Child welfare agencies are under pressure to determine where AI can meaningfully improve practice, and where it should not be used. The challenge is identifying problems worth solving, protecting professional judgment, understanding risk, engaging the workforce, and generating evidence that a solution works before scaling it. Generative AI could help child welfare agencies reduce administrative burden, make trusted information easier to use, and give staff more time to focus on children and families. But agencies need to identify which child welfare decisions should remain entirely human, which administrative tasks can be augmented, and what evidence agencies need before expanding AI use to strengthen child welfare practice.

Mathematica helps state child welfare agencies move from AI possibilities to defensible adoption decisions, from setting outcomes and identifying promising opportunities through testing, evaluation, and continuous improvement.