A recent survey shows half of AI agents deployed by enterprises fail despite passing evaluations, revealing a significant trust gap.