AI made coding faster. Did delivery get faster?

Most engineering teams adopted AI coding tools in the last two years. Developers say they save real time: in Atlassian's 2025 developer experience survey, 68% save more than 10 hours a week with AI. The same survey found that half of developers still lose more than 10 hours a week to friction outside coding: finding information, switching between tools, waiting.

So the time saved in the editor often leaks out somewhere else.

What the research shows

Google's 2025 research on AI-assisted software development, based on nearly 5,000 technology professionals, describes the pattern: AI raises individual output (more tasks completed, more pull requests merged), but team delivery stays flat, and instability goes up. Its central point: AI doesn't fix a team, it amplifies what's already there.

In practice that looks like:

• More and bigger pull requests, so review becomes the new bottleneck. On the average team, a change already spends more than half of its way to production in review (LinearB 2026 benchmarks, 8.1 million pull requests).
• More rework: code that was fast to write but slow to review, or that comes back as a bug.
• Releases that need fixing afterwards, because big changes are harder to review well (reviewers find fewer defects beyond about 400 lines, SmartBear study at Cisco).

Four questions to ask your own data

1. Items finished per week, before and after the tools came in. If individual output went up, did team output follow?
2. Time to first review. Google's engineering practices put the maximum at one working day. Has it grown?
3. Pull request size. Are more pull requests now over 400 changed lines?
4. Bugs and hotfixes. Is a bigger share of finished work bug fixing, or are more releases followed by a hotfix?

If output went up but these four got worse, the gain is stuck in the system, not lost. The fixes are usually small: a size limit for pull requests, a review-first habit, and shared team rules for AI-written code.

Guide values for each measure, with sources: delivery benchmarks. A free config that labels every pull request with its review time: review labels.

How we check it

In a Delivery Health Check we read these measures from the team's own tracker and pull request data (timing and size only: no code, and names are replaced by codes), add a 5-minute anonymous team survey, and talk to three people. One week, then your top 3 delivery problems, each fix written as a team agreement, and a free re-measure after 30 days.

Sources: Atlassian State of Developer Experience 2025; DORA 2025 State of AI-assisted Software Development (Google); LinearB 2026 engineering benchmarks; Google engineering practices, speed of code reviews; SmartBear, best practices for peer code review. Checked October 2026.