How to measure AI developer productivity in 2025 | Nicole Forsgren

What It Covers

Nicole Forsgren explains how AI tools are accelerating code generation but not overall developer productivity proportionally, due to persistent bottlenecks in processes, broken builds, and unreliable systems that create friction throughout development workflows.

Key Questions Answered

• •Productivity metrics fail with AI: Lines of code becomes meaningless as a metric because AI generates verbose code easily. Companies must track which code comes from humans versus AI to measure survivability rates, quality, and avoid training bias loops in fine-tuned systems.
• •DORA and Space frameworks need adaptation: DORA's four metrics (deployment frequency, lead time, MTTR, change fail rate) still assess pipeline performance, but miss AI-era feedback loops. Space framework remains relevant because it measures satisfaction, performance, activity, communication, and efficiency without prescribing specific metrics.
• •Trust and code review dominate workflow: Developers now spend significantly more time reviewing AI-generated code than writing it. Teams must evaluate for hallucinations, reliability, and style consistency. Non-deterministic LLM outputs require validation that deterministic compilers never needed, fundamentally changing daily work structure.
• •Flow state requires new approaches: Senior engineers create effective workflows by architecting systems upfront, assigning parallel tasks to multiple AI agents with clear API conventions, then reviewing integrated results. This planning-heavy approach produces near-production code faster than traditional iterative coding methods.
• •Quick wins start with listening: Before implementing tools or automation, conduct listening tours asking developers about yesterday's friction points. Companies often discover process changes (like replacing physical approval walks with emails) that eliminate delays without engineering investment, delivering immediate productivity gains.