An agentic system processes ~10,000 queries per day across four tools (web_search, database_lookup, document_analyzer, report_generator). Costs are surprising — sometimes a query costs $0.02, sometimes $2.50. Monthly bills swing wildly. The CFO wants to know what's driving the cost variance, but no one on the team has visibility.
What's the right addition?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Cost visibility, like accuracy visibility, has to be stratified to be useful. A single average tells you nothing about which queries are expensive and why. Stratifying by query type, tool pattern, and outcome surfaces the actionable picture: "research-synthesis queries with 8+ tool calls average $1.80 each and represent 12% of volume but 70% of cost" tells you exactly where to focus. Stratification reveals concentration risk that aggregate metrics hide — the same lesson applied to cost spending instead of accuracy or errors.
A single average flattens the variance the CFO is asking about. "Average cost is $0.60" tells you nothing useful when individual queries range from $0.02 to $2.50.
A $0.50 cap that rejects expensive queries means the most complex (and often most valuable) queries silently fail — without you knowing what the system is being asked to do that would cost that much. Caps have a place, but not before diagnosis.
Smaller models on simple lookups make sense; smaller models on research synthesis often produce worse results that cost more in retries and rework. Uniform downgrades can increase total cost while reducing quality.