AI Coding Tools Triple Code Output But Barely Lift Shipped Software

A comprehensive study tracking more than 100,000 software developers has uncovered a striking disconnect: while AI coding assistants can nearly triple the volume of code developers produce, the number of finished software releases increases by only 30%. The findings, detailed in research by Demirer, Musolff, and Yang, offer one of the first empirical tests of why task-level AI productivity gains fail to translate into proportional increases in final output.

The research examined three generations of AI coding tools adopted between 2021 and 2026. Autocomplete tools that suggest code as developers type increased commit activity by roughly 40%. Synchronous agents like Claude Code, which write and edit code in real time alongside developers, pushed the cumulative effect to 140%. Asynchronous agents that work autonomously on assigned tasks brought the total gain to approximately 180% at the commit level.

Yet this dramatic expansion in coding activity attenuates sharply as work moves through the software production hierarchy. The same developers who nearly tripled their commits worked on only 50% more projects and shipped just 30% more releases. For sync agents alone, a more than sevenfold increase in raw lines of code translated to only a 20% rise in shipped releases.

Why it matters

This research directly challenges optimistic forecasts that extrapolate task-level AI gains to economy-wide productivity growth. When production stages are complementary—meaning output depends on multiple sequential steps—automating one stage provides only limited gains if human effort still constrains downstream stages. For business leaders investing in AI tools, the findings suggest that productivity improvements require addressing entire workflows, not just accelerating isolated tasks. The binding constraint in software development is shifting from writing code to the human-intensive work of reviewing, integrating, testing, and releasing it.

The bottleneck hypothesis confirmed

The researchers combined public GitHub histories with internal Microsoft usage data, using a matched event-study design that compared each adopter to a control developer with nearly identical activity one year earlier. This approach avoided the common pitfall of comparing adopters to supposed non-adopters who may be using AI tools without detection.

The pattern fits what economists call the "weak links" or bottleneck hypothesis: when production consists of complementary stages and AI accelerates only some of them, final output remains limited by the stages humans still perform. AI-written code still requires human expertise for review, integration, testing, and release—stages where judgment and coordination matter as much as speed.

Supply expands, usage doesn't

The researchers also examined four major application marketplaces including the Apple App Store and Google Play. Monthly new iOS app releases roughly doubled between early 2025 and April 2026, and other platforms showed moderate increases. However, total user engagement with each monthly cohort of new apps remained flat or declined across every marketplace examined.

By early 2026, more than 5% of public GitHub commits were attributable to Claude Code alone—a lower bound on AI's true share since most AI usage leaves no visible trace. Yet despite this surge in developer activity and new app releases, there is no detectable increase in software consumption.

The disconnect suggests either that marginal AI-era apps are lower quality, clearing the publication bar only because entry costs fell, or that consumer attention and app discovery represent additional bottlenecks that don't expand with supply.

While a 30% increase in shipped releases would be remarkable for any workplace technology, the gap between coding gains and final output confirms what Robert Solow observed about computers in 1987: "You can see the computer age everywhere but in the productivity statistics." Four decades later, his quip applies with renewed force to generative AI.

These findings were first reported by Mert Demirer, Lukas Musolff, and Lynn Yang in NBER Working Paper 35275 and discussed on VoxEU.