The story most of us have absorbed goes like this: once a machine can technically do your job, the machine gets your job. It sounds like a law of nature. If the software can spot the defect, sort the parcel, or grade the produce, then the human doing that work is on borrowed time.
That assumption skips a step. Being able to do a task is not the same as it making financial sense. A machine can be perfectly capable and still cost far more than the person already doing the work, once you count everything it takes to build, install, and keep it running. When a team at MIT added up that full cost, the picture looked very different from the mass-job-loss version.
The paper we’re leaning on here looks at only one slice of AI, so it isn’t the last word on the whole technology. But it’s careful, and it reframes a debate that needs reframing.
What the MIT team measured
In January 2024, researchers at MIT published a working paper with a dry title: “Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?” The key word is “beyond.” Most earlier research asked a narrower question: could an AI system, in principle, perform this task? That tells you what’s possible. It says nothing about what a business would actually choose to do.
The team, led by Neil Thompson of MIT’s FutureTech project, modeled the real decision a company faces. As Thompson describes it, the model starts with the performance of real tasks, asks what AI system would be needed to do them, and then asks whether a business would actually go for it. They focused on computer vision, the kind of AI that reads and interprets images, partly because its costs are easier to pin down than other kinds.
Thompson put the headline finding plainly. “In many cases, humans are the more cost-effective way, and a more economically attractive way, to do work right now,” he told CNN. That “right now” matters, and we’ll come back to it.
The numbers that undercut the panic
“We find that only 23% of worker compensation ‘exposed’ to AI computer vision would be cost-effective for firms to automate because of the large upfront costs of AI systems,” the paper reports. Roughly three-quarters of the work computer vision could technically touch is cheaper to keep with people.
Zoom out to the whole US economy and the share gets smaller. Computer vision could technically automate tasks worth about 1.6% of US worker wages, not counting farming. But once you add up the cost, only around 0.4% of wages would actually be cheaper to automate today. A sliver of a sliver.
The gap shows up at the job level too. One write-up notes that while around 36% of US non-farm jobs have at least one task a camera could handle, only about 8% have a task where automating would actually pay off. The distance between “could be done” and “worth doing” is where the whole argument lives.
Why the costs tip the balance back to people
An AI system is not a one-off purchase. It has to be built, trained on the right data, fitted into how a workplace already runs, and then kept working year after year. The paper uses a hypothetical small bakery to make this concrete. Checking the quality of ingredients is a small part of a baker’s day, and the time and wages saved by adding cameras and an AI system fall well short of what the upgrade would cost.
For a large employer automating a high-volume task, the numbers can work. But for the many jobs where a vision task is just one part of a varied day, done by a handful of people, the fixed cost never gets spread thin enough to beat a wage. Small businesses especially rarely have the volume to justify a system that has to be right almost every time.
This is what gets missed when capability is treated as destiny. The bakery, the appraiser, the teacher with a task a camera could technically handle are not being kept on out of sentiment. In many cases they’re kept on because, at today’s prices, replacing them would cost more than they’re paid.
What would actually change the math
None of this means the workforce is frozen in place. The authors are clear that job losses are coming; their point is about pace and order. “Overall, our findings suggest that AI job displacement will be substantial, but also gradual,” they write, “and therefore there is room for [government] policy and retraining to mitigate unemployment impacts.” Gradual and gated by cost is a very different forecast from mass job loss any minute now.
Two things would speed it up. One is falling costs to build and run these systems, though even that is slow. Thompson notes that even if costs drop fast, by around 20% per year, it would still take decades for many vision tasks to become cheap enough to automate. The other is AI sold as a shared service, where the heavy build cost is split across many customers instead of paid by each company alone. As Thompson put it, “What we’re seeing is that while there is a lot of potential for AI to replace tasks, it’s not going to happen immediately.”
Where it does happen, it won’t land evenly. The study projects that “there will be more automation in retail and healthcare, and less in areas like construction, mining or real estate,” Thompson said, based on where the costs favor machines.
Two caveats keep this honest. The paper covers computer vision only, not the large language models behind the recent wave of chatbots, so it says nothing about text-and-language work. And as far as we can see, the analysis has not yet been peer-reviewed, which makes it a serious contribution rather than a settled verdict. Read it as a correction to a lazy assumption, not a promise about your particular job. The useful shift is from asking whether a machine could do a task to asking whether, all costs counted, anyone would actually pay it to.