Fine-tune vs RAG: the real economics for medical agents

RAG is the right starting point for almost every medical AI project. It is fast to build, easy to evaluate, and forgiving when the underlying knowledge base shifts.

But at production volumes — say, 50,000 agent runs a day — three things start to bite. Latency from retrieval round-trips. API costs from frontier models that re-read the same context. And a creeping inability to reason in your team's voice.

This is where a fine-tuned SLM stops being a nice-to-have and starts paying for itself. We sketch out the math for a typical PV team in this article and show why the breakeven is closer than most CFOs expect.