The AI coding productivity paradox: 30% more code, no change in delivery

2026.10.20Source · Industry research

Three reports published on the same day, 22 October, reach the same conclusion from different angles: AI coding agents raise both the volume of code and the review workload, but the value actually delivered does not rise in proportion.

The numbers do not line up

A Harvard working paper covering 718 companies and 300 million work events found that after adopting AI coding agents, lines of code rose 30%, commits 20% and pull requests 23%, while the completion rate of Jira issues and epics showed no statistically significant change. The paper puts the gap down mainly to review: the average time from PR to merge rose 49%, the share of PRs sent back for changes nearly doubled, and comments per PR rose 35%.

Review became the new bottleneck

LinearB’s benchmark across 8.1 million PRs is blunter: PRs submitted by AI agents wait an average of 17.6 hours for review — 5.25 times the 3.4 hours of ordinary PRs — and 32.7% merge within 30 days, against 84.4% for ordinary PRs. In another Faros AI dataset drawn from 22,000 developers, median PR review time rose 441%.

The bill for code quality

In the same set of reports, defect counts and incident rates per PR both rise. LocalStack’s survey found 65% of teams at least doubling code output, 54% reporting more defects in AI-generated code, and about 40% of AI-generated code never reaching production. The reason is not mysterious: the model does not understand your system’s conventions and architectural decisions, so what it writes “looks right”.

What it means for enterprises

  • Do not use lines of code or PR counts as productivity metrics; that is activity, not output.
  • Schedule review capacity as the scarce resource. Without more reviewers, AI only makes the queue longer.
  • Give AI-generated code its own quality gates: static checks, unit-test coverage and a limit on change scope — all three, not two.
  • Start with modules that have test coverage behind them; do not try it first on a legacy system with no regression safety net.

Want to talk through what you just read?

We work end to end, from product design to launch. Tell us about your scenario and we can start there.

Shenzhen XZS Technology Co., Ltd. · Tel 0755-82494862 · 13760424861

Contact us

← Back to News