Windsurf's AI-code metric inflates productivity claims — showing 98% AI-written code that doesn't match how developers actually work

Windsurf (a VSCode-based IDE with AI code-generation features owned by Codeium) displays a metric called 'PCW' (percent code written) on its analytics dashboard, claiming customers typically see 85–95% of their code generated by AI. One user testing the tool saw 98%, implying the AI wrote 49 times more code than the developer wrote manually.

The metric counts bytes of code persisted in commits, attributing them to either the developer or Windsurf's AI. However, the author's testing revealed a major flaw: the tool resets its history when the editor restarts, so code deleted after a restart doesn't subtract from the AI's byte count. This means the metric doesn't actually track what happened across multiple work sessions—it only measures what's visible within a single editor session.

Enterprise customers using Windsurf to justify AI tool spending are seeing misleading numbers. A 98% PCW figure might convince executives that AI is replacing half their developers' work, when the actual metric is skewed by session resets and doesn't account for code developers write, test, then delete before committing. Teams evaluating whether these tools save time or just create unreviewed code are getting false confidence in the data.