Software engineers using AI coding agents produce more code than ever before, but the quality is often compromised. According to data from companies in the "developer productivity insight" space, developers who use tools like Claude Code and Codex generate 80-90% of accepted code initially, but require revisions, undercutting claims of increased productivity.
Engineering managers are seeing high code acceptance rates due to AI-generated code, but the real-world acceptance rate drops significantly when engineers have to revise that code. For example, Waydev's CEO Alex Circei notes that his firm's customers see 80-90% initial code acceptance, but the actual acceptance rate drops between 10-30%.
Analytics companies like Waydev and GitClear are tracking these dynamics and finding that while AI tools increase productivity in terms of volume, they don't necessarily lead to better quality or value. Faros AI found a 861% increase in code churn under high AI adoption, while Jellyfish discovered that engineers with large token budgets produced the most pull requests but at an unsustainable cost.
These statistics suggest that companies are still figuring out how to use AI tools efficiently and effectively. Atlassian's acquisition of DX last year highlights the growing need for engineering intelligence startups to help organizations understand the return on investment on coding agents.
