Automation rarely kills jobs directly. The ATM didn’t eliminate bank tellers—it enabled more branches, which needed more tellers. What destroyed teller jobs was the iPhone, arriving decades later to combine with ATM-era automation and make branches themselves obsolete. This pattern—automation creating jobs until a second technology finishes them—is the most important framework for assessing what AI will actually eliminate.
What the ATM Paradox Actually Tells Us
When Chemical Bank installed the first ATM in Rockville Centre, New York in 1969, analysts predicted the end of the bank teller. The logic was intuitive: machines that dispense cash 24 hours a day would make human cash handlers redundant. What happened instead is one of economics’ most cited counter-intuitions.
ATMs did reduce the number of tellers required per branch—from roughly 21 down to about 13, according to research by economist James Bessen published in Finance & Development.1 But that cost reduction made opening a branch dramatically cheaper. Banks responded not by closing branches but by opening more of them. Urban bank branches grew by 43% during the period of peak ATM proliferation. Fewer staff per branch multiplied across more branches meant total teller employment grew, rising from approximately 300,000 in 1970 to over 600,000 by the early 2000s.2
The teller’s job description also evolved. With routine cash transactions automated, tellers shifted toward relationship banking—answering questions, cross-selling financial products, and handling complex service requests that no machine could resolve. Automation didn’t hollow out the role; it elevated it.
This is the mechanism economists call complementarity: the automated task (cash dispensing) lowered costs enough to expand the market, and humans remained essential for the tasks that remained. It’s a pattern that echoes across industrial history and carries direct implications for AI.
The Second Technology: Why the iPhone Finished What ATMs Started
The bank teller story doesn’t end in 2000. Between 2010 and 2018, Bank of America reduced its total headcount from 288,000 to 204,000—a reduction of 84,000 positions driven substantially by mobile banking adoption.3 US bank branches, which peaked at roughly 99,550 in 2009, have declined by nearly 30% since.4 Bureau of Labor Statistics projections forecast a further 15% decline in teller employment—approximately 53,000 positions—by 2032.5
What changed? The iPhone arrived in 2007 and mobile banking apps followed shortly after. Suddenly customers could deposit checks by photographing them, check balances, transfer funds, and receive alerts—all without visiting a branch. This didn’t just automate another set of teller tasks. It made the branch itself optional.
Bank of America CEO Brian Moynihan described the iPhone as a “game changer” that “effectively allowed customers to carry a bank branch in their pockets.”6 That wasn’t metaphor. Once the branch became a physical convenience rather than a necessity, the economics reversed: maintaining costly real estate and staff for dwindling foot traffic no longer made sense.
The ATM had automated cash. The smartphone automated the reason to go anywhere in the first place. The second technology didn’t just continue the first wave of automation—it changed the unit of analysis from the teller’s tasks to the branch’s existence.
A Pattern That Predates Banking
This two-wave structure appears throughout industrial history. Daron Acemoglu and Simon Johnson’s 2024 MIT research on machinery and labor during the early Industrial Revolution provides the clearest historical parallel.7
The introduction of spinning machines in the late 18th century disrupted cottage spinners but created a boom in demand for hand weavers, who now had cheaper, more abundant yarn to work with. Cotton’s share of British GDP grew from roughly 1% in the early 1780s to 7-8% by 1811-13. The spinners who lost work had somewhere to go.
Then came the power loom. A single power loom could outproduce 10 to 20 handweavers. Worse, these machines required factory buildings—eliminating cottage-industry weaving as a viable fallback entirely. Factory-scale production created relatively few new jobs to absorb 200,000+ displaced handloom weavers. Between 1814 and 1819, nominal weekly earnings for a weaving family of six fell by half.8
The pattern is structural: the first automation creates complementary demand. The second automation eliminates the complementary roles too—and often does so by changing the physical or economic infrastructure, not just the tasks.
