AI coding agents added 23% more pull requests but no more finished features. Review is the bottleneck in your automations too

More output is not more finished work. A new study of 718 software firms shows exactly where the gains disappear.

What happened

Harvard researchers Fiona Chen and James Stratton studied 300 million work events from the engineering analytics platform Jellyfish, covering more than 700,000 employees at over 700 firms from 2021 through March 2026, as reported by Ars Technica.

After firms introduced AI coding agents, total lines of code rose 30 percent, commits rose 20 percent and pull requests rose 23 percent. But the resolution rate for Jira issues and epics, the actual features shipped, did not change in a statistically significant way.

The reason was review. The time between a pull request being submitted and merged grew 49 percent on average. The share of pull requests with changes requested nearly doubled, and comments per pull request rose 35 percent. The share of workers doing code review went up 14 percent. AI review did not close the gap: although 80 percent of firms used some AI code review by March 2026, AI agents wrote only 23.3 percent of review comments.

The paper's abstract describes this as a bottleneck from code review that limits how much of the productivity gain passes through to output.

My take

This is not just a software story. I see the same pattern in business automations.

An AI step that drafts 200 follow up emails, CRM notes or proposal sections a day looks productive. If a sales manager has to check each one, you have moved the work, not removed it. Worse, if quality is uneven, every draft gets read more carefully, which is exactly what the study found with code.

So when I design an AI workflow, I budget for review as its own stage:

1. Measure finished outcomes, like replies sent or deals updated correctly, not drafts produced.
2. Route by confidence. Clear cases go straight through with logging, unclear ones go to a person.
3. Make review fast. Show the reviewer what changed and why, in one screen, with approve and edit buttons.
4. Keep batches small enough that a person actually reads them instead of rubber stamping.

The win from AI comes when the whole pipeline moves faster, and the slowest step decides that.