MIT ran the numbers on automation, and the result surprised a lot of people: for most jobs, humans are still the cheaper option. Here’s what that research actually says, and when AI automation is worth your money instead of just your patience.
The study everyone keeps citing
It’s called “Beyond AI Exposure,” and it comes from MIT’s FutureTech group, led by researcher Neil Thompson with colleagues from MIT, the Productivity Institute, and IBM. The paper came out in early 2024, and it’s been making headlines ever since, with Bloomberg, TechCrunch, and Forbes all covering it. It even made the rounds again this week, because the question it answers hasn’t gone away: when does automating a job with AI actually save money?
Here’s what makes the study different from the usual AI hype. Instead of asking “what could AI do someday,” the researchers asked “what makes economic sense today.” That changes everything.
The numbers, in plain English
The study focused on computer vision tasks, the kind of work where AI looks at images or video and decides something: quality inspection, security monitoring, reading documents, spotting defects. Those are the tasks AI is genuinely good at. And even there, the economics are brutal.
Only 23% of the wages paid to humans for those vision tasks would be economically attractive to automate with AI. For the other 77%, hiring a human is still the smarter deal.
| Vision-task wages in the US | Share |
|---|---|
| Economically attractive to automate with AI | 23% |
| Cheaper to keep a human doing | 77% |
Thompson’s quote from the coverage says it all: “Humans are still the better economic choice for doing these parts of jobs.”
For normal businesses, these tasks live on spreadsheets and clipboards: logging receipts, counting stock, checking photos of deliveries. If your operation has any of those, the study is basically a pricing guide for whether AI is worth the setup.
Why most tasks aren’t worth automating
The study even tested a generous scenario. What if an AI system cost just $1,000 to set up? Still not worth it for a big chunk of tasks, especially low-wage, multitasking roles where the human is already cheap and the AI would need constant hand-holding. And here’s the kicker: even if AI costs keep dropping 20% per year, the researchers figure it would take decades before most of those tasks become economically efficient to automate.
That’s the uncomfortable truth about AI automation: the technology works, but the accounting doesn’t. It’s also the part nobody puts in the sales pitch.
The real cost stack of automation
Why is automation so expensive in practice? Because the sticker price is only the beginning. You pay to build the system, or buy it. Then comes integration with everything else you use, and that’s a project of its own. Running it is a monthly subscription, forever. On top of that, you pay people to babysit it when it does something weird, and you pay for the mistakes it makes while it learns. That’s the cost stack, and most business owners only see the first layer.
The pattern is familiar if you’ve ever bought any “automation” tool. The demo looks flawless. Then onboarding takes three weeks, the outputs need human review for a month, and the integration breaks whenever your other software updates. Suddenly the “$50 a month” tool has cost you $2,000 in setup time and it still can’t handle the edge cases your part-time employee handled without thinking.
When AI automation IS worth it
None of this means automation is a scam. It means the math has to work, and the math works when the workload has the right shape. AI automation pays off when:
- The volume is high. Thousands of similar tasks a week, not a dozen.
- The task is repetitive and stable. Same inputs, same process, same output, week after week.
- The data is clear. Text, images, structured records the AI can actually read.
- The human version is expensive. Specialist salaries, overtime, or tasks that take hours.
- Mistakes are cheap to catch. A human reviews the output, so errors don’t reach customers.
- The setup cost gets absorbed. One integration that pays for itself across thousands of runs.
Customer service triage, invoice processing, lead qualification, social media scheduling, report generation. Those are the shapes that fit. We mapped out the full ladder from simple helpers to autonomous workers in our 8 stages of AI automation guide, and the earlier stages are exactly where the economics work best.
A quick way to run the numbers
Here’s a back-of-the-envelope calc you can do in five minutes. Take one task and write down how many hours it eats per week, multiplied by what you pay per hour. Say reporting takes 8 hours a week at $35 an hour, that’s $280 a week, roughly $14,500 a year. Now price the tool, add setup time and a month of babysitting, and divide by a year. If the annual automated cost isn’t clearly lower, you stop here and save yourself the trouble. Most of the time, that’s where the MIT math kicks in and the spreadsheet lies: the automation was never going to pay for itself.
When it’s a waste of money
Flip the checklist and you get the danger zone. Low volume, judgment-heavy work, tasks that change every week, cheap labor already doing the job fine, and anything where an AI mistake is expensive or embarrassing. Automating those is how you spend $3,000 to save $200, and it’s a trap I see constantly in small business content.
There’s also a subtler version. Some tasks look automatable because they’re boring, but they’re actually a mix of five different jobs stuffed into one role. A receptionist answers the phone, greets people, schedules, handles mail, and mediates conflicts. An AI can do maybe two of those five. Splitting the role into “automatable” and “not automatable” parts is real work, and that work has a cost too.
That’s why our build vs buy framework keeps coming back to the same advice: start with one narrow, high-volume task, measure the time cost before you automate, and measure it again after. If the numbers don’t move, the automation was decoration, and the four roles framework explains where AI fits and where to skip it entirely. For anyone selling professional services, the AI sales automation guide shows a realistic before-and-after of what good automation looks like.
The takeaway
MIT’s research is forty years of common sense compressed into one paper: automation pays when it’s cheaper than the human, and usually it isn’t yet. So run the math before you buy. Pick one high-volume, repetitive task, count what it costs you today, and only automate when the numbers actually move. That’s the difference between AI automation that earns its keep and AI automation that just adds a subscription. You can dig into the original MIT FutureTech study and the TechCrunch coverage if you want the full picture, but the takeaway fits in one line: automate the boring, repetitive, high-volume stuff, and leave the thinking to humans for now.
One more thing: re-run this calc every year. AI costs fall around 20% annually, so a task that doesn’t pay for itself today might in eighteen months. The math isn’t a one-time decision, it’s a habit.
