Artificial intelligence coding tools can increase developers' coding activity by as much as 180%, yet those gains translate into only a 30% rise in finished software releases, according to a new study from Wharton professor Leon Musolff and MIT researchers Mert Demirer and Liyuan Yang. The disconnect reveals that human bottlenecks in reviewing, integrating, and distributing code now overshadow the speed gains from AI-assisted writing. The research tracked more than 100,000 developers on GitHub between 2022 and 2026, comparing their output before and after they adopted three successive generations of AI coding assistance.
The productivity boost varied sharply depending on which generation of AI tool developers used. Autocomplete systems that suggest the next line of code lifted coding activity by 40%. When developers added "sync agents" — tools that edit code alongside them in real time — the cumulative gain jumped to 140%. "Async agents," which work independently from a simple prompt, pushed the total increase to 180%. Despite that dramatic rise in code generation, software projects grew by just 50%, and actual releases climbed only 30%. Researchers also examined the four largest software marketplaces — Apple App Store, Google Play Store, Chrome Web Store, and SourceForge — and found a sharp surge in new applications starting in mid-2025. On Apple's App Store, monthly new releases rose from roughly 30,000 before AI coding agents arrived in early 2025 to about 100,000 per month by April 2026, yet total usage across all four major app stores remained flat or declined.
The authors write in the paper that "in software, the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it." Musolff told Wharton that if AI capabilities froze at today's level, "these results would be a bit of a cold shower," yet he noted that the tools are improving rapidly and that a 30% increase in software releases represents gains "very few technologies you can invest in today" can deliver. The research found that AI can't yet match human judgment when it comes to code review. "If the same AI that wrote the code also reviews it, that doesn't really solve the problem," Musolff said, adding that "the review just isn't of the same quality."
The study suggests the gap between coding speed and finished software may narrow as each new generation of AI tackles later stages of the development process. If the technology can produce higher-quality code that requires less human oversight, the report notes, today's bottlenecks could prove temporary. But even when software ships, another hurdle emerges: user adoption. The flood of new applications may make discovery harder, Musolff said, or it could be that "even once you've shipped an app, there's another skill involved: iterating with users." The research combined public GitHub records with Microsoft data on developers' use of the tools to pinpoint when each developer first adopted the technology, allowing the team to measure productivity changes before and after adoption. The findings show that writing code is no longer the main constraint — unless AI can also automate the labor-intensive work of reviewing, integrating, and releasing software, the productivity revolution will remain incomplete. Companies banking on AI to transform development pipelines will need to solve for the entire workflow, not just the coding step, if they want the full promise of these tools to materialize. Organizations that treat code generation as the finish line rather than the starting gate risk discovering that their investment yields impressive activity metrics but disappointing business outcomes.