A Harvard-affiliated study by researchers Fiona Chen and James Stratton, reported by TechRadar Pro, looked at what AI actually changes in software development. Its finding comes in two parts: with AI, developers produce significantly more code, but the volume of software actually completed does not increase in a statistically significant way. The reason put forward: human code review is becoming the new bottleneck.

What the study measures

According to TechRadar Pro, the authors observe an increase of around 30% in the amount of code produced when developers use AI. In detail, lines of code rise by 30%, commits by 20% and pull requests by 23%. In other words, all upstream activity measures increase by 20 to 30%.

The most interesting result, however, lies elsewhere: the researchers detect no statistically significant increase in the volume of software actually completed. It is this gap between what is written and what is delivered that they sought to explain.

AI agents and human review: the new bottleneck

The explanation lies in the step that follows writing the code. According to the study, the time needed to review and merge pull requests increases by 49%, the number of comments left by reviewers by 35%, and the share of employees taking part in reviews by 14%. On average, the review process goes from 7.03 days to 10.48 days.

The mechanism is easy to understand: every proposed change must be reviewed and approved by a human before being merged. If AI multiplies the proposals, the reviewers' queue grows accordingly. "Vibe coding", the practice of letting AI write a large part of the code, thus seems to move the bottleneck further down the chain rather than remove it. The study's conclusion, as summarised by TechRadar Pro, points in this direction: productivity gains are absorbed by downstream constraints, to the point of making code review the new sticking point.

Findings, hypotheses and open questions

It is important to distinguish what the study observes from what it suggests. The increases in code, commits, pull requests and review time are measurements. On the other hand, the idea that AI-generated code requires substantially more corrections is presented as a claim by the researchers, which raises an obvious question about the real scale of the productivity gain.

Other points remain at the question stage. The researchers ask whether this shift in workload is leading companies to assign more employees to code review. They also wonder about the next phase of AI adoption: AI systems are already used for some review tasks, but humans remain heavily involved in validation. Nothing in the sources indicates that AI can currently take over this step in their place.

On employment, the study indicates that AI adoption does not appear to be accompanied by a significant reduction in jobs in the data studied. This finding applies only to that data and does not allow conclusions to be drawn for the sector as a whole.

The limitations to keep in mind

The TechRadar Pro article specifies neither the exact title of the publication, nor the sample size, nor the companies or period studied, nor how AI use was measured. Nor does it say whether the study was peer-reviewed. Yet these elements are essential for judging the scope of the results. It should also be noted that the summary published by TechRadar Pro presents some figures slightly differently depending on the passage; here we use those from the body of the text, which are consistent with the change from 7.03 to 10.48 days.

What this changes for development teams

For a team adopting AI agents, the lesson is first and foremost organisational: measuring productivity by the number of lines or commits is not enough, since these indicators rise without delivery keeping pace. What matters is the time between writing a change and putting it into production, and the review capacity available to absorb it.

According to TechRadar Pro, the issue is therefore less about replacing developers than about the best way to use these tools to help employees and eliminate unnecessary tasks. The human role does not disappear: it shifts towards review and validation.

Sources