One of the most attractive promises of AI-assisted software development is that significantly more code can be produced in the same amount of time. AI coding assistants and increasingly autonomous development agents can indeed help a lot in achieving this. But does faster coding really go hand in hand with faster software development?

Let’s start with the fact that AI tools for coding do not all work the same way. Assistants typically help by producing code snippets and supporting the developer’s work, while AI agents can carry out longer, more complex tasks partly on their own. The difference also shows up in measurements.

Fiona Chen and James Stratton, researchers at Harvard University, examined the development activity of 718 companies based on an analytics platform’s database of more than 300 million events. The data covered the activity of some 700 000 employees between January 2021 and March 2026. The study examined the impact of introducing coding assistants and AI agents separately. After agents were introduced:

  • the number of lines of code created rose by 30%;
  • the number of commits rose by 20%;
  • the number of pull requests rose by 23%.

At first glance, these look like major performance gains. The authors, however, did not treat the code written as the final outcome. Development-related Jira tasks counted as done when the corresponding code was deployed to production, and closing an epic marked the completion of a larger feature or development unit. Here the study found no statistically significant increase. So the picture changes significantly when we look not only at the production of code but at the entire development process.

In other words: more code and more changes were produced, but this did not lead to a measurable, proportional increase in completed development work. This does not prove that the involvement of AI agents brought no benefit at all. It shows that, based on the data examined, the significant gain in coding output attributable to AI agents did not translate into a proportional, reliably measurable increase across the entire software delivery process.

What does this mean in practice? Imagine a development team that, with the help of AI, can make more changes and write more code in the same amount of time. What happens next? If several of those changes need fixing, or integrating them into the system requires more review and coordination, faster coding does not necessarily result in more completed development tasks.

Finished code is only an intermediate stage. The real result is when a change meets the business requirements, runs reliably, passes all checks without errors and can be safely integrated into the existing system and infrastructure.

That is why it is not enough to measure the impact of AI by the amount of generated code or the number of commits. We also need to see how many tasks actually make it to completion, how long it takes and how much rework they need.

In the next part, we look at what happens in the development process when code is produced faster than before, but review and fixing do not speed up along with it.

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!