A new software-engineering paper on arXiv gives enterprise AI coding mandates a more useful question than "does AI write code faster?" The better question is what happens after the code arrives.

The July 2026 study, titled AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise 2x Mandate, analyzes an AI-forward company that pushed to double merged pull requests per engineer. The authors studied 802 developers and 196,212 pull requests from January 2024 through April 2026.

The headline number is striking. Per-developer throughput eventually reached 2.09 times the pre-mandate baseline in April 2026. The authors link part of that gain to adoption and accumulated use of AI coding tools, while warning that tool usage was not randomly assigned. In other words, the study is strong evidence about an observed rollout, not a clean laboratory proof that any company can flip the same switch and get the same result.

That caveat matters because the paper's most useful finding is not simply the throughput lift. It is the shape of the new constraint. The study says the gains were broadly shared across seniority levels, concentrated more in newer code than legacy repositories, and not clearly separable across model generations. The organizational change was bigger than a model upgrade story.

The review system had to absorb the extra volume. Per-reviewer load roughly doubled, automated review overtook human review, and merge and revert rates stayed broadly steady on the short-term measures the authors could observe. The authors also report that AI-authored pull requests took longer from first human review to merge, suggesting the delay moved downstream into review rather than disappearing.

That is the lesson for engineering leaders. A team that measures only merged pull requests can miss where the work went. More generated code still has to be reviewed, understood, tested, secured, owned, and maintained. If review queues, incident rates, revert rates, security checks, and long-term maintainability are not tracked alongside throughput, the productivity story can look cleaner than the system actually feels.

The study does not say AI coding tools are fake productivity. It says the productivity curve depends on adoption, repetition, and complementary process changes. It also says automation is now part of the review path, which makes review quality and governance more important, not less.

For software teams, the practical takeaway is straightforward: do not govern AI coding by merge count alone. Pair authoring speed with review capacity, queue depth, human-review coverage, incident and revert signals, and ownership health. The bottleneck has moved, and the dashboard has to move with it.

Sources:

https://arxiv.org/abs/2607.01904

https://arxiv.org/html/2607.01904