Beyond One-shot: AI Agents for Learning in Field Experiments
Researchers have developed a tool-augmented AI agent capable of learning from experimental data to design improved interventions. In a two-stage field experiment involving healthcare prescription messaging, the AI method autonomously extracted principles from initial data to generate new message variants. This approach significantly outperformed human-AI collaboration, with the best AI-generated message achieving a 69.8% click-through rate, a 6.5 percentage point increase over the baseline. AI
IMPACT Demonstrates AI's potential to move beyond one-shot evaluations to cumulative learning in experimental design.