Where AI Sits in This Cycle
By this framework, most current AI deployment looks like the ATM phase. Generative AI automates specific tasks within jobs—drafting emails, summarizing documents, generating initial code, analyzing data—while leaving the broader role intact or even expanded. Research from OpenAI, OpenResearch, and the University of Pennsylvania estimated that approximately 20% of US job tasks are exposed to current AI capabilities.9 MIT economist David Autor argues that well-deployed AI could function as a “force multiplier for expertise,” enabling more workers to perform higher-value tasks—precisely the teller-to-relationship-banker shift.10
Early productivity data supports the ATM-phase reading. A 24% decrease in AI-exposed skills per job posting in the top automation-exposure quartile is matched by a 15% increase in AI-augmented skills in roles most susceptible to AI collaboration—suggesting task displacement within roles, not wholesale job elimination.11
But the second-technology risk is real, and it’s already identifiable in several sectors.
What AI Will Actually Automate: The Second-Wave Candidates
The question isn’t whether AI will automate tasks within existing jobs—it already is. The question is which roles face a second-wave disruption where AI combines with another technology or systemic shift to make the role itself redundant.
| Sector | First Wave (Task Automation) | Second-Wave Trigger | Risk Level |
|---|---|---|---|
| Legal | Document review, contract analysis | AI + legal tech platforms making in-house counsel viable | High |
| Radiology | AI-assisted image reading | Teleradiology + AI enabling off-shoring/centralization | High |
| Software QA | Test generation, bug detection | AI-generated code + AI testing making dedicated QA unnecessary | High |
| Customer Service | Chatbots, FAQ handling | LLM agents + CRM integration replacing tier-1 and tier-2 support | High |
| Medical Coding | Automated ICD-10 assignment | EHR AI integration + regulatory approval removing manual review | Medium |
| Journalism (commodity) | AI drafts, summaries | AI + SEO automation + content farms eliminating low-differentiation writing roles | Medium |
| Financial Analysis | Data aggregation, report generation | AI + real-time data feeds reducing entry-level analyst pipelines | Medium |
| Teaching | Personalized tutoring assistance | AI + remote education infrastructure—but human mentorship creates floor | Low-Medium |
The Structural Conditions for Second-Wave Disruption
Three conditions converge when automation shifts from task-level to role-level destruction:
1. The complementary demand evaporates. ATMs made branches cheaper; mobile banking made branches unnecessary. When the demand driver for the human role disappears—not just individual tasks—the role becomes structurally vulnerable. For customer service, the demand driver is customers who need human interaction to resolve issues. As LLM agents improve at complex resolution, that driver shrinks.
2. Infrastructure reorganizes around the automation. Power looms required factories; factories created a new economic unit that displaced the cottage. Mobile banking required apps and smartphones; their adoption created a new economic unit (the digital-first bank) that displaced the branch. When AI enables a new organizational form—the legal team of 5 that does the work of 50, the radiology reading center that serves 200 hospitals—the infrastructure shift follows the capability.
3. Transition costs drop below replacement costs. Banks kept branches open for years after mobile banking existed because switching costs—retraining customers, managing branch leases, navigating union contracts—were high. When those costs fall below the ongoing cost of the existing structure, the transition accelerates quickly. In radiology, the transition cost is FDA approval; in customer service, it’s contact center renegotiation cycles.
Acemoglu’s concern that firms are currently optimizing AI for job replacement rather than productivity augmentation maps onto this framework.12 When corporate incentives align with infrastructure reorganization, the second-wave accelerates.
What This Means for Practitioners
McKinsey’s estimates project that between 400 million and 800 million individuals globally may need to find new jobs by 2030 due to automation, with up to 30% of hours worked in the US economy potentially automatable.13 These are the top-line numbers. The more useful question is whether your sector is in the ATM phase or approaching the mobile-banking phase.
Signals you’re still in the ATM phase: Your role requires human judgment at the point of decision, customer relationships drive value, or physical presence remains economically necessary. AI is making you faster but isn’t changing what the organization needs from you.
Signals you’re approaching the mobile-banking phase: Your role’s physical presence or institutional structure is the main reason it exists, the AI tools in your workflow are becoming capable enough that the human review step feels increasingly redundant, or new organizational models (smaller teams, centralized functions, remote-first) are becoming viable in your sector.
The workers most at risk aren’t those whose individual tasks AI can perform—they’re those whose entire organizational role becomes economically unnecessary when AI combines with a second enabling factor. The tellers didn’t lose jobs when ATMs appeared. They lost jobs when smartphones made branches a cost center without sufficient revenue justification.
