AI coding tools compress the cheapest part of shipping software while leaving the expensive parts exactly where they were, so teams feel faster and ship at the same pace.
The Details:
Writing code was roughly a fifth of the real shipping pipeline. Review, integration, testing, and deployment judgment made up the rest. AI shrinks the fifth to almost nothing, which does not touch the other four fifths. It just makes them visible for the first time, because they were always hiding behind the time it took to write the code in front of them.
Review, approval, and ownership are not friction to be automated away. Someone still has to understand a change, trust it enough to sign off, and own the fallout if it breaks. Treating those steps as waste produces fast-moving code paired with slow, shaky confidence, which is a worse outcome than the original delay.
The single number worth tracking is the time between a change being done and it reaching production, not lines written or PRs opened per day. That number exposes exactly where a pipeline is stuck: oversized pull requests, a test suite that gives slow or dishonest feedback, or deployment gates guarding against risk the team already solved.
Capping pull request size, even when the model could hand over far more, cuts review time to hours instead of days, shrinks blast radius when something breaks, and stops engineers from batching work to justify a heavier review. AI still helps here, summarizing intent and flagging risky diffs, but only if reviewers still hold the whole change in their head rather than approving a summary.
Bottom Line: The speed AI gives you is real, but it shows up at the moment code becomes shippable, not the moment it gets typed.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
