The AI Coding Paradox: Velocity Gains Erased by Review Bottlenecks
AI-generated image · US National Wire
New research suggests that while AI agents accelerate code generation, the resulting surge in volume is creating a technical debt bubble that neutralizes productivity gains.
For enterprise leaders, the promise of AI coding agents has been centered on velocity. The theory is simple: automate the writing process to ship software faster. However, new data suggests this speed is an illusion, as the operational cost of verifying AI-generated code is creating a massive bottleneck in the production pipeline.
As Ars Technica first reported, a study conducted by Harvard University researchers Fiona Chen and James Stratton reveals that the efficiency gains realized during the initial coding phase are being "absorbed" by downstream constraints. While AI agents can churn out functional code at scale, the human effort required to review that code for accuracy has surged, meaning firms are not seeing a statistically significant increase in actual software output.
**The Volume vs. Value Gap**
The study utilized aggregated analytics from Jellyfish, analyzing 300 million work events—including commits and pull requests—across more than 700 software development firms and 700,000 employees from 2021 through March 2026. The findings highlight a stark disconnect between raw output and completed features:
* **Increased Volume:** The introduction of AI coding agents led to a 30% increase in total lines of code, a 23% rise in pull requests, and a 20% increase in total commits. * **Stagnant Output:** Despite the volume spike, the resolution rate for "Issues and Epics" (the wholesale software features tracked in tools like Jira) did not change in a statistically significant way.
**The Review Bottleneck**
From an operational lens, the "speed" of AI is being neutralized by a ballooning review process. Ars Technica reports that the average time between a pull request submission and its merge increased by 49% after AI agents were introduced.
This friction is driven by a decrease in initial code quality. The researchers found that the number of comments per pull request rose by 35%, and the share of pull requests requiring changes nearly doubled. To manage this influx of problematic code, firms saw a 14% increase in the share of workers dedicated to performing code reviews.
**The Human Cost of Automation**
While 80% of the firms studied implemented some form of AI code review by March 2026, the technology has not yet automated the verification process to a meaningful degree. AI agents were responsible for only 10.8% of all pull requests and 23.3% of all review comments, leaving the vast majority of the critical verification work to human developers.
Ultimately, this shift has not resulted in a leaner workforce. By cross-referencing Jellyfish data with LinkedIn, Chen and Stratton found no significant employment changes attributable to AI. For the B2B operator, the takeaway is clear: increasing the volume of code without a corresponding increase in verification efficiency does not result in more software—it simply results in more code to review.