Much research about the economic trade-offs of artificial intelligence has focused on the relatively marginal cost of production using AI compared with employing people. That’s consistent with how economists have explored previous types of technology, according to Danial Lashkari, PhD ’11, an affiliate researcher at MIT FutureTech.
However, that approach fails to account for the fixed cost of the resources necessary to get started with AI. “Unlike other technology, there’s a substantial training cost for AI that doesn’t rise with scale of production,” he said.
A new paper from Lashkari, FutureTech colleagues Wensu Li and Neil Thompson, and Christina Qiu at the University of Miami looks at the fixed and marginal costs associated with tasks automated using computer vision technology.
The researchers found that, given the up-front cost of training models, AI’s advantage is best realized when production occurs at a large scale. As a result, bigger enterprises are far more likely to be early adopters than smaller companies with fewer resources.
Right now, scale helps AI investments pay off
Companies considering task automation have typically weighed the marginal costs of human labor against the cost of machines. On this metric alone, AI would appear to be an easy sell, said Thompson, a principal research scientist at the MIT Sloan School of Management and director of MIT FutureTech. That’s especially true given the rapid progress that foundational AI models have made since their debut earlier this decade.
For most use cases, though, AI models aren’t plug-and-play. “There’s a lot of customization that has to happen for an individual company,” Thompson said. “If you have 1,000 products or processes that OpenAI has never seen, that often requires tailoring the model for your specific case. That tailoring costs something.”
The FutureTech research team investigated whether the fixed cost of training an AI model, the anticipated future drop in compute costs, and the marginal cost of human vs. machine “shifts the frontier of automation” to the point where AI is a cost-effective substitute for human labor.
The researchers considered 659 tasks that were capable of being automated by computer vision as defined by O*NET, a U.S. Bureau of Labor Statistics database of occupational information. This category of tasks was chosen because the main requirements for computer vision automation are mapped to two factors that can be easily measured: the number of potential outcomes and the acceptable error rate. Monitoring a retail chain’s in-store security cameras has few potential outcomes and a high error-rate tolerance; detecting tumors in radiological images is quite the opposite.
The researchers found that scaling AI accounted for 63% of its comparative advantage from one task to another across computer vision tasks in 2023 — and would still account for 60% of the variation across tasks in 2050.
The adoption rate for computer vision automation also reflected the advantage of scale. In 2025, fewer than 5% of firms had adopted the technology, though that figure exceeded 12% when firms were weighted based on number of employees. The researchers project an unweighted adoption rate close to 23% by 2035 but an employment-weighted adoption rate of nearly 60%. The reason: Larger companies that can scale the technology are likely to adopt computer vision earlier and more extensively than smaller ones.
“There are the fixed costs to get the system working, and the fixed costs dominate,” Thompson said. A multinational retailer has the resources to train a computer vision model to monitor cameras in thousands of stores to ultimately reap the benefits of efficiency gains. The local corner store cannot afford to make the same investment.
AI Executive Academy
In person at MIT Sloan
Register Now
Real output gains will lessen over time
Adoption will rise over time as compute prices fall and models become easier to fine-tune. Pushing these trends forward, the researchers predict that real output will increase over the next decade, peak in 2040 as technology matures, and then increase much more slowly.
There are three explanations for this anticipated slowdown.
- There are only so many tasks that computer vision can automate, Qiu noted. Once an organization has automated all of the tasks that computer vision can accomplish, there will be little additional opportunity for output growth. “The slowdown is almost mechanical,” she said.
- The human employees that remain will become harder to replace as more work is automated. Computer vision may be able to indicate the presence of a tumor, but it takes a radiologist with a medical background to determine malignancy and effectively communicate with a patient, Lashkari said. Also, some roles — like backcountry trail guide, for example — are filled by few humans to begin with, and their expertise would be hard to replicate.
- Patterns of automation adoption by company size, industry, or geography are going to evolve substantially, Lashkari said, noting, “The current state may be indicative of what’s happening in 10 or 20 or 30 years.” What’s more, automation is likely to proceed faster if model training costs are shared among firms within an industry rather than concentrated within a single firm.
Computer vision is one part of a broader automation framework
It’s worth noting that the paper focuses exclusively on computer vision tasks, which represent about 3.5% of total tasks in the economy as defined by O*NET. That means it doesn’t account for more dynamic drivers of gains, Qiu said. These could include the introduction of new tasks that computer vision can automate or the emergence of new economic sectors where computer vision automation can take hold quickly.
That’s why the researchers see their work on computer vision automation as one part of a broader framework, Lashkari said. Subsequent research will examine large language models, which are poised to help automate complex tasks with many potential outcomes and low error-rate tolerance. (To continue with the radiology example, an LLM could assist with analyzing tumors identified by computer vision models.)
“Our approach is to break down a big problem into constituent components,” Lashkari said. “We want to look at computer vision and automation in parallel with the question of how AI could be used to create new methods of production and scientific progress.”
Read the paper: “AI, Scale, and Task-Based Theories of Automation”
Neil Thompson is a principal research scientist at the MIT Computer Science and Artificial Intelligence Laboratory and the MIT Initiative on the Digital Economy, where he is head of the Artificial Intelligence, Quantum and Beyond research group. He is also the director of the MIT FutureTech lab.
Danial Lashkari is an affiliate researcher with MIT FutureTech and an associate professor in the department of economic development at the London School of Economics. He studies different long-run drivers of economic change, including technological breakthroughs such as IT and AI, and the innovation activity of firms.
Wensu Li is a research scientist at MIT FutureTech. Her research focuses on the economic implications of technical change, with particular interest in labor markets, automation, and the deployment of AI technologies. Her recent work explores topics such as partial vs. full automation and the cost-effectiveness of computer vision systems and LLMs.
Christina Qiu is an MIT FutureTech affiliate researcher and an assistant professor of economics at the University of Miami. She studies how digital technologies reshape economic outcomes, including in developing countries where markets and institutions remain informal.
