The initial speed is tempting, but you're conflating two fundamentally different problems. Your BI platform feels heavy because your data model isn't built for agile, operational queries, not because the technology is inherently slow. A tool like Perplexity doesn't solve the data modeling problem, it just hides it behind a conversational interface.
You're right to separate deep analysis from ad-hoc insight. But for the question "what were the main causes of our P1 incidents last quarter?", the bottleneck isn't the reporting tool, it's that the data isn't structured for immediate querying. Perplexity would require you to first structure all those postmortems into a clean, consistently formatted corpus, which is the same upfront work you're trying to avoid. It just moves the labor from building a dimensional model to curating documents.
The real comparison is: would you rather maintain a semantic layer in your data warehouse, or maintain a perfect, static set of documents for an LLM to ingest? Both require rigor, but only one gives you an audit trail and deterministic results.
You're right about the setup speed, but I think you're underselling the Looker modeling lift for a reason. Those "weeks of data modeling" create a governed, version-controlled asset. The two-day Perplexity ingest just creates a pile of tokens.
I've seen exactly this pattern in a payments compliance audit. The "curated internal wikis" ingested for quick questions became the source of a regulatory finding because the chatbot surfaced an outdated policy draft as the current standard. There was no lineage, no certification, just a fast, confident answer from a document that shouldn't have been in the index anymore.
The real comparison isn't deployment effort, it's operational debt. Perplexity's overhead isn't the pipeline to Postgres, it's the constant curation, cleaning, and validation of the source corpus to prevent that fast answer from being dangerously wrong. That's a permanent human cost your BI platform pushes to the front through its modeling phase.