The Problem
Standard observability captures the request in, the response out, and the latency between. That's enough for traditional services. It's not enough when an LLM is in the loop.
When an AI-augmented endpoint misbehaves, the questions you need to answer are: What did the model see? Which tools did it call? Which prompt template was active? Did the cache hit? How many loops did it run? What did each loop cost? None of these are visible in standard request/response logs.
You also have a P&L problem you didn't have before. Token spend is now a real line item, and product teams will (correctly) want to know which feature is burning the budget. Vendor billing tells you total spend; it doesn't attribute it to features.
The Core Insight
Log what the model saw, not just what your service returned.
The model is now a component in your system. Treat it like any other production component: capture its inputs, outputs, intermediate state, and resource usage. The log schema below is the floor; pick anything above it.