Why Most Machine Learning Projects Fail To Reach Production
infoQ, Monday, February 2nd, 2026
This article is a summary of my talk at QCon San Francisco 2024. For the past decade, I have worked as an applied scientist and machine learning engineer across multiple domains, including social media, fintech, and productivity tools.
Most ML projects fail to reach production. Five recurring pitfalls drive failures in ML projects: choosing the wrong problem, data quality/labeling issues, the model-to-product gap, offline-online mismatch, and non-technical blockers.
Define a clear business goal before starting, and validate that it truly needs ML. Translating business goals into ML requires heavy data engineering, objective-function design, and sometimes expensive infrastructure, making late pivots costly.
Treat data as a product: prevent leakage and bias, invest in labeling and golden sets, and build evaluation pipelines early to avoid brittle releases.
Manage uncertainty with a balanced portfolio: ship low-risk/high-impact wins to justify investment, while incubating riskier bets that can be game-changing.
Encourage early collaboration and active engagement of cross-functional teams. Successful ML teams align stakeholders, scope an MVP, build end-to-end early for A/B testing, and iterate based on monitoring.