When ATMs proliferated across the United States from the 1970s through the 1990s, economists expected bank teller employment to collapse. Instead, it nearly doubled. The number of tellers grew from roughly 300,000 in 1970 to over 600,000 by the early 2000s—even as ATM counts surged past 400,000. Then the iPhone arrived, and the real destruction began.
This is the framework you need for understanding AI and employment. Not the naive version—“AI automates tasks, workers become redundant”—but the historically grounded one: automation often creates jobs until a second, complementary technology eliminates the reason those workers existed in the first place.
The ATM Paradox: What the Numbers Actually Show
The ATM story is counterintuitive enough that economists still use it as a cautionary tale against linear automation predictions. Between 1988 and 2004, as banks deployed ATMs at scale, the number of tellers required per branch fell—from roughly 20 to 13, a reduction of about one-third.1 Standard economics would predict that total teller employment therefore declined.
It didn’t. Urban bank branches increased by 43% during peak ATM proliferation.1 Lower staffing costs per branch meant lower costs to operate a branch, which made previously unprofitable locations viable. Banks expanded their physical footprint. More branches needed more tellers, even if each branch needed fewer.
Total teller employment grew substantially through this period. The workers who remained shifted from cash handling—now delegated to machines—toward relationship banking, product cross-selling, and loan origination. The job changed. Employment didn’t decline.
The iPhone Moment: When the Second Technology Strikes
Bank branches peaked at roughly 99,550 in 2009.2 Then mobile banking arrived—not the primitive WAP-era experiments of the early 2000s, but the iPhone-enabled, app-driven, full-service digital banking that emerged from 2008 onward.
The effect was decisive. By 2023, the number of US bank branches had fallen to under 78,000—a decline of more than 21,000 locations, or roughly 22%, from the peak.2 When people stopped visiting branches, branches stopped needing to exist. When branches closed, tellers became redundant. The Bureau of Labor Statistics now projects bank teller employment—which stood at approximately 347,400 in 2024—will decline by 13% through 2034.3
This is the second-technology effect. The ATM automated what tellers did. The iPhone automated the reason people visited tellers. The first created a new equilibrium. The second collapsed it.
The key distinction: task automation expands adjacent roles; purpose automation eliminates them.
History Rhymes—The Pattern Across Technologies
Bank tellers are not an isolated case. The two-phase pattern appears consistently across automation waves:
Word processors and typists. When word processors arrived in the late 1970s, typing pool employment held relatively stable. Organizations produced more documents, and specialists remained valuable because not everyone operated the machines. The second technology was the personal computer combined with accessible software: once executives typed their own emails, the specialist intermediary disappeared. Secretarial and administrative support employment fell from roughly 4 million to under 2 million between 1980 and 2010.4
Spreadsheets and bookkeepers. VisiCalc and Lotus 1-2-3 appeared between 1979 and 1983. Accounting employment continued growing through the 1980s and 1990s—more analysis became feasible, demand expanded. The second technology is still unfolding: cloud accounting software (QuickBooks, Xero), combined with direct bank feeds and automated transaction categorization, is now eliminating the bookkeeping tier of the profession. The higher-skill accounting functions remain; the routine reconciliation work is disappearing.
Manufacturing robots and factory workers. Industrial automation reduced assembly-line roles from the 1970s onward, but employment in manufacturing held up through the 1980s as productivity gains expanded the total market. The second technology—combined offshoring enabled by container shipping and trade liberalization—eliminated the comparative advantage that sustained US manufacturing employment at scale. According to David Autor’s landmark research on job polarization, blue-collar production, craft, and operative jobs fell by roughly 16% as this combination took hold.5
The State of AI: Mapping the Phases
Most AI deployments today sit firmly in the first phase—the ATM phase. The pattern by sector:
| Sector | Current Phase | What AI Automates Now | What Would Trigger Phase 2 |
|---|---|---|---|
| Software development | ATM | Code generation, documentation, testing | Autonomous agents that handle business requirements → deployment end-to-end |
| Customer service | ATM → transitioning | Tier-1 queries, FAQ resolution | Full-context conversational AI eliminating escalation paths entirely |
| Legal (document work) | ATM | Contract review, due diligence, e-discovery | Regulatory systems that interface directly with AI; AI-certified legal filings |
| Accounting (bookkeeping) | Phase 2 arriving | Transaction categorization, reconciliation | Already underway via cloud + AI combination |
| Radiology (routine reads) | ATM | Flagging anomalies, prioritizing worklists | AI-as-primary-reader accepted by regulators and insurers |
| Content/copywriting | ATM → transitioning | Drafting, SEO optimization, templated content | Distribution systems that algorithmically personalize AI content at zero marginal cost |
| Data analysis | ATM | Query generation, visualization, pattern detection | Autonomous insight pipelines that replace analyst judgment loops |
The software development row deserves attention. A 2023 controlled study found that developers using GitHub Copilot completed programming tasks 55.8% faster—the speed gains are real and significant.6 A parallel study found that AI assistance in customer service improved issue resolution rates by 14.4% on average, with newer workers seeing gains of up to 30%.7 These are ATM-phase dynamics: productivity gains per worker, expanding what teams can accomplish.
