ATMs Didn't Kill the Bank Teller. The iPhone Did.

June 10, 2024

In 1970, there were about 300,000 bank tellers in the United States. Then came the ATM. Everyone knew what this meant. A machine that could count cash faster and more accurately than a human, available at 2am, never calling in sick. The teller was finished.

Except the opposite happened. By 2010, employment was up to nearly 600,000. Not double, but a real, sustained increase, and it held for roughly 30 years while ATM installations climbed toward 400,000 nationwide.¹

Here's the mechanism, and it's worth sitting with because it's not intuitive. ATMs made it cheaper to run a bank branch. Cheaper branches meant banks opened more of them. Tellers per branch fell from around 20 to about 13 between 1988 and 2004, but urban branch count over that same stretch went up 43%. Fewer tellers per branch, way more branches. Do the math and total teller employment climbed for a generation.²

The tellers who survived didn't survive by out-counting the machine. They survived because the job quietly turned into something else. Economist James Bessen of Boston University, who did the research that made this whole story famous, found that banks started treating tellers as part of the "customer relationship team." Economist David Autor summarized the new bar in one sentence: a teller who can tally currency but can't provide relationship banking isn't going to fare well at a modern bank.³ Cross-selling a mortgage. Untangling a small business owner's cash flow problem. Being the human a customer trusts with something a machine can't be trusted with. The ATM took the arithmetic. The bank kept the relationship, and it turned out the relationship was worth more.

So far this is the story everyone tells. Automation ate the boring part, humans moved up, everybody wins, cue the inspirational LinkedIn post.

Here's the part almost nobody tells.

The ATM never killed the bank teller. Something else did, decades later, and it wasn't even trying to.

It was the iPhone.

Once a customer could deposit a check and check their balance from a couch, the entire interaction left the building. Not the counting. Not one task inside the job. The whole reason to walk into a branch. Teller employment, after 30-some years of growth, is now in a genuine, sustained decline. The Bureau of Labor Statistics currently projects a 13% drop over the 2024–2034 decade, on top of declines already logged in prior projection cycles.⁴

The ATM was a better tool for doing the job. The iPhone made the job's reason to exist optional.

Those are not the same threat. And a person who successfully out-adapted the first one had no particular reason to see the second one coming, because it didn't come from the direction they'd been watching.

So which one is AI, right now, for the desk job everyone's asking about?

Let's use software engineering, since it's the one closest to home for a lot of us.

Current AI tools, Copilot, Cursor, Claude Code, whatever you're using this month, are extremely good at the ATM job. They take a clear task and execute it fast. Write this function. Draft these tests. Refactor this file. What they're not doing, at least not reliably yet, is deciding what should be built, catching the mistake three layers up in the architecture, or owning it when the thing breaks in production at 3am. That's still a human problem, and if anything, it's a more valuable human problem now that the typing part is cheap.

There's even an early version of the same growth effect we saw in banking. Cheaper to build software, so more software gets attempted, so more total demand shows up for people who can direct and verify it well. Junior engineers today are increasingly expected to review AI output the way a lead reviews a junior's pull request, which is a genuinely higher-order skill than writing the code themselves would have been five years ago. The floor moved up. That's the uplevel story working as advertised.

There's a real crack in that story worth naming honestly, though. Anthropic's own economic research found that hiring of young workers, ages 22 to 25, into the occupations most exposed to AI has slowed by around 14% since ChatGPT launched, even while overall unemployment in those fields hasn't moved much yet.⁵ That's not proof the uplevel story is failing. But it's a real early signal that the entry ramp into these jobs, the stuff junior people used to cut their teeth on, is narrowing before we've figured out what replaces it. Worth keeping in view, because it's exactly the kind of quiet leading indicator nobody was watching in banking either.

So call it: today's AI is the ATM. It automated the task. So far, the role has adapted and arguably gotten more interesting because of it.

Which raises the actual question. What's the iPhone?

Here's the trap almost everyone falls into when they try to answer this, including me a draft ago. They picture a better coding tool. A more autonomous agent, a smarter Copilot, something that writes more of the function for you. But that's just imagining a faster ATM. The whole point of the iPhone story is that the real threat didn't come from inside banking at all. Nobody at Chase or Wells Fargo built it. It came sideways, from a phone company and a hardware team who weren't trying to disrupt tellers and probably never once thought about them. The thing that ends a role rarely looks like the thing that's currently automating it, just faster.

