MIT IDE research group lead and MIT FutureTech Research Project Director Neil Thompson presents his research at the 2026 IDE Annual Conference in April.
The pressure around getting meaningful results from AI is no longer abstract. Today many organizations are putting AI to work—but most aren’t seeing what they expected.
For example, a 2025 MIT report finds that despite a nearly $40 billion investment in generative AI, 95% of organizations were getting zero return. Similarly, according to a 2026 working paper from the National Bureau of Economic Research, eight in 10 senior business executives say AI has had no impact at all on either their organizations’ employment or productivity.
What could lead to better results? New papers by MIT Initiative on the Digital Economy (IDE) researchers argue that AI automation is more complex and nuanced than many business leaders think. Rather than being a simple matter of replacing people with agents, they say, automation affects different skills, people and industries in different ways.
Automation Is Not a Switch
Many companies today are exploring whether AI can automate tasks and boost productivity. Instead of asking how to automate with AI, companies should be asking whether a task should be automated at all. And if so, to what extent?
MIT IDE Research Lead Neil Thompson and MIT IDE Research Scientist Martin Fleming are among the co-authors of a new working paper, Economics of Human and AI Collaboration that explores AI’s workforce impact. In it, they make the case for moving beyond the binary choice of automate or don’t automate. In its place, they offer a continuum of three possibilities: no automation, partial automation, or full automation.
“For two and a half centuries, we thought about automation as capital replacing labor,” Fleming said at the MIT IDE 2026 Annual Conference. “Our view is that we now ought to think about automation along a spectrum, from no automation to partial automation to full automation.”
To choose, the researchers explain, it helps to consider a task’s level of complexity. What’s the share of work time and subtasks that can be automated?
The Future of Work Is About Finding the AI Goldilocks Zone
In general, tasks with few subtasks and low complexity are good for high levels of AI substitution. Conversely, tasks with many subtasks and high complexity favor only limited automation.
Another important factor is cost. Developing and using AI models is expensive. And most organizations expect, even demand, a return on their investment (ROI).
Adding yet another layer of nuance, the researchers say cost is closely related to deployment scale. There’s often a “just right” level of automation. Beyond that point, the cost of scaling AI further can outweigh the benefits. In these cases, full automation is prohibitively expensive, even when it’s technically feasible.
One way around the cost crunch, the researchers advise, is spreading AI development expenses across many users. With more users benefitting, the ROI should be faster and higher. This can also expand the set of tasks that are economically viable to automate. In addition, as training data, model size and compute are scaled up, model performance should improve.
Another option is partial automation. By stopping short of full automation, an organization lets AI do the tasks it can do economically while leaving humans to do the rest.
“The notion of full automation is relatively limited to a relatively small number of tasks,” Fleming said. “Partial automation is pervasive across all of these tasks, and that becomes really the way we need to think about automation in the presence of artificial intelligence. It’s much more of a partial process.”
Yet another option is the emergence of third parties offering specialized AI functions as a service. For an example, consider a SaaS vendor that offers AI models designed for commercial banking.
To examine how these ideas work in the real world, the authors used data from the U.S. Department of Labor to explore the potential for AI computer vision. Surprisingly, they found that total share of U.S. labor compensation attractive for computer-vision automation is less than 4%.
This low number leads the researchers to conclude that AI will be transformative for specific occupations, especially those with standardized, low-complexity components. Yet, they write, “its aggregate labor market impact may be more gradual than purely exposure-based estimates suggest.”
AI Automation’s Rising Tide
Another common misconception about AI and employment is the idea that AI will hit us like a crashing wave, with sudden discontinuous jumps in capabilities even for tasks on which AI previously performed poorly. Instead, a team of researchers—it includes Thompson, Fleming and MIT IDE research scientist Matthias Mertens, postdoctoral associate Wensu Li, and statistician/data analyst Brittany Harris—finds that AI’s workforce impact more closely resembles a slowly rising tide.
That’s even the title of the researchers’ new working paper, Crashing Waves vs. Rising Tides. In it, they propose that AI automation follows a continuum between two extremes: crashing waves where AI capabilities surge abruptly over small sets of tasks, and rising tides where the increase in AI capabilities is more continuous and broad-based.
To determine which end of the continuum is more common, the researchers examined more than 3,000 broad-based tasks derived from U.S. Labor Department categories. All were text-based and therefore addressable by large language models (LLMs). Based on more than 17,000 evaluations by workers from these jobs, the researchers found little evidence of crashing waves, but substantial evidence for rising tides.
That’s a good thing, the researchers write, because crashing waves—where AI moves quickly from almost always failing to almost always succeeding—“would lead to harsh surprises for human workers.”
The research team shows that AI progress remains rapid. If current trends continue, AI could complete most text-based tasks at an 80% to 95% success rate by 2029 at a minimally sufficient quality level, they predict. This kind of broad, incremental “rising tide” pattern is ultimately good news, as it gives governments, workers, and firms time to adapt and prepare for steadily improving capabilities.
