Chris Mason

The ATM/bank teller story is the most-cited historical analogy in the automation debate, and for good reason. When automatic teller machines were rolled out across the United States in the 1970s through 1990s, the prevailing assumption was that bank tellers would disappear. Machines that could dispense cash, accept deposits, and display balances seemed to make the human teller obsolete. That’s not what happened. As economist James Bessen documented in his influential IMF analysis (2015), ATMs reduced the number of tellers needed per branch from 21 to about 13. But this cost reduction made it economically viable to open more branches. Bank branches in urban areas increased 43% during the period of ATM proliferation. Fewer tellers per branch, but more branches, meant that total teller employment grew. Classic Jevons Paradox. What this chart shows and why the AI story is critically different: The left panel depicts the classic ATM/teller Jevons Paradox (as technology increases the efficiency with which a resource is used, the total consumption of that resource increases rather than decreases): automation made each branch cheaper to operate, banks expanded their branch networks, and total teller employment grew. This is the productivity effect in action. But notice what happened after 2010: mobile banking - a different, more complete form of automation — finally caused the teller decline that ATMs never produced. When the technology graduated from automating some of the teller’s tasks to automating nearly all of them, the Jevons effect broke down. The right panel shows AI’s impact on programming. The critical difference is speed. The ATM transition played out over approximately 40 years. The AI coding transition took roughly 3 years. When the pace of automation is slow, the economy has time for the Jevons Paradox to operate — businesses reorganize, new markets emerge, displaced workers retrain. When automation arrives in 3 years instead of 40, there is no time for these adjustment mechanisms. MIT Sloan research provides the academic framework for this observation: when AI automates only some tasks within a role, employment in that role often increases — because the remaining human tasks become more valuable. This is the ATM phase. But when AI can perform most tasks within a role, employment declines by approximately 14% at the firm level. This is the mobile banking phase. The central question for every AI-exposed occupation is: are we in the ATM phase (partial automation, Jevons effect, potential growth) or the mobile banking phase (near-complete automation, structural decline)? For programmers, the answer appears to be the latter. For many other knowledge-work categories, we are at the tipping point.