Why AI Founders Now Track Cost Per Task, Not Just Gross Margin

A founder's dashboard metric — cost per completed task — is prompting a rethink of how AI-native products measure unit economics, as traditional gross margin figures no longer tell the full story. Unlike conventional SaaS, where serving an extra user costs nearly nothing, AI products average around 52% gross margin because every response requires a paid model call. Agentic AI compounds this further, since a single user-facing task can trigger five to ten model calls internally, making costs scale closer to exponentially than linearly with usage. Token prices have fallen sharply — Anthropic's Opus dropped from $15 to $5 per million input tokens — but the savings are offset when each task multiplies that lower price across multiple calls. Engineering techniques like semantic and prompt caching can cut inference bills by up to 73%, making cost-per-task tracking a more predictive signal than gross margin alone.