A lack of auditability, data leakage and high error rates all contributed to organizations shutting down their AI agents.