AI Coding Tools Boost Developer Output 180%, But Releases Rise Just 30%

Artificial intelligence tools are dramatically accelerating how fast software developers write code, but those productivity gains largely evaporate by the time applications reach users, according to new research from MIT Sloan School of Management.

A study examining more than 100,000 developers on GitHub found that AI coding assistants increased coding activity by up to 180% cumulatively across different tool types. Yet this surge translated to only 50% more projects and just 30% more actual software releases compared to developers working without AI assistance.

The research, conducted by MIT Sloan associate professor Mert Demirer along with Leon Musolff of the University of Pennsylvania and Liyuan Yang of Boston University, highlights a critical disconnect: human processes and bottlenecks remain unchanged even as AI accelerates early-stage work.

Why it matters

Organizations investing in AI coding tools expect productivity improvements to flow through to business outcomes. This research demonstrates that without redesigning workflows and addressing downstream constraints, companies will capture only a fraction of AI's potential value. The findings apply beyond software development to any knowledge work where AI automates one stage but leaves manual decision-making processes intact.

How AI tools perform across development stages

Modern software development unfolds in six stages, from writing code to releasing finished products. The researchers evaluated three categories of AI tools:

• Autocomplete tools that suggest code as developers type, boosting activity by 40%
• Sync agents that write and edit code in real time alongside developers, adding 100 percentage points for a cumulative 140% increase
• Async agents that work autonomously on assigned tasks, pushing the cumulative effect to 180%

These gains materialized primarily in early stages like writing and organizing code. Later stages involving code review, merging changes, and launching applications remained constrained by human processes.

"We're hearing they can do in a matter of minutes what used to take an entire day," Demirer said, referring to coding tasks. But technical and mechanical bottlenecks that require human intervention prevent those early gains from flowing through.

More apps, but not more users

Analyzing data from four popular app stores, researchers observed a broad increase in new applications following the early 2025 release of agentic AI coding tools. However, app downloads and user reviews showed no corresponding rise.

This pattern suggests AI is enabling developers to ship more software, but much of it attracts little or no user base. "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," the researchers write.

Breaking through the bottlenecks

Demirer points to two strategies for capturing more value from AI coding tools:

First, organizations should reduce team sizes since AI boosts individual output. Smaller teams face fewer coordination challenges and can move faster. Companies can then reallocate resources to address end-of-cycle bottlenecks like merging projects, releasing products, and ongoing maintenance.

Second, leaders should identify where AI can facilitate human workflows beyond just automating individual tasks. This might include generating prototypes, summarizing meetings, or streamlining approval processes.

"Look at where people are spending too much of their time and see how AI can resolve that," Demirer said. The key is deploying AI tools that ease downstream constraints rather than simply accelerating one stage while leaving manual decision-making workflows unchanged.

The findings were first reported by MIT Sloan in the paper "Writing Code vs. Shipping Code — Productivity Effects Across Generations of Coding Tools."