The AI Coding Paradox: Why More Lines of Code Aren't Delivering More Software

For enterprise leaders, the value proposition of AI coding agents has always been rooted in velocity. The promise is simple: automate the rote work of writing code to accelerate the delivery of software features. However, new data suggests that this efficiency is an illusion, as the time saved during the initial drafting phase is being swallowed by a critical bottleneck in the production pipeline: human verification.

As Ars Technica first reported, a study conducted by Harvard University researchers Fiona Chen and James Stratton reveals that while AI agents are highly effective at churning out volume, they are not necessarily increasing actual software output. The researchers analyzed aggregated analytics data from Jellyfish, covering more than 700,000 employees across 700 software development firms from 2021 through March 2026. The dataset included 300 million individual work events, such as commits and pull requests.

From a raw production standpoint, the AI tools appear to be a success. Data highlighted by Ars Technica shows that the adoption of AI coding agents drove an average increase of 30 percent in total lines of code generated, along with a 20 percent jump in total commits and a 23 percent rise in pull requests.

But for the B2B operator, the metric that matters is the resolution of "Issues and Epics"—the actual software features tracked in tools like Jira. The researchers found that the introduction of AI tools did not result in a statistically significant change in the resolution rate for these features. Furthermore, there was no "compositional shift" in the size or complexity of these issues, meaning firms aren't using the extra capacity to build more complex software; they are simply producing more code to reach the same end goal.

***Opinion:*** *From an ROI perspective, this is a red flag. If the output of finished software remains flat while the volume of intermediate code increases, the AI isn't accelerating the product roadmap—it is simply shifting the labor burden from the writer to the reviewer.*

The study identifies the code review process as the primary point of failure. After AI agents were introduced, the average time between a pull request submission and its merger into the codebase increased by 49 percent. The quality of the AI-generated output necessitates more rigorous human scrutiny: the number of comments per pull request rose by 35 percent, and the share of pull requests requiring changes nearly doubled.

This shift in workload is reflected in staffing. According to the study, the portion of the workforce dedicated to performing code reviews grew by 14 percent. Despite these shifts, the study found no evidence that firms reduced employment due to AI, a finding the researchers supported by cross-referencing Jellyfish data with LinkedIn records.

While 80 percent of the firms in the study utilized some form of AI code review by March 2026, the human element remains dominant. AI agents were responsible for only 10.8 percent of all pull requests and 23.3 percent of all review comments. Until AI can reliably review its own work, the "agentic coding" era may simply be a trade-off where coding speed is neutralized by the exhaustive effort required to ensure the code actually works.