But the second-technology trajectory is visible. The question is not whether AI agents can write code—they demonstrably can. The question is when they can reliably translate ambiguous business requirements into tested, deployed, maintained software systems without human review as a critical path. That’s the iPhone moment for software development employment.
The Acemoglu Problem: Not All Automation Leads to Phase 1
There is a crucial caveat to the optimistic ATM-phase reading: not every technology follows the pattern.
Daron Acemoglu and Pascual Restrepo’s 2019 paper “The Wrong Kind of AI?” argues that the current wave of AI investment is skewed toward pure labor substitution rather than creating new tasks where humans have comparative advantages.8 Automation always reduces the labor share of value-added income, even when it raises aggregate productivity. The net employment effect—the ATM-phase expansion—only materializes when automation creates enough new complementary tasks to offset displacement.
Acemoglu’s 2024 macroeconomic analysis found that even under optimistic assumptions, AI’s effect on total factor productivity is likely below 0.66% over a decade.9 That’s non-trivial but far from transformative—and it suggests that the efficiency gains are not necessarily flowing through to workers or new job creation at scale.
This matters for the two-phase framework: if the economic incentives systematically favor labor replacement over task augmentation, the ATM phase may be shorter and weaker than historical precedents suggest. The second technology may not need to wait for the iPhone; it may arrive as a configuration decision in the first technology’s architecture.
What AI Will Actually Automate—And the Timeline
The honest assessment, grounded in the historical pattern:
High probability, near-term (2025–2028): Roles defined primarily by information retrieval and synthesis under structured conditions—insurance claims processing, lower-tier legal document review, financial data analysis, basic customer service. These are in late ATM phase or early Phase 2 transition. The World Economic Forum’s 2025 Future of Jobs report estimates AI will create roughly 11 million new jobs globally while displacing approximately 9 million—a net gain at the aggregate level, but concentrated disruption within specific occupational categories.10
High probability, medium-term (2028–2033): Roles where the second technology is the AI itself operating as an autonomous agent. This includes significant portions of software development support, medical imaging interpretation, contract generation, and investigative journalism. The trigger is not AI capability improving—it is organizational trust and regulatory frameworks catching up to capability. Once a hospital is willing to rely on AI imaging reads without radiologist sign-off, the employment mathematics change structurally.
Uncertain, contingent on breakthroughs: Physical-world roles requiring dexterous manipulation, roles requiring genuine relationship trust (certain therapy, complex negotiation, senior advisory), and roles where accountability structures demand human decision-makers by law or strong social norm.
Carl Frey and Michael Osborne’s foundational 2013 Oxford study estimated 47% of US occupations at high risk of computerization.11 The OECD’s task-based follow-up analysis reached just 9%, because many nominally automatable occupations contain non-routine components that vary too much for current systems.12 Both are correct—the gap is a matter of timing and the second-technology question. The 47% is a ceiling assuming Phase 2 arrives. The 9% is closer to near-term observable reality.
What to Do with This Framework
For workers: the relevant question is not “can AI do any part of my job?” but “is there a second technology that will eliminate the reason my role exists?” If you work in a role that exists because humans must physically visit a location, sign something in person, or interact through a constrained channel—examine what technologies are converging on that channel constraint. That is where Phase 2 pressure builds.
For organizations: the ATM phase creates a false sense of security. Banks that didn’t adapt their branch strategy before mobile banking hit were caught flat-footed. The window between Phase 1 and Phase 2 is real but finite, and it is not always possible to see Phase 2 approaching until the transition is already underway.
