The AI Coding Paradox: Why More Code Doesn't Mean More Software

For enterprise leaders, the value proposition of AI coding agents has always been centered on velocity. The promise is simple: automate the rote work of writing syntax to accelerate the delivery of software features. However, as Ars Technica first reported, a deep dive into the actual mechanics of the development lifecycle suggests that this velocity is an illusion. We aren't necessarily accelerating the production of software; we are simply shifting the labor cost from the writing phase to the review phase.

According to reporting from Ars Technica, a comprehensive study conducted by Harvard University researchers Fiona Chen and James Stratton highlights a critical friction point in the modern DevOps pipeline. By analyzing aggregated analytics data from Jellyfish, the researchers examined 300 million individual 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 suggest that while AI agents are highly efficient at generating code, they are not yet efficient at generating *finished* software.

### The Volume Trap

From a raw output perspective, the impact of AI coding agents is undeniable. Ars Technica reports that firms introducing these tools see total lines of generated code climb by 30 percent.

This growth is paired with an average 20 percent jump in total commits and a 23 percent increase in the number of pull requests.

In a traditional ROI model, these metrics would signal a massive leap in productivity. But in the context of enterprise software, lines of code are a vanity metric. The real measure of productivity is the resolution of "Issues" and "Epics"—the actual features and fixes tracked in management software like Jira. The Harvard study found that the resolution rate for these high-level deliverables did not change in a statistically significant way after the introduction of AI tools. Additionally, the researchers noted that the size and complexity of the tasks being finished showed no "compositional shift."

In short: firms are producing significantly more code, but they aren't delivering more software.

### The Review Bottleneck

The discrepancy between increased code volume and stagnant software output is explained by a massive surge in the human review cycle. Because developers cannot trust the accuracy of AI-generated output, the burden of verification has intensified.

As detailed by Ars Technica, the average "review process" time—the window between a pull request being submitted and its eventual merge into the codebase—balloons by 49 percent on average following the introduction of AI agents. The friction is evident in the granular data: the number of comments per pull request increases by 35 percent, and the share of pull requests requiring changes nearly doubles.

This creates a zero-sum game. The time saved by allowing an AI agent to autonomously write a block of code is effectively "absorbed" by the downstream requirement for a human to scrutinize that code more intensely. The result is a production process where the bottleneck has simply migrated from the keyboard to the review screen.

### The Labor Shift

If the goal of AI integration is to reduce headcount or drastically lower the cost of development, the current data suggests those goals are not being met. The Harvard researchers found that the share of workers performing code reviews increased by 14 percent after AI agents were introduced.

When cross-referencing Jellyfish data with LinkedIn records, the researchers found no evidence that firms were reducing employment or significantly increasing software output. The efficiency gains realized during the initial coding phase are being offset by the increased human effort required to ensure the code is functional and safe.

### The Limits of AI Review

One might argue that the solution is to deploy AI to handle the review process as well. While 80 percent of the firms in the study had implemented some form of AI code review by March 2026, the impact remains marginal. According to the data cited by Ars Technica, AI agents were responsible for only 10.8 percent of all pull requests and 23.3 percent of all review comments. The vast majority of the critical verification work remains a human responsibility.

### Analysis: The ROI Reality Check

From an operational lens, this suggests that many organizations are paying a premium to implement AI agents that don't actually shorten the lead time to delivery. If the total time from "idea" to "merged code" remains stagnant because the review cycle has expanded to meet the volume of AI output, the actual ROI of these tools is questionable.

We are seeing a transition where the role of the software engineer is shifting from a creator to an editor. While 95 percent of the firms in the study have already implemented AI coding agents, the industry is still in a learning phase. The long-term viability of these tools depends on whether teams can optimize the "coding time versus review time" trade-off. Until AI can reliably handle the review process with the same rigor as a human architect, the productivity promise of agentic coding will remain a mirage.