AI boosts worker productivity — but does it increase final output?
cybersentinelnews.com · post 61067 · 2026-09-09
CAMBRIDGE, Mass., Sept. 9 — Artificial intelligence tools are dramatically speeding up the way software developers write code, but a new study from MIT Sloan School of Management finds that early-stage gains largely evaporate by the time a product actually ships. The research, led by MIT Sloan associate professor Mert Demirer, University of Pennsylvania assistant professor Leon Musolff, and Liyuan Yang (formerly of MIT Sloan, now a PhD student in economics at Boston University), tracked more than 100,000 developers on GitHub.
It examined how three categories of AI tools affected productivity across six stages of software development: writing code, organizing it into files, committing changes, submitting pull requests, merging changes, and releasing finished products. The numbers tell a striking story of acceleration at the front end.
Autocomplete tools increased coding activity by 40%. Adding sync agents pushed that cumulative gain to 140%, and further use of async agents raised it to 180%. Yet those early-stage surges were muted by human bottlenecks later in the pipeline — reviewing changes and launching apps, for instance.
The boosted coding work translated into only 50% more projects and 30% more actual releases compared with developers not using AI. Demirer noted the apparent magic of the tools: “We’re hearing they can do in a matter of minutes what used to take an entire day.” But he added that “some bottlenecks are truly technical and mechanical tasks” that humans must complete because AI isn’t good at them.
The researchers point to unchanged human processes as the reason for the disconnect. 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. Looking at data from four popular app stores, the team observed a broad increase in new apps since early 2025, when agentic AI coding tools became generally available.
But there was no corresponding rise in app downloads or user reviews — a sign that supply-side productivity is producing applications with little or no user base. Demirer recommends making teams smaller since AI boosts productivity, reallocating resources to address end-of-life-cycle bottlenecks such as merging projects, releasing products, and maintaining and updating them. He also suggests companies decide whether to reskill workers or hire new ones.
The paper’s findings reflect the limitations of workflows designed before AI; a determining factor for its impact, the authors conclude, is whether tools ease downstream constraints rather than simply automating one stage.