So here are three candidates that aren't "AI writes better code," because I think that one's actually the least interesting answer available.

Candidate one: the unit that employs engineers disappears, not the coding task. Today you need an "engineering org" mostly because shipping software requires coordinating a dozen specialized humans inside one company: engineers, PMs, designers, QA, DevOps. If one non-technical person with an AI agent stack can spin up, run, and wind down a working software business alone, competitively, the disruption isn't that coding got automated. It's that the container that used to employ coders, the company with a dedicated eng team, stops being the default unit of production. That's structurally identical to the iPhone: the branch didn't get more efficient, the reason to visit one disappeared.

Candidate two: the interface itself stops needing to be built. Almost everything a software engineer makes today exists because a human has to look at a screen and click things. If interaction shifts to something ambient or conversational, you describe what you want to an agent and it composes the capability on the spot from raw data and APIs, with nothing persistent left behind, then "product with a UI someone engineered" stops being the object of the work. Not because AI writes UIs faster. Because the category of thing engineers spend their careers building quietly stops being how people use computers. That change wouldn't come from an AI coding lab. It'd come from wherever the next interaction paradigm actually gets invented, which might not be a software company at all.

Candidate three: a liability mechanism nobody's watching removes the human, not a capability jump. The reason a person still has to review AI-written code before it ships isn't purely technical, it's that somebody has to be legally and professionally accountable if it breaks something expensive. If an insurance product, a legal standard, or a regulatory framework emerges that lets companies run AI-generated code without a human in the loop, because the risk is now priced and transferred rather than personally owned, the human reviewer's job doesn't get automated by a smarter model. It gets removed by a policy decision made in an insurance office, not an AI lab. Boring, plausible, and exactly the kind of thing nobody currently arguing about AI and jobs is paying attention to.

Here's the honest tension, and I think it's the actual question worth sitting with.

Banking had something close to fixed demand. People need a finite amount of banking done, and no amount of cheap ATMs was going to make anyone need a second checking account. So when the iPhone came for the channel instead of the task, there was nowhere for demand to expand into. It just shrank.

Software has never worked that way. Every time it got dramatically cheaper to build, in the PC era, the internet era, the mobile era, demand didn't shrink into a smaller pool of engineers doing higher-level work. It exploded into entirely new categories of software nobody had thought to build yet, because building had just become affordable enough to try. Cheaper dev tools didn't just uplevel developers, they created the app economy, which needed more developers than existed before.

So the real question isn't "which of these three candidates is the iPhone." It's a level above that:

Is software more like banking, where demand is basically fixed and getting cheaper eventually just means needing fewer people? Or is it the one domain that's never once behaved that way, where every "iPhone moment" so far has made the pie bigger instead of the slice smaller?

I don't think there's a clean answer to that yet, and I'd trust anyone who tells you they're certain a little less, not a little more. But it's the question I'd actually be asking, instead of the one everyone's asking, which is just "will AI take my job." The more useful version is: does my job have a floor above it to retreat to, the way the teller did the first time. And separately, is the thing threatening it a better tool, or a reason to stop needing the tool at all.

Those require completely different answers. Betting on the wrong one is how you win the fight you can see and lose to the one you can't.

Notes

¹ Teller employment growth alongside ATM proliferation: James Bessen (Boston University), documented in Learning by Doing: The Real Connection between Innovation, Wages, and Wealth (2015), and summarized in "Toil and Technology," Finance & Development, IMF, March 2015: https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

² The 43% urban branch increase and the ~20-to-13 tellers-per-branch shift are from Bessen's original data, as discussed in the AEI writeup "What the Story of ATMs and Bank Tellers Reveals About the 'Rise of the Robots' and Jobs": https://www.aei.org/economics/what-atms-bank-tellers-rise-robots-and-jobs/

³ David Autor's summary of the new teller job requirement, and Bessen's "customer relationship team" framing, are discussed in this recap: https://markcrosling.medium.com/bank-tellers-and-the-atm-the-role-that-changed-instead-of-vanishing-d1830b380d71

⁴ U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, Tellers, 2024–34 projections: https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm

⁵ Massenkoff, M. & McCrory, P., "Labor Market Impacts of AI: A New Measure and Early Evidence," Anthropic, March 2026: https://www.anthropic.com/research/labor-market-impacts. The report also names computer programmers among the occupations with the highest AI exposure.

Notes

Further reading