AI writes code faster, but software doesn’t ship faster

AI coding agents can help developers produce code much faster, but a new Harvard working paper suggests that faster coding does not automatically mean teams deliver more finished software. The research points to a possible bottleneck in code review, where the additional changes generated by AI require more time and effort from human developers.

What the Study Found

Researchers Fiona Chen and James Stratton analyzed approximately 300 million engineering events from 718 firms and 725,938 workers between January 2021 and March 2026. Their research examined coding activity, pull requests, software issues, project tasks, and the adoption of AI coding tools.

After firms adopted AI coding agents, the researchers estimated a 30% increase in lines of code, a 20% increase in commits, and a 23% increase in pull requests.

However, those increases did not clearly translate into more completed software work. The researchers found no statistically significant increase in the measured outcomes for resolved software issues and completed project tasks. This does not prove that AI produced no productivity benefits; it means the study did not establish a statistically significant improvement in those outcomes.

Why Code Review Slows Teams Down

The main challenge was not generating code but handling the additional changes afterward. Following the adoption of AI coding agents, the average time between submitting a pull request and merging it increased by an estimated 49%.

Review time increased by approximately 3.45 days, from a baseline of 7.03 days to about 10.48 days.

Reviewers also requested changes on nearly twice as many pull requests, while comments per pull request increased by 35%. The share of workers involved in review activity also rose by approximately 14% relative to the baseline.

These findings apply specifically to AI coding agents. The researchers did not find a statistically significant increase in review time following the adoption of AI coding assistants.

In practical terms, AI can generate code quickly, but developers still need to determine whether the changes are correct, secure, compatible with the existing system, and ready to merge. When the volume of generated code grows faster than a team's review capacity, pull requests can spend longer waiting for feedback and approval.

Does This Mean AI Coding Agents Do Not Work?

No. The study does not prove that AI-generated code is inherently poor quality or that AI agents cannot help build production-ready software.

Instead, it shows that, across the firms studied, increased coding activity did not produce a statistically significant increase in the measured indicators of completed software work.

Nor does it measure every aspect of software success, such as user satisfaction, reliability after release, or the business value of individual features.

The findings should therefore be interpreted as evidence of a potential bottleneck in software production—not as proof that AI coding tools are ineffective.

The key lesson is that AI coding tools should be judged by delivered software, not just code volume. A team may see more commits and pull requests, but that does not guarantee faster releases or better products.

Teams adopting AI agents should invest in:

• Strong automated testing and CI/CD pipelines

• Clear code-review standards and review ownership

• Smaller, easier-to-review pull requests

• Better validation of AI-generated changes before merge

• Tracking completed issues and releases, not only lines of code

AI can accelerate the first stage of software development, but shipping reliable software still depends on review, testing, integration, and release. The real competitive advantage comes from improving the entire engineering workflow—not just writing code faster.