I've been working in "genBI" for a few years now and this has been exactly the trajectory. We started by focusing on selling shiny viz with an AI engine to produce it. But now making a fancy chart indeed costs 0.50, except it's going to be the wrong chart with the wrong data that will be read once and lost.

The hard part with all of this is building and maintaining a semantic layer, but more importantly capturing all the prompts and queries users are firing in to make these shiny HTMLs. Since making them is so cheap, what you want to do is capture that usage and convert that into improvements to the actual BI layer (the semantic definitions, the skills and knowledge base files, even the charting engine rules) so you can provide a consistent, correct source of truth for everyone at your company (or anyone relying on you for analytics, even your end users).

The data model underneath all these charts, especially a self-improving one based on processing usage data into quality evals and fine tuned models, is now the key. The charts can be replaced with HTMLs made by Claude for sure.

(Disclosure: I'm the founder of Upsolve so I think about these problems every day)