Why did ATMs make bank tellers more common, not fewer — and when does automation flip from helping workers to replacing them?
Why do ATMs and automated systems often complement rather than replace human workers?
This explores why automating part of a job, like cash dispensing at banks, often keeps or even grows human roles instead of eliminating them, and when that pattern breaks down.
This explores why automating part of a job, like cash dispensing at banks, often keeps or even grows human roles instead of eliminating them, and when that pattern breaks down. The best-known case is the bank teller. ATMs spread widely between 1995 and 2010, yet US teller employment went up. Autor explains this by looking at the whole chain of work: when a machine makes some steps cheaper or more reliable, the human steps that remain become more valuable, so firms keep paying for them Does automating some tasks make other human tasks more valuable?. Cheaper branches meant more branches, and tellers shifted toward relationship and sales work.
Gans and Goldfarb sharpen the point with the 'O-ring' idea. In many jobs the tasks multiply together rather than add up, so one weak link drags down the whole result. When machines take over some links, workers spend their time on the bottleneck that's left, and the quality and value of that work rise Does automating some tasks actually increase the value of remaining work?. A side effect is that the usual 'AI exposure' scores probably overstate job loss, because they assume the remaining tasks keep their old value. A maritime shipbroking firm shows the same thing in practice. Its manual classification work wasn't eliminated. It was turned into checking the algorithm and correcting it, and that feedback loop became a strategic asset for the company Does automation eliminate manual work or repurpose it?.
Complementing workers is not automatic, though. It's a design and business choice. Acemoglu, Autor and Johnson argue that firms often earn more by automating expertise than by building tools that create new tasks for workers, so worker-friendly AI gets underfunded Why do firms build automating AI instead of pro-worker AI?. Brynjolfsson warns that leaning toward replacement can lock in a situation where workers lose bargaining power and can't win it back Does automating labor concentrate power more than augmenting it?. There's already evidence of uneven replacement: firms with heavy AI exposure swap out freelance workers for AI faster and more cheaply than other firms do Do firms substitute labor for AI at different rates?. Shneiderman's framework pushes back on the idea that this is a forced trade-off. Automation and human control are separate dials, and both can be set high at once Can automation and human control increase together?.
The part you may not expect is that the human role that's left can be harder, not easier. Bainbridge's classic 'ironies of automation' argument says machines take the routine work and leave people to monitor and step in during failures, while the skills they'd need to do that fade from lack of practice Does automation actually remove humans from the control loop?. A study mapping thousands of workplace AI risk scenarios found that working alongside AI doesn't protect against this skill erosion on its own Does AI augmentation protect workers from skill erosion?. The Taobao customer-service experiment shows the limits clearly. Humans who took over from a chatbot could fix technical problems, but they couldn't undo customer frustration if they stepped in too late When human workers take over from AI, does it actually fix the problem?. When automated systems fail, the blame often lands on the nearest human operator, even when that person had little real control. Elish calls this the 'moral crumple zone' Do automated systems shift blame onto their human operators?.
One last twist concerns which workers are safest. In human-AI teams, the person most at risk of replacement isn't the least accurate one. It's the one whose mistakes most closely match the AI's mistakes Does error diversity matter more than accuracy for human replacement?. People stay valuable next to automation because they're different from the machine, not because they're better at the same thing. The ATM didn't replace tellers because tellers ended up doing work the machine couldn't.
Sources 12 notes
Autor argues that when automation makes some production steps more reliable or cheaper, it increases the value of remaining human links—evidenced by US bank teller employment rising despite ATM proliferation from 1995 to 2010.
When job tasks multiply together rather than add (O-ring production), automating some tasks concentrates worker time on what remains, raising its quality and value. This means standard exposure indices underestimate worker resilience by assuming static task value.
A maritime shipbroking firm did not eliminate manual classification when introducing algorithmic support. Instead, it repurposed this work to generate ground truth for auditing, creating a feedback loop between auditing and algorithm alteration that became a strategic capability.
Acemoglu, Autor and Johnson argue that automating expertise generates higher economic returns for firms than creating new tasks, creating a collective-action gap where individual profit-maximization conflicts with worker welfare.
Brynjolfsson argues that AI designed to automate rather than augment human labor shifts bargaining power to those controlling the technology, risking a stable equilibrium where workers have no leverage to improve outcomes.
Show all 12 sources
Higher AI-exposed firms replace online labor marketplace workers with AI tools faster and at lower cost than less-exposed firms, suggesting returns to scale in internal AI capability rather than uniform technology diffusion.
Shneiderman's HCAI framework decouples automation from control into two independent dimensions, allowing both to be high simultaneously. Evidence comes from task-specific examples like patient-controlled analgesia, where automation and oversight coexist without conflict.
Bainbridge argues that automating industrial processes doesn't remove the operator—it leaves them monitoring and handling failures they're now less equipped to manage. Manual control timing and process knowledge decay through disuse, making takeover exactly when it's most needed least likely to succeed.
Research mapping 8,356 workplace AI risk scenarios found that augmentation mode does not inherently prevent harm. Overreliance on AI agents can gradually erode worker skills and their capacity to provide meaningful oversight, undermining augmentation's core safety justification.
A randomized field experiment on Alibaba's Taobao found that when workers took over after AI failures, service quality recovered for capability mismatches but not for customer frustration. Timing mattered: early intervention succeeded; late escalations after frustration accumulated failed to recover quality.
Elish demonstrates that when automated systems malfunction, investigators and media assign responsibility to nearby human operators regardless of their actual authority or information access, protecting the technology's reputation while the operator absorbs moral and legal liability.
Choudhary and colleagues show that the human most at risk of replacement in an AI ensemble is not the least accurate but the one whose prediction errors most closely align with the AI's. Ensemble performance depends on error dispersion, not individual skill.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Simple Macroeconomics of AI
- Why Are There Still So Many Jobs? The History and Future of Workplace Automation
- O-Ring Automation
- Automation, AI, and the Intergenerational Transmission of Knowledge
- Agents, robots, and us: Skill partnerships in the age of AI
- Human Performance Consequences of Stages and Levels of Automation: An Integrated Meta-Analysis
- Augmenting the algorithm: Emerging human-in-the-loop work configurations
- A Model for Types and Levels of Human Interaction with Automation