What would you trace in a production AI request?

I use standard distributed tracing concepts and adopt emerging GenAI semantic conventions carefully because their stability can change. Sampling keeps all failures and high-risk paths while controlling normal-volume cost, and trace links preserve asynchronous work. Evaluation results and later user outcomes attach to the same operation identity so quality can be diagnosed alongside latency and spend. Retention, deletion, tenant isolation, and incident access are designed before collecting content-rich telemetry.