A key performance indicator (KPI) measures progress toward a business goal. For a production ML system, business KPIs answer a question that model-quality metrics alone cannot: is the system delivering value? Revenue or conversion rate may be useful KPIs when a model supports user acquisition. section 14.4.4
When monitoring an ML system,
- Track both absolute and relative KPI values to get a fuller view of its effect. section 14.4.4
- Choose KPIs with data scientists and business stakeholders so monitoring reflects business needs. section 14.4.4
- Use a proxy or interpretable check when a KPI is difficult to measure or the system’s effect is hard to isolate. section 14.4.4
A falling KPI is a reason to investigate, not proof that the model caused the decline. Check the model, data, and software for the cause. Business KPIs sit above those layers in monitoring because they show whether the whole system meets its goals, while lower-level checks help explain what went wrong. section 14.4.4 section 14.6





