Cost Per Task
Cost per task is the all-in price of one completed unit of AI work — model tokens, tool calls, retries, infrastructure — used to compare AI workforce ROI against human equivalents.
Token-cost dashboards mislead. The number that matters to a CFO is what one done thing costs: one outbound email sent, one ticket resolved, one invoice categorized. Cost per task includes failed attempts, retries, the cheaper model that escalated to the expensive one, the embedding refreshes, the tool API calls. Most AI employee economics only become legible once you measure cost per task. A model that's 30% cheaper per token but causes 2x retries is more expensive per task. The right comparison isn't 'AI vs. humans by cost per hour'; it's 'AI vs. humans by cost per outcome'.
Example
Acme's AI SDR averages $0.18 per personalized cold email sent (model + enrichment APIs + retries). A human SDR sending the same email costs ~$3.50 loaded. The AI wins on cost per task by ~20x at the volume Acme runs.
How OpenLabor uses it
OpenLabor surfaces cost per task per mission, not just token spend, so you can compare to the human baseline.
How is cost per task different from token cost?
Token cost is one input. Cost per task includes retries, failed runs, multi-model escalation, and tool API spend.
Should I optimize cost per task aggressively from day one?
No. Optimize quality first. A cheap-per-task system that produces unreliable output costs more in oversight than it saves in compute.
Related: token-cost, headcount-equivalent, llm.
AI Labor Glossary