Does AI help humans make better decisions? A methodological framework for experimental evaluation