Here’s a use case: you load an LLM instance with a large amount of highly specific factual data over the course of several weeks (ingestion of a large number of documents, daily KT sessions, call with screen sharing transcription, docs artifact generation), and use that LLM to generate answers to people’s requests under high time pressure (urgent project deadline) which is prohibitive to even type the message, nevermind to comb though the data which you’d have to cram in your head, then never use again. Both sides are aware it involves AI. I have used this process against a control group (similar tasks, no AI use) and the result was clearly superior.