AI Sped Up Coding Faster Than It Sped Up Delivery
A 2025 study of 750k coding tasks found that AI assistance boosted coding speed by 26%, but delivery speed—testing, reviewing, and deploying code—only improved by 3.4%. The research suggests AI's main gains are in writing code, not in downstream processes like reviews or deployments, and that nearly 40% of time saved on coding was offset by increases in delivery-related activities.
Background
- The article challenges a common assumption in software engineering: that making code faster to write (via AI coding assistants like GitHub Copilot, Cursor, or ChatGPT) automatically makes the whole delivery pipeline faster.
- It argues that coding is only one part of software delivery — the rest includes reviewing, testing, debugging, deploying, and monitoring. If those downstream steps don't speed up proportionally, they become bottlenecks.
- The core insight: AI has accelerated code *generation* more than it has accelerated code *verification and delivery*, creating a new kind of imbalance in engineering workflows.
- This matters because teams investing heavily in AI coding tools may not see the productivity gains they expect unless they also address the slower parts of the pipeline — which are harder to automate.