The assumption that a machine that can do your job will take your job skips a financial step. A January 2024 working paper from MIT's FutureTech project found that, for computer vision tasks, only about 23% of the wages exposed to automation would actually be cheaper to automate today, once full system costs are counted.
Researchers led by Neil Thompson modeled the actual choice a business faces: what AI system would be needed to perform a task, what it would cost to build and maintain, and whether that beats paying a human. The answer, for most work, is no. "In many cases, humans are the more cost-effective way, and a more attractive way, to do work right now," Thompson told CNN.
A sliver of a sliver
Across the entire US economy, computer vision could technically automate tasks worth about 1.6% of worker wages, excluding farming. But once installation and maintenance costs are included, only around 0.4% of wages would actually be cheaper to automate. The gap between "could be done" and "worth doing" is the whole story.
The same pattern appears at the job level. About 36% of US non-farm jobs have at least one task a camera could handle, but only about 8% have a task where automation would pay off. The researchers considered the full cost of building, training, and running these systems year after year-and in most workplaces, that fixed cost never spreads thin enough to beat a wage.
A hypothetical small bakery illustrates the point. Checking ingredient quality is a minor part of a baker's day. The time and wages saved by installing cameras and AI would fall far short of what the upgrade costs. Small businesses rarely have the volume to justify a system that must be right almost every time.
What would change the math
The authors are not claiming job losses won't happen. "Overall, our findings suggest that AI job displacement will be substantial, but also gradual, and therefore there is room for policy and retraining to mitigate unemployment impacts," they write. Cost is a major brake on how fast that plays out.
Two forces could accelerate it. The first is falling costs: even a 20% annual price decline would take decades for many vision tasks to become cheap enough to automate, Thompson said. The second is shared services that spread the heavy build cost across many customers at once.
The impact will not land evenly. The study projects more automation in retail and healthcare, where the cost math favors machines, and less in construction, mining, and real estate.
Two caveats keep this honest
The study only covers computer vision-not the large language models behind chatbots-so it says nothing about text-based work. It also has not been peer-reviewed yet. Treat it as a serious correction to a lazy assumption, not a settled verdict about any specific occupation.
The useful shift in thinking, for researchers and everyone else, is to stop asking whether a machine could do a task and start asking whether, with all costs counted, anyone would actually pay it to. That is also the question that matters for career planning and for businesses considering new automation investments. Training focused on task-level economics, rather than hype about capability, better reflects the actual adoption timeline. For research professionals, the takeaway is: compute the full lifecycle cost of a task before deciding that full automation is inevitable.
The paper connects directly to broader questions about which work gets automated and why, and if you are assessing automation's impact on your own field, it is worth tracking AI Agents & Automation developments closely.
Your membership also unlocks:
