Algorithm tools have the potential to improve public service efficiency, but our understanding of how experts use algorithms is limited, and concerns about resulting bias are widespread. We randomize access to algorithm support for workers allocating Child Protective Services (CPS) investigations. Access to the algorithm reduced maltreatment-related hospitalizations, especially for disadvantaged groups, while reducing CPS surveillance of Black children. Child injuries fell by 29 percent. Workers improved their scrutiny of complementary information emphasized by the algorithm, and targeted investigations to children at greater risk of harm irrespective of algorithm-predicted risk. Algorithm-only counterfactuals confirm human-algorithm complementarity for both efficiency and equity.