Agentic workloads behave nothing like the reporting workloads Snowflake accounts were designed around. A dashboard runs the same query every morning. An agent takes a different path every run, and one prompt fragments into spend across Cortex inference, Cortex Search, and the warehouses it ultimately queries.
No single view tells you what that interaction cost or who was behind it. This session covers building attribution across the whole path from Snowflake's native surfaces, where those surfaces fall short, and how to read cost next to P50, P95 and P99 before deciding an expensive model is worth it.

Takeaways:
•⁠ ⁠Why agentic workloads break the cost assumptions behind existing Snowflake chargeback models
•⁠ ⁠How to reconstruct a full prompt-to-query attribution chain from available telemetry
•⁠ ⁠Which decisions (model selection, warehouse routing) actually move cost and latency, and how to evaluate them
•⁠ ⁠Patterns for moving agents from dev to production without losing budget or SLA control