I recently audited an IT department for a company’s board.
Everyone there had been using AI tools for a long time. Larger projects were still late.
More and more research shows why writing code faster doesn’t automatically mean delivering projects faster.
The 2025 DORA report already described AI as an amplifier: where the process works, AI helps. Where it doesn’t, AI can make the problems bigger. The 2026 DORA report found that time saved on writing code often moves to verifying it.
An NBER study covering more than 500,000 developers shows it well. The authors estimated that autonomous agents increased the number of commits by 240%, but the number of releases by only 30%.
In the company I audited, the problem showed up even earlier. Not in writing code, testing or deployment, but where IT meets the business. Requests came from many directions, but nobody had settled who owns what or what comes first. So people worked on tasks they thought were important, not necessarily the ones that mattered most to the company.
Instead of counting commits alone, it’s worth measuring how long we wait for:
- decisions,
- clarified requirements,
- review and testing,
- and finally, deployment.