Part 10 of 12 · 1 min read
Quick notes
- Unit economics start with the unit: a task. The number that matters is cost per resolved task, which is cost per task divided by the share the agent resolved.
- Cost per task has parts: model tokens, tools and APIs, infrastructure, and credits or support for failures. Which part dominates depends on the product: tokens for a research agent, failures and resolution rate for a support agent.
- Resolution rate is the biggest lever in the support model, far ahead of model price or tokens per task. Beware a metric that rewards escalations.
- Four price structures, each with a catch: per seat exposes you to heavy users, per task charges for failures, per resolved task needs a shared definition of "resolved", and a platform fee with overage protects against the tail.
- Price between a floor (cost over one minus target margin) and a ceiling (the customer's value), instrument cost per task from day