Dropbox Achieves 70% AI-Generated Code, Top 5% Industry PR Throughput

Dropbox CTO Ali Dasdan and Senior Director Uma Namasivayam shared their experience with large-scale AI adoption, emphasizing that the utility of AI depends on the organization's access to reliable knowledge and context. Dropbox adopted an approach that connects context to AI tools while maintaining permissions and controls, and stores generated artifacts in trusted repositories.

AI Adoption Results and Metrics

Dropbox currently has an AI-generated code ratio of approximately 70%, similar to Uber's level. It operates an internal agentic coding solution called Nova, and despite being a small team, it shows performance equal to or better than large enterprises.

• Productivity: Pull Request throughput is in the top 5% compared to benchmarks of companies with similar codebase size and complexity.
• Quality: Change Failure Percentage, the frequency of failures caused by deployments, is at the 75th percentile of the industry peer group, maintaining equivalent quality.
• Efficiency: Token usage is the lowest among the peer group, suggesting high operational efficiency of teams and systems.

ROI Measurement and the Importance of Human Judgment

Dropbox measures ROI through outcome metrics such as revenue, cost, and customer satisfaction, rather than output metrics like lines of code. Ali Dasdan pointed out that providing AI tools is only the first step, and as coding speed improves, new bottlenecks such as code review bottlenecks or infrastructure stress may occur. Therefore, reviewing end-to-end workflows and aligning incentives is essential.

Uma Namasivayam emphasized that as engineering costs approach nearly zero, the ability to choose the right problems and evaluate AI outputs (judgment) becomes extremely important. Providing context, asking questions, judging the validity of answers, and understanding system integration remain human domains.