From Cost per Query to Cost per Task

An ordinary AI interaction is often described as a request followed by a response. Agentic systems change the unit of analysis: completing a task may require planning, retrieval, tool calls, retries, verification and delegation to other agents.

The cost of one completed task can therefore vary widely, even when users perceive two requests as equivalent.

Measure completed work

Cost per token and cost per model call remain useful engineering measures, but product leaders should also understand total cost per successfully completed task. That measure naturally incorporates failures, retries and verification overhead.

Build economic guardrails

Useful controls include limits on repeated calls, explicit spending budgets, verification steps and escalation paths for costly or ambiguous tasks. Well-designed systems balance completion quality, reliability, latency and cost.