch expansion, and with how the finding has been simplified in popular discourse to dismiss concerns about AI displacement. ↩

8. David B. Humphrey, “Delivering Deposit Services: ATMs Versus Branches,” Federal Reserve Bank of Richmond Economic Quarterly, vol. 80, Spring 1994, pp. 59-81. This paper is cited far less often than Bessen in the popular discourse, which is unfortunate, because it speaks to the cost question that Bessen’s mechanism depends on. ↩

9. This is a genuinely important distinction, and I want to be fair to Bessen here. It is possible to read Humphrey and Bessen as compatible at the per-branch level: labor costs fell (Bessen), but total deposit delivery costs held steady because ATM capital and operating costs replaced the labor savings, and because transaction volume expanded (Humphrey). The question is whether the labor savings alone were sufficient to incentivize new branch openings. Given that labor is one cost among many in running a branch (rent, compliance, technology, management), a 35% reduction in teller staffing, while real, may not translate to a branch-opening incentive of the magnitude the popular story implies. ↩

10. This is perhaps the simplest and most damaging observation for the Bessen mechanism. The annualized growth rate was higher in the decade before ATMs arrived at scale than in the two decades after. The correlation between ATMs and branches is real, but the direction of causation that the standard story requires is hard to reconcile with the timeline. ↩

11. David B. Humphrey, “Cost and Technical Change: Effects from Bank Deregulation,” Journal of Productivity Analysis, vol. 4, iss. 1-2, June 1993, pp. 9-34. Also available as Federal Reserve Bank of Richmond Working Paper 90-5 (1990). ↩

12. Regulation Q turned branches into the competitive instrument of banking. If you cannot offer a better interest rate, you offer a more convenient location. The entire branch infrastructure of American banking was built under this incentive structure. When the structure changed, the infrastructure did not shrink. It just became expensive to maintain. ↩

13. Booz-Allen & Hamilton, “Managing Delivery System Economics,” a study for the American Bankers Association, October 1987. Cited in Humphrey (1993). The original study is not publicly available, which is a shame, because “half of all bank branches are unprofitable” is the sort of finding that ought to feature more prominently in discussions of banking efficiency. CRA compliance obligations created a distinct incentive to keep branches open in lower-income areas regardless of profitability, which complicates any simple reading of branch counts as reflecting market demand. ↩

14. The $20 billion figure comes from Humphrey’s review of industry estimates. The negative technical change finding is robust across three different measurement approaches (time trend, time-specific index, and cross-sectional cost function shifts), which is part of what makes it striking. Banks spent heavily on technology and became measurably less efficient. The explanation is that interest rate deregulation imposed costs faster than technology could reduce them. This is the Solow productivity paradox applied to banking: you can see the computers everywhere except in the productivity statistics. ↩

15. The 1981 peak of 602,541 is from the CPS ASEC. The CPS ASEC is a single-month (March) survey, so individual year figures carry more noise than decennial census benchmarks. That said, the 1981 figure is not an obvious outlier in the way that, say, the excluded 2014 figure (~805,000) is. It sits at the top of a clear upward trend running from 193,000 in 1960 through 520,000 in 1980. ↩

16. This distinction between the destruction of existing jobs and the suppression of job creation is one that labor economics handles well in theory and poorly in public discourse. The BLS (Bureau of Labor Statistics) tracks employment levels, not counterfactual employment levels. There is no monthly report on “jobs that would have existed.” This makes suppression effects hard to detect without careful analysis, which in turn makes them easy to deny. The ATM story benefits from this asymmetry. ↩

17. Weighted median of INCWAGE (total annual pre-tax wage and salary income for the prior calendar year) for all persons coded OCC1990=383 with INCWAGE > 0 and INCWAGE < 99,999,998, weighted by ASECWT. Deflated to 1980 dollars using BLS CPI-U annual averages. Note that INCWAGE captures prior-year earnings, so the “1980” figure reflects 1979 wages. The series is best read as an annual index of relative purchasing power. See IPUMS CPS (cps.ipums.org). ↩

18. Real wage stagnation was not unique to bank tellers during this period. Real wages for many occupations in the bottom two-thirds of the income distribution were flat or falling through the 1980s. The teller wage decline is therefore not attributable solely to ATMs; it coincided with broader trends in wage compression, union decline, and labor market deregulation. That said, the teller-specific pattern (deskilling through task automation, followed by a delayed upskilling premium) is distinctive enough to be informative about the mechanism, even if the magnitude is partly explained by macro factors. ↩

19. Bessen deserves credit for this part of the analysis. His observation that the teller role was recomposed toward relationship banking and advisory work, with an attendant increase in skill requirements, is the strongest element of his argument and is well-supported by the wage data. The 1998 wage jump is consistent with the upskilling hypothesis. Where I diverge from the popular reading of Bessen is on what happened to the median worker during the transition. The endpoint was arguably positive. The path to get there was eighteen years of real wage decline. The popular version of the story skips the path and presents only the endpoint. ↩

20. A useful comparison: the S&P 500 (Standard & Poor’s 500 stock index) returned approximately 1,100% in nominal terms over the same 1980-2001 period. The productivity gains from the automation of banking did go somewhere. They went to capital. ↩

21. This is the distinction between what economists call “task-specific” and “general-purpose” technologies. An ATM replaces a specific task (dispensing cash) within a broader occupation. An LLM can perform or assist with legal research, financial analysis, content creation, code generation, customer service, and administrative coordination, among other things. The number of residual tasks left to the human worker after a general-purpose cognitive technology is deployed is, by definition, smaller than the number left after a task-specific one. ↩

22. The demand elasticity question is the central empirical unknown in AI labor economics right now. The optimistic case requires that when AI cuts the cost of, say, legal research by 80%, the total volume of legal research purchased rises by more than 5x to offset the labor savings. For some outputs this is plausible (more people might buy legal services if they cost less). For many professional services the evidence suggests demand is relatively inelastic at the margin: people purchase roughly the amount they need, and cheaper production does not generate proportionally more demand. ↩

23. I use “just-so story” language deliberately here. In evolutionary biology, a just-so story is a post hoc narrative that explains an observed outcome with a plausible but untested causal mechanism. The popular version of the Bessen narrative has this structure: we observe that branch counts and teller counts both went up during the ATM era, and we construct a causal chain (ATMs lowered costs, lower costs enabled branches, branches required tellers) that fits the correlation. Humphrey’s cost data and the branch expansion timeline both weaken the first link in that chain. The correlation remains. The causal mechanism, in its strong form, does not hold up. ↩

24. Humphrey (1993, 1994). I want to be clear about the limits of this critique. Humphrey’s work does not prove ATMs had zero marginal effect on branch costs. It shows that aggregate deposit delivery costs were a wash, because increased transaction volume consumed the per-transaction savings. It is possible that ATMs had some marginal cost benefit at the branch level that is not visible in the aggregate data, and that this benefit contributed in some small way to branch-opening decisions. The point is that the strong form of the mechanism, in which ATMs made branches substantially cheaper and thereby caused the branch expansion, is not supported by the best available cost data from the period, and is contradicted by the timeline of branch growth. ↩