AI spend is often discussed as token price. For a consumer business, that is a weak frame. The better question is which process those tokens replace, accelerate, or compress.
If an agent loop takes over support, operations decisions, QA, research, or workflow coordination, its cost should be compared with payroll, contractors, mistakes, SLA misses, and cycle time.
That calculation does not make AI spend automatically cheap. It removes the false comparison of “expensive tokens versus free humans”. Human processes also have latency, management overhead, and error cost.
New Runtime read: agent economics are not calculated in a vacuum. They sit next to the unit economics of the process the agent actually changes.