There's basically two places where I see it as being useful -- as part of an automation pipeline where you need some translation layer from natural language to structured data or vice versa, and when the user knows it's an LLM and _wants to interact with an LLM_. I never want to see LLM generated content as an end user when I haven't specifically asked for it.
The use case our startup is working is one more that I think really works well: why are you showing me these results?
Our product basically takes any search output and uses our internal knowledge graph for our clients to "explain" to the customer "why these results for this query/question?"
We use the same knowledge graph to then extend our client's native search results with much more relevant results using the LLM to better match results to the semantic intent of the question.
It's all quite nifty and transparent; the user wouldn't know that there's an AI involved without labeling it as AI generated.