Finding the AI Automation That Saves the Most Time
Which industries are most likely to be affected by a rising tide of LLM automation? To find out, the same team of researchers collected survey data evaluating LLM performance on over 11,000 labor-market tasks drawn from government classifications.
Next, they used GPT-4 to screen tasks for whether doing them with AI had the potential to save at least 10% of the time it takes for a human. For each text-based task, the researchers constructed two subtasks that were then completed by 40 LLMs.
The output from the LLMs was then evaluated by humans with relevant job experience. They rated each model response on a scale of 1 to 9, with 1 indicating that a job needs to be redone from scratch, 7 indicating success without edits at a minimally sufficient quality level, and 9 indicating that a job was done with above-average performance.
The researchers also examine AI performance across occupational groups. While there are some differences in how well AI handles text-based tasks across occupations—for example, it performs somewhat worse on legal tasks—the central finding is that performance is surprisingly similar across many job families, consistent with a “rising tide” pattern. Across job categories, AI completes text-based tasks at a minimally sufficient quality level with success rates ranging from 47% to 73%.
As Thompson emphasized when presenting this paper at the IDE’s recent Annual Conference, task loss and job loss are not the same thing. Several steps need to happen to put this automation into practice. And full automation—getting to the “last mile”—is not imminent.
“This last mile problem is a big deal,” Thompson said. “You actually have to think about a lot of pieces in order to get this into the workplace.”
Will AI Automate Your Expertise?
Many companies have acknowledged that humans will need to remain in the loop for many job roles and many tasks. While AI does some things well, human expertise is harder to replace.
The paper, Expertise, was cowritten by David Autor, an MIT economist, and Neil Thompson. It digs further into this idea, asking the question: Does AI automation raise the value of labor in the tasks that remain, or lower it?
Autor and Thompson argue that it depends on whether removing a task raises or reduces the expertise required for the remaining non-automated tasks. Once a task is automated, the relevant expertise is no longer needed anywhere in the economy. But while the task may be relatively expert for one occupation, it may be one of the least expert tasks in another. Therefore, automation can simultaneously replace experts in some occupations while assisting experts in others.
“Automation,” Autor and Thompson write, “simultaneously replaces experts and augments expertise.”
The authors provide an example by comparing the fate of two jobs over the past 40 years: accounting clerks and inventory clerks. Routine tasks in both jobs could be automated by computers, but the expertise required for the remaining tasks differs. For accounting clerks, the tasks that cannot be automated—these include decision-making and problem-solving—require specialized knowledge and training. By contrast, for inventory clerks, the tasks that cannot be automated—such as counting and weighing inventory—require far less training and certification.
This nuance has an impact on both wages and job demand, the researchers write. When automation eliminates nonexpert tasks, it raises wages and reduces employment. By contrast, when automation eliminates expert tasks, it lowers wages and increases employment.
“Task loss is not the same as saying your wages are going to go down,” Thompson said. “You have to think about the expertise side of things.”
The authors say this framework also resolves the puzzle of why routine task automation has lowered employment but often raised wages in routine task-intensive occupations. Indeed, the researchers predicted that for accounting clerks, wages would rise while the number of jobs would fall. Conversely, for inventory clerks, the researchers predicted wages would fall as the number of jobs would rise. These predictions align with the actual wage and employment changes in these occupations.
Higher Wages or Higher Employment?
Looking ahead, the authors expect occupations that become more expert due to automation will see falling employment, but rising wages. Conversely, occupations that become less expert due to automation will see rising employment, but falling wages.
We see this already with the impact of GPS. It lowered the expertise level needed for driving, leading to the rise of ride-sharing drivers. Today there are far more Uber and Lyft drivers than taxi drivers. But where taxi drivers need deep knowledge of city streets—London cabbies still must pass “The Knowledge” exam, demonstrating their mastery of the city’s complicated streets, landmarks and routes—ride-sharing drivers are essentially freelancers dependent on GPS and paid by algorithm.
So the pressing question is not “will AI take your job?” Instead, ask whether AI will devalue your expertise. Alternatively, will AI make your expertise even scarcer—and therefore more valuable?
To be sure, these findings come from recent papers of MIT researchers, not immutable laws written in stone. And the AI environment is changing by the month, if not by the week or even day.
Yet their research points in one clear direction: Trying to lower costs by simply using AI to eliminate jobs doesn’t work. A more nuanced approach to implementing AI should lead to far better results.
Read the papers:
- Economics of Human and AI Collaboration by Wensu Li, Atin Aboutorabi, et al.
- Crashing Waves vs. Rising Tides by Matthias Mertens, Adam Kuzee, et al.
- Expertise by David Autor & Neil Thompson
Peter Krass is a contributing writer and editor to the IDE.