The ATM was supposed to end the bank teller. It didn't. Between 1970 and 2010, the number of bank tellers in the US actually grew, from roughly 300,000 to over 600,000. The machine that was supposed to replace them freed them up to do something more valuable: sell financial products, handle complaints, build relationships with customers. The ATM reduced the cost of running a branch, so banks opened more branches, and hired more tellers.
Then the iPhone happened. Mobile banking did what the ATM never could. Why drive to a branch at all? Branch visits collapsed. Banks started closing locations at scale. Teller headcount peaked around 2010 and has been falling since. By 2024, the Bureau of Labor Statistics projected a 15% decline in teller jobs through the decade.
Same industry. Two technologies. Opposite outcomes. The difference wasn't automation. It was whether the technology changed the reason to show up.
The Wrong Lesson Everyone Keeps Learning
When people talk about AI eliminating jobs, they're usually making the ATM argument while actually describing the iPhone scenario. These are not the same thing.
The ATM argument goes: AI will handle the repetitive tasks, humans will move up the value chain, everyone wins. It's clean, optimistic, and has historical support. The iPhone argument is harder: some technologies don't just change how work gets done, they eliminate the premise of a job category. When there's no branch to visit, the teller position doesn't transform. It just ends.
The honest answer is that AI is doing both things simultaneously, in different industries, at different speeds. The trick is figuring out which one you're in.
For knowledge work, the evidence so far looks more ATM than iPhone. GitHub's own data from Copilot deployments showed that developers using AI assistance completed tasks 55% faster. Those developers didn't lose their jobs. Their employers got more output. Some of those employers hired more developers. The parallel to branch economics is almost too neat.
But the jobs that were already marginal, where the human's role was to be a reliable interface between two systems, those are getting hollowed out faster than anyone expected.
What Actually Transforms vs. What Just Disappears
Here's a useful frame. The teller's job survived the ATM because the valuable part of a teller's job turned out to not be cash handling. It was trust, judgment, and the ability to have a conversation about a second mortgage. The ATM couldn't do that. So banks found out, somewhat by accident, what tellers were actually for.
The smartphone exposed that what tellers were actually for was showing up in a place. Once the place became optional, the job became optional.
AI is running the same experiment right now, across every category of knowledge work. It's revealing what the job is actually for. In some cases, like data entry, report formatting, first-pass research, the answer is: not much, the human was a processing layer. Those roles are contracting.
In other cases, the experiment is producing the opposite result. Tasks that looked simple are turning out to require more human involvement after AI gets involved, not less. A customer service agent using an AI assistant handles more tickets, but the tickets they personally touch are harder, stranger, and higher stakes than before. The volume goes up. The complexity goes up. The job doesn't disappear; it gets compressed into its most demanding form.
Where Human Pages Fits Into This
We built Human Pages on a specific hypothesis: AI agents will generate demand for human work in categories that don't exist yet, the same way ATM economics generated demand for branch salespeople.
Here's a concrete example of what that looks like in practice. An AI agent running a research workflow for a venture firm hits a wall. It needs someone to call three portfolio company CEOs and ask a question that requires actual judgment to phrase correctly in context. It can't do that call. It can't even write the script without knowing which of the 40 ways to ask this question won't cause a relationship problem. So the agent posts a job on Human Pages. A human picks it up, makes the three calls, returns structured notes. The agent continues its workflow.
The human in that scenario isn't competing with the AI. The AI created the job. It needed a teller.
This happens dozens of times a day on the platform, in different forms. An agent needs someone to verify a physical address exists. An agent needs a human to watch a 40-minute video and identify the three most emotionally resonant moments. An agent needs someone to translate a nuanced message into Mandarin with specific cultural context the model keeps getting wrong. These aren't jobs that existed five years ago. They're not jobs any human would have thought to post on a traditional job board. They're gaps that only became visible once an AI started trying to close them and couldn't.
The Category Nobody Named
Bank tellers didn't call themselves ATM-complementary workers. They just kept showing up and discovered, over a decade, that their job had shifted. The transition was messier than the retrospective analysis makes it look.
We're in that messy part now. There's no clean taxonomy for what humans will do in an AI-driven economy, because the jobs are emerging in real time, shaped by what AI agents are actually attempting and where they're actually failing.
What the ATM story does tell us is that the failure mode isn't automation. The failure mode is irrelevance. Tellers survived because they turned out to be useful for something the machine couldn't do. The ones who didn't survive were in branches that closed because the premise of the branch itself became unnecessary.
The question worth sitting with isn't whether AI will take your job. It's whether the thing your job is actually for, the part that requires you specifically, survives the experiment. Some jobs will find out they were always about something deeper than the task. Others will find out the task was the whole point.
History suggests both outcomes are real. The hard part is knowing which one you're in before the answer becomes obvious.
Top comments (0)