Exactly this - I've a colleague who just fell into "solution mode" and started whiteboarding a LLM framework he could fine-tune with RAG and some 3rd party vector database. It's six months later and there's still nothing working or even a deliverable schedule.

Compare this to another colleague, with an almost exactly similar use case, who just downloaded an open-source LLM, wrote some glue code and set it loose in production. It's not pretty but it (mostly) gets the job done.

As the old addage goes "Perfect is the enemy of good".