The ATM lesson is ultimately optimistic for the near term and cautionary for the medium term. Automation creates jobs until the second technology arrives—and the second technology for many sectors is either already here or actively being built.
Frequently Asked Questions
Q: Did ATMs actually increase bank teller employment? A: Yes. ATMs reduced the cost of running a bank branch, which enabled banks to open more branches, increasing total teller headcount from approximately 300,000 in 1970 to over 600,000 by the early 2000s despite widespread ATM adoption.
Q: What finally caused bank teller employment to decline? A: Mobile banking, enabled by smartphones after 2009, made branch visits optional for most transactions. This caused US bank branches to decline nearly 30% from their 2009 peak of ~99,550, with teller employment projected to fall a further 15% by 2032.
Q: Is current AI in the “ATM phase” or the “iPhone phase” for most jobs? A: For most knowledge work roles, current AI resembles the ATM phase—automating specific tasks within jobs while often expanding total demand for the role. The iPhone phase (where the role itself becomes structurally unnecessary) is approaching in specific sectors including tier-1/2 customer service, commodity legal work, and certain diagnostic imaging specialties.
Q: Which jobs face the highest risk from second-wave AI disruption? A: Roles where AI capability is already high and a second enabling factor—platform infrastructure, regulatory approval, or organizational restructuring incentives—is actively converging. Legal discovery, customer service center work, commodity content production, and entry-level financial analysis face the clearest near-term second-wave conditions.
Q: How long did it take between ATM deployment and the iPhone eliminating branches? A: Roughly four decades. ATMs proliferated from 1969 through the 1990s; the iPhone arrived in 2007; significant branch closure acceleration didn’t begin until the mid-2010s. Historical second waves can take years to decades to materialize after the enabling technology appears—which means current AI exposure doesn’t necessarily translate to immediate displacement, but the structural conditions are being set now.
Footnotes
-
Bessen, James. “Toil and Technology.” IMF Finance & Development, March 2015. https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm ↩
-
Perry, Mark J. “What the Story of ATMs and Bank Tellers Reveals About the ‘Rise of the Robots’ and Jobs.” American Enterprise Institute, 2016. https://www.aei.org/economics/what-atms-bank-tellers-rise-robots-and-jobs/ ↩
-
“Why ATMs Didn’t Kill Bank Teller Jobs, But the iPhone Did.” David Oks, Substack. https://davidoks.blog/p/why-the-atm-didnt-kill-bank-teller ↩
-
FDIC Bank Structure Data. Branch Office Closings. https://banks.data.fdic.gov/bankfind-suite/oscr/branch_office_closings ↩
-
“Bank Tellers Are Going Away. What’s Next?” Troy Group. https://www.troygroup.com/blog/bank-tellers-are-going-away-whats-next ↩
-
Ibid. ↩
-
Acemoglu, Daron, and Simon Johnson. “Learning from Ricardo and Thompson: Machinery and Labor in the Early Industrial Revolution—and in the Age of AI.” MIT Economics, April 2024. https://economics.mit.edu/sites/default/files/2024-04/Learning%20from%20Ricardo%20and%20Thompson%20-%20Machinery%20and%20Labor%20in%20the%20Early%20Industrial%20Revolution%20-%20and%20in%20the%20Age%20of%20AI.pdf ↩
-
Ibid. ↩
-
Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” OpenAI, OpenResearch, University of Pennsylvania, 2023. ↩
-
Autor, David. “Applying AI to Rebuild Middle Class Jobs.” SSRN, 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4722981 ↩
-
“Displacement or Complementarity? The Labor Market Effects of Generative AI.” Harvard Business School Working Paper 25-039, 2025. https://www.hbs.edu/ris/Publication%20Files/25-039_05fbec84-1f23-459b-8410-e3cd7ab6c88a.pdf ↩
-
Acemoglu, Daron. “The Simple Macroeconomics of AI.” MIT Economics, April 2024. https://economics.mit.edu/sites/default/files/2024-04/The%20Simple%20Macroeconomics%20of%20AI.pdf ↩
-
McKinsey Global Institute. “Generative AI and the Future of Work in America.” 2023. https://www.mckinsey.com/mgi/our-research/generative-ai-and-the-future-of-work-in-america ↩