In the 1970s, America employed about 300,000 bank tellers. Then came the ATM. By 2010, over 400,000 machines dotted the country. Yet, teller numbers climbed to 600,000. This paradox—a replacement technology leading to more jobs—is now a favorite anecdote for those predicting a smooth AI transition. But the full story is less comforting, and for data and ML engineers building these systems, it’s essential reading.

As David Oks detailed in a recent analysis, ATMs cut staff per branch, lowering operating costs. Banks responded by opening 40% more branches between 1988 and 2004, hiring more tellers overall. But the job itself transformed. Tellers became salespeople for credit cards and mortgages; the title remained, the work did not.

Crucially, the reprieve wasn’t permanent. Tellers peaked in 2007. By 2022, numbers fell to 470,000, with a 15% decline projected this decade. The ATM didn’t kill the job overnight—it took 30 years, aided later by digital banking and AI. Furthermore, branch expansion was fueled by banking deregulation, a policy choice that interacted with the technology. Crediting the job growth to the machine alone is a mistake.

For engineers, the parallels and divergences with AI are stark. ATMs automated one physical task (dispensing cash), leaving a clear human role. Modern AI models, from LLMs to agentic systems, have blurry boundaries, capable of drafting, coding, analyzing, and servicing across numerous domains within a single job. Deployment is also measured in quarters, not decades, compressing any adjustment period.

Recent data echoes a different pattern. Major banks now deploy AI in research and operations to hold headcount flat while boosting output, a logic of efficiency, not expansion. Occupational forecasts for 2026 show declines in roles like paralegals and data clerks—areas where AI capabilities directly overlap.

The lesson isn’t that AI causes mass unemployment. New jobs will emerge. But the ATM tale shows automation’s effects are mediated by institutional responses—policy, market strategy, retraining—that are not guaranteed. The transition is a policy problem, not merely a technological one. For those designing the tools, understanding this history is the first step toward shaping a future where technology serves more than just the bottom line.

Source: Webpronews