Large language models are useful assistants but a poor default for business decisions, because the work of making them reliable falls on the firm that adopts them.