AI Coding Tools Boost Output 180% But Software Releases Lag at 30%

The productivity paradox in AI-assisted development

Artificial intelligence coding assistants are delivering dramatic productivity gains for individual developers, but those improvements are evaporating before they reach customers. New research from Wharton and MIT shows that while the most advanced AI tools increase coding activity by 180%, actual software releases grow by just 30%.

The study, conducted by Wharton professor Leon Musolff alongside MIT researchers Mert Demirer and Liyuan Yang, tracked more than 100,000 developers on GitHub from 2022 to 2026. The researchers measured productivity changes as developers adopted three successive generations of AI coding tools: autocomplete systems, synchronous editing agents, and autonomous async agents.

The gap between code generation and finished software points to a fundamental shift in software development constraints. Writing code is no longer the primary bottleneck—reviewing, integrating, and releasing that code now represents the binding constraint on productivity.

Why it matters

This research challenges the assumption that AI coding tools will automatically accelerate software delivery at the same rate they speed up code writing. For technology leaders planning AI adoption strategies, the findings suggest that investment in coding assistants must be paired with solutions for downstream bottlenecks. Companies may need to rethink their entire development pipeline, not just the coding phase, to capture the full value of AI tools.

Where the bottlenecks emerge

The researchers combined public GitHub activity records with Microsoft data on tool adoption to identify precisely when developers started using AI assistants. Autocomplete systems that suggest the next line of code increased coding activity by 40%. Synchronous agents that edit alongside developers in real time pushed the cumulative gain to 140%. Autonomous agents working from prompts reached 180%.

Yet even that 180% surge in coding activity translated to only a 50% increase in software projects and 30% more releases. "In software, the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it," the authors wrote.

Musolff acknowledges the findings could seem discouraging. "If the world froze at today's level of AI capabilities, these results would be a bit of a cold shower," he said. But he notes that each new generation of tools tackles later stages of the development process, and a 30% increase in releases still represents substantial gains.

The user adoption problem

Even successfully released software faces another hurdle: finding users. The researchers examined the four largest software marketplaces—Apple App Store, Google Play Store, Chrome Web Store, and SourceForge—and found a surge in new applications since mid-2025. Monthly new releases on Apple's App Store jumped from around 30,000 to roughly 100,000 by April 2026.

Total usage across these platforms, however, remained flat or declined. "It could simply be that it's much harder to discover new applications when there's such a flood of them," Musolff said. "Alternatively, even once you've shipped an app, there's another skill involved: iterating with users."

Some companies are developing AI tools to review machine-generated code, but Musolff questions whether AI can match human judgment in this role. "If the same AI that wrote the code also reviews it, that doesn't really solve the problem. The review just isn't of the same quality," he said.

The findings were first reported by Knowledge at Wharton, the business journal of the Wharton School at the University of Pennsylvania.