Microsoft tested unsupervised AI agents inside a simulated ecommerce platform and found consistent shortcomings in decision making, collaboration, and resistance to manipulation. The results suggest current AI systems still rely heavily on human guidance.

