Has anyone else found that "AI-native" audience tools are mostly LLMs making decisions they shouldn't be making?
Spent the last few months building an audience analysis system, and the biggest design question wasn't "where do I add AI", it was "where do I NOT let AI make decisions."
Ended up with a hard rule: LLMs handle interpretation and narrative generation, but never matching, scoring, or recommendation. Rules and deterministic algorithms handle those, with the LLM proposing inputs to deterministic decision rules rather than making the decisions itself.
Curious if anyone else has hit the same wall — where AI tools that "just work" on demos fall apart when you actually need auditable output an analyst can defend to a client. Or has anyone found AI-native approaches that genuinely hold up?
(For context: an 8-minute walkthrough video is at https://www.loom.com/share/278e4db305714400be0941e23e7b9b6d and the system is at https://mk-intel-delta.vercel.app/ if anyone wants to poke at the actual implementation. Happy to discuss the engineering decisions.)