When ATMs spread across American banking in the 1970s and 1980s, economist James Bessen documented something that surprised almost everyone predicting the teller's extinction: the number of tellers per branch fell by roughly a third, from about 21 to 13, but the total number of tellers barely moved. Banks used the labor savings to open more branches, cheaper to staff, and pushed tellers into a role ATMs couldn't touch: selling accounts, loans, and advice. Automation didn't eliminate the job. It changed what the job was.

Bank tellers are rated High risk, with a horizon toward 2027, the most urgent classification and timeline AI Doomsday has covered in this series. The question this article actually has to answer is whether the ATM paradox repeats, automation reshaping the role while headcount holds, or whether this time is genuinely different.

The early evidence says different. Bank branches in the United States have contracted by roughly 15% since 2017, a net decline with no corresponding wave of new branch openings to offset it the way there was after ATMs. The compensating expansion that saved the job in the 1990s has not shown up this time, and the reason is straightforward: AI is now capable of handling the relationship and advisory layer that ATMs never could.

Key Points

  • Bank tellers are rated High risk with a Low-term horizon toward 2027, the most urgent classification in this series, reflecting a decline the U.S. Bureau of Labor Statistics projects at 13% through 2035.
  • Economist James Bessen's well-known "ATM paradox" found that automation in the 1970s-90s cut tellers per branch by a third without reducing total employment, because banks opened more branches and pushed tellers into sales roles ATMs couldn't perform.
  • That offsetting expansion has not repeated this time: U.S. bank branches have contracted roughly 15% since 2017, with no compensating wave of new branch openings.
  • Bank of America's Erica assistant has handled more than 3 billion cumulative interactions since 2018 across 50 million active users, while Wells Fargo's Fargo assistant grew from 21.3 million interactions in 2023 to 245 million in 2024.
  • Even as teller headcount shrinks, the evolved "universal banker" role, blending routine service with sales and advisory work, carries a vacancy rate of roughly 19%, about four times the national average, a hiring gap AI is being used to bridge rather than a hiring surplus signaling the job's disappearance.

What a Bank Teller Actually Does

The core of the job, processing deposits, withdrawals, and routine transactions at a branch counter, has always been simple enough to automate in principle. What kept it a human job for decades was everything layered around that core: answering questions, resolving errors, building the kind of trust that gets a customer to open a second account or ask about a mortgage. That layer, not the transaction processing, is what the ATM never reached and what conversational AI is now reaching directly.

The ATM Paradox, and Why It Isn't Repeating

Bessen's research remains the reference point for understanding why automation predictions about this job have failed before. ATMs cut the number of tellers needed per branch by roughly a third, but 1990s deregulation and branch-expansion strategies meant banks opened new locations faster than automation shrank staffing at existing ones. Total teller employment stayed roughly flat for years after the technology that was supposed to eliminate the role became ubiquitous. Tellers who kept their jobs increasingly did so by becoming something closer to junior salespeople.

THE DIFFERENCE THIS TIME

The mechanism that saved the job before required two things: a task ATMs couldn't do, and a business reason to keep opening branches to create room for humans doing it. AI conversational assistants remove the first condition by handling much of the advisory and relationship work directly. Branch closures remove the second: down roughly 15% since 2017, with net closures continuing at hundreds of locations a year rather than the expansion that offset automation the last time around.

What AI Is Already Doing

Bank of America's Erica, launched in 2018, has processed more than 3 billion cumulative interactions and serves roughly 50 million active users, handling about 58 million interactions a month with a 98% rate of users finding what they came for. Internally, the bank reports Erica cut IT service desk call volume by half and reduced live chat volume in its corporate banking arm by 42%. Wells Fargo's comparable assistant, Fargo, grew from 21.3 million interactions in 2023 to 245 million in 2024, more than a tenfold increase in a single year, with roughly 80% of that usage in Spanish since the tool's expanded rollout.

These are not pilot programs or demos. They are production systems handling transaction volumes that dwarf what any branch network of human tellers could process, for exactly the category of routine inquiry, balance checks, transfers, bill payments, that used to require a person behind a counter or a phone line.

The Labor Market Reality

The U.S. Bureau of Labor Statistics projects a 13% decline in teller employment from 2025 to 2035, a loss of roughly 44,700 positions from a current base of about 339,200, driven explicitly by video kiosks, mobile check deposit, and online banking. That is one of the steepest projected declines of any occupation AI Doomsday tracks, though the agency still projects roughly 26,800 openings a year from retirements and turnover in the shrinking pool of remaining positions.

The twist is that the evolved version of this job, the "universal banker" role blending transaction processing with sales and advisory work, is simultaneously hard to staff. Industry reporting puts the vacancy rate for these roles at around 19%, roughly four times the national average, with banks using AI knowledge tools to help newer, less experienced staff cover the advisory gap left by branch positions that don't attract candidates. AI is estimated to free up roughly 12.7% of a banker's workday, previously spent on routine queries, for the relationship and sales work that increasingly defines what's left of the role.

How to Use AI as Bank Branch Staff Now

For routine service: assume AI chatbots will keep absorbing basic transaction and inquiry volume. The branch role that survives is not the one that competes with Erica or Fargo on speed; it's the one handling what those tools explicitly can't, complex disputes, fraud situations, and decisions that require discretion.

For career direction: the universal banker path, sales and advisory skills layered onto transaction competence, is where the vacancy data says the actual opportunity sits. Treat AI tools as the knowledge base that lets you operate credibly in that advisory role faster than traditional training would.

What I Think

The 2027 horizon and High-risk classification look justified to me, and the historical comparison to ATMs is actually what convinces me, not what reassures me. The ATM paradox held because automation left a clear task for humans to do and a business incentive to keep hiring them to do it. Conversational AI removes the first condition, and branch economics have removed the second. Both pillars that saved this job before are gone at the same time, which is a genuinely different situation than the one Bessen studied.

What I'd watch is whether the universal banker role becomes a real, durable destination for displaced tellers or a much smaller landing pad than the number of people leaving branch teller work. A 19% vacancy rate in the evolved role sounds like opportunity, but it's also consistent with a role that pays or trains too little relative to what it now demands, in which case AI isn't creating a soft landing so much as revealing how few good landings exist.

"ATMs took the transaction and left the relationship. AI is taking the relationship too. That's the difference between a job that survived automation once and a job that might not survive it twice."