Why wouldn't they be? LLMs need a lot of content for training and there's multiple orders of magnitude less to train on of if you limited it to insurance-specific content, so you'd probably get a really crappy LLM. And training from scratch is really expensive anyway.

At best they'll be using fine tuned enterprise OpenAI / Anthropic models, more likely a regular model with a custom prompt.