As enterprises race to deploy agentic AI, a new fault line is emerging. The risk is no longer model accuracy, but whether organisations can actually run autonomy safely, predictably, and at scale.