In the previous part we saw that after AI agents were introduced, coding activity rose, but the amount of software delivered did not rise in proportion. Examining the development process in detail, the Harvard University study also identified one specific factor that may limit growth in the end result: code review.

It is not enough for an AI agent to help write the code. Code still has to be reviewed and tested, its behaviour has to be verified, and we have to make sure it fits into the existing system. Solutions produced by AI also have to be checked against the project’s requirements, its architecture and the expectations that often go unwritten but are fundamental to how the system works.

After AI agents were introduced, the researchers measured the following changes in code review:

  • the average turnaround time of pull request reviews rose by 49%. The baseline average was 7,03 days;
  • the share of reviews requesting changes nearly doubled. The baseline share was 13%, and it is estimated to have risen by 12 percentage points;
  • the number of comments per pull request rose by 35%. The baseline average was 1,66 comments.

Why can this become a bottleneck?

In the development process, writing code and reviewing code are complementary tasks. Faster coding can send more changes to the review stage, but review capacity does not necessarily grow at the same pace. Review is also not a simple formal check. The reviewer has to understand what the change is meant to achieve, whether it meets the requirements, whether it fits the existing codebase correctly and whether it causes problems in other parts of the system. If a shortcoming is found, the developer has to fix the solution, which then has to be checked again.

Expert accounts in the study also describe such situations: the code is ready quickly, but experienced people are available for review only to a limited extent. In one example quoted in the study, a junior developer reported that, because of the workload of the experienced engineer who approves the code, they sometimes had to wait up to a week for a review. This illustrates the heart of the problem well: faster code production does not remove the need for checking, and if review capacity is limited, part of the speed-up is lost to waiting and further rounds of fixes.

The study also showed that companies responded to the increased need for checking partly by reallocating the workforce available. After AI agents were introduced, the share of employees who performed code review rose by 14%. Alongside growing coding capacity, organising the review and involving the right experts has therefore become an important task, and human reviewers kept their central role even as AI-based review tools spread.

So AI does not necessarily remove the constraints of development. Instead, it may simply shift them: rather than writing code, the bottleneck of the process becomes review, bug fixing or integration. Therefore, the question is not simply how quickly AI can produce code. How we organise the checking process, and which tasks need what level of professional control, matters as well.

In the next part, we look at how AI can be used so that faster coding becomes a real advantage.

Source: Fiona Chen and James Stratton, Harvard University: Artificial Intelligence in the Firm: Bottlenecks in Software Production. This post is based on the version of 2026. 08. 04.

Was your software built with AI? Have it reviewed by an expert!