AI boosts worker productivity — but does that translate to final outputs?

Organizations are eager to find out whether artificial intelligence can boost worker productivity. Studies involving accountants and manufacturing firms have shown that it can — and new research has found that AI is doing the same for software developers.

“We’re hearing they can do in a matter of minutes what used to take an entire day,” said Mert Demirer, an associate professor of applied economics at the MIT Sloan School of Management.

But this boost in productivity doesn’t necessarily translate to final outcomes. A new paper by Demirer and University of Pennsylvania assistant professor Leon Musolff and Liyuan Yang, formerly of MIT Sloan and now a PhD student in economics at Boston University, concluded that although AI tools substantially boost productivity when developers write code, the effect is muted when it comes to shipping code.

The researchers found that developers using AI tools were able to increase their coding activity, with a cumulative effect of up to 180%. But those gains diminished throughout the production cycle, and the boosted coding work led to just 50% more projects and 30% more actual releases compared with developers who didn’t use AI.

Additionally, while adoption of AI tools translated to an increase in releases to popular app stores, it didn’t lead to a corresponding boost in app downloads or user reviews. 

The researchers said that human processes — which remained unchanged — accounted for the disconnect between AI productivity gains and actual outcomes. “Some bottlenecks are truly technical and mechanical tasks” that humans have to step in and complete because AI isn’t good at doing them, Demirer said. 

AI tools work well early in software development

Modern software development tends to unfold in six stages: writing lines of code, organizing code into files, committing changes to code, submitting pull requests to review changes, merging changes into a project, and releasing the project as a finished product.

Researchers looked at data from more than 100,000 developers on GitHub and assessed how their use of three categories of AI tools impacted productivity in each stage of software development:

• Autocomplete tools, which suggest code as the developer types 
• Sync agents, which write and edit code in real time alongside the developer
• Async agents, which can be assigned a task and work autonomously without oversight from the developer

Using the different AI tools improved developers’ productivity in writing code. Using autocomplete tools increased coding activity by 40%. The cumulative effect including sync agents boosted coding activity by 140%, and additional use of async agents boosted it by 180%.

Higher productivity doesn’t mean more products

The researchers found that the productivity gained from AI in earlier stages of development, such as writing code, were attenuated by human bottlenecks at later stages, such as reviewing changes and launching apps. Different AI tools required different layers of human involvement in later stages, so the end productivity effects varied. 

Researchers didn’t calculate an impact on software releases for async agents because, by design, they cannot release software directly.

Looking at data from four popular app stores, the researchers observed a broad increase in new apps since early 2025, which was when agentic AI tools for coding became generally available. However, there was no corresponding increase in app downloads or reviews.

This suggests that the increase in supply-side productivity is largely leading to applications with little or no user base, according to the researchers. “Developers must still test, polish, and iterate toward market fit, and these higher-level tasks may remain constrained even when AI lowers the cost of writing code,” they write.

AI Executive Academy

In person at MIT Sloan

Register Now

2 ways AI can break software development bottlenecks

Demirer said the paper’s findings reflect the limitations of workflows designed for an era that predates AI. The remedy is to determine where AI tools can give employees back their time.

An obvious first step is to make teams smaller, since AI tools boost productivity. Smaller teams are nimbler and present fewer collaboration and communication challenges.

From there, Demirer said, organizations can reallocate resources to addressing bottlenecks at the end of the life cycle. For software developers, this means merging projects, releasing finished products, and maintaining and updating products. An important decision follows: Reskill some workers, or bring in new hires for those tasks, keep all developers in their roles, and have more teams capable of working on more projects.

Leaders should also see where AI can facilitate human workflows, whether it’s in fleshing out prototypes, summarizing meetings, or reducing the red tape associated with approvals. The researchers concluded that a determining factor for the impact of AI tools on productivity is whether they’re able to ease downstream constraints, rather than just automating one stage of a process only for future stages to be held up by manual decision-making workflows. 

“As long as process aren’t defined by technical or institutional knowledge, AI tools can resolve them in a very particular way,” Demirer said. “Look at where people are spending too much of their time and see how AI can resolve that.”

Read the paper: “Writing Code vs. Shipping Code — Productivity Effects Across Generations of Coding Tools”

Mert Demirer, PhD ’20, is the Ford Foundation International Career Development Assistant Professor and an assistant professor of applied economics at the MIT Sloan School of Management. He works in the field of empirical industrial organization, and his research focuses on firm productivity, firms’ use of digital technologies, antitrust issues, and the productivity effects of generative AI.