David Autor’s research suggests that technological change consistently creates more middle-class opportunity when it creates new tasks—work that didn’t exist before—rather than purely automating existing ones.5 AI assistants to radiologists, AI-augmented legal counsel, AI-enabled research positions: these expand the economic pie. AI replacing radiologists, AI replacing junior lawyers, AI replacing analysts: these redistribute the pie. The policy and investment choices made in the next five years will determine which dynamic dominates.
The ATM lesson is not that automation is harmless. It is that automation’s harm often arrives from a direction you weren’t watching.
Frequently Asked Questions
Q: Didn’t ATMs eventually cause bank teller job losses? A: Yes—but through a second technology, not ATMs directly. Mobile banking reduced branch visits, which triggered branch closures, which eliminated teller positions. ATMs alone produced a net increase in teller employment for decades by lowering branch operating costs and enabling geographic expansion.
Q: Is AI more like the ATM or the iPhone for knowledge workers? A: Currently more like the ATM for most sectors—AI is lowering the cost of cognitive work, likely expanding demand for skilled workers in the near term. The iPhone moment arrives when AI agents can handle complete workflows autonomously, eliminating the organizational reason for human intermediaries. That transition is closer in some fields (document review) than others (software architecture).
Q: How should I evaluate my own job’s automation risk? A: Apply the two-technology test. First: does AI automate specific tasks I perform? (Near-term risk: low if your role involves judgment, relationships, or variable physical tasks.) Second: is there an emerging second technology that could eliminate the reason my role exists—the channel constraint, the information asymmetry, the oversight requirement? That second question is the one that determines long-run structural risk.
Q: What does Acemoglu’s “Wrong Kind of AI” argument mean practically? A: It means that the economic incentives driving AI investment currently favor replacement over augmentation. Unlike previous automation waves that created new human task categories alongside destroying old ones, current AI investment is disproportionately focused on pure labor substitution. Without deliberate policy intervention or business model shifts that favor augmentation, the “ATM phase” employment benefits may be smaller and shorter than historical precedents suggest.
Q: When should workers take AI-driven displacement seriously rather than dismissing it as hype? A: When two conditions align: (1) AI demonstrates reliable performance on the core judgment tasks of a role, not just peripheral tasks; and (2) organizational or regulatory frameworks begin approving AI outputs without mandatory human review. The first condition is approaching or already met in several knowledge-work categories. The second condition is the actual displacement trigger—and it tends to arrive suddenly once it starts.
Footnotes
-
Bessen, James. “Toil and Technology.” IMF Finance & Development, March 2015. Based on research in Learning by Doing: The Real Connection Between Innovation, Wages, and Wealth. Yale University Press, 2015. ↩ ↩2
-
FDIC. Quarterly Banking Profile, 2024. Branch count data from FDIC Summary of Deposits; peak of ~99,550 in 2009, declining to under 78,000 by 2023. ↩ ↩2
-
Bureau of Labor Statistics. Occupational Outlook Handbook: Tellers, 2024–2034 projection cycle. https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm ↩
-
Bureau of Labor Statistics. Historical occupational employment data, administrative and secretarial occupational categories, 1980–2010. ↩
-
Autor, David, and David Dorn. “The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market.” American Economic Review, 103(5), 2013. See also Autor, David. “Work of the Past, Work of the Future.” NBER Working Paper 25588, 2019. ↩ ↩2
-
Peng, Sida, et al. “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.” arXiv
.06590, February 2023. ↩ -
Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. “Generative AI at Work.” arXiv
.11771, 2023. Published in Quarterly Journal of Economics, 2025. ↩ -
Acemoglu, Daron, and Pascual Restrepo. “The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand.” Cambridge Journal of Regions, Economy and Society, 13(1), 2020. NBER Working Paper 25682, 2019. ↩
-
Acemoglu, Daron. “The Simple Macroeconomics of AI.” NBER Working Paper 32487, 2024. ↩
-
World Economic Forum. Future of Jobs Report 2025. January 2025. Sample: 1,000+ employers, 14M+ workers across 55 economies. ↩
-
Frey, Carl Benedikt, and Michael A. Osborne. “The Future of Employment: How Susceptible Are Jobs to Computerisation?” Oxford Martin School Working Paper, September 2013. Published in Technological Forecasting and Social Change, 2017. ↩
-
Arntz, Melanie, Terry Gregory, and Ulrich Zierahn. “The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis.” OECD Social, Employment and Migration Working Papers, No. 189, 2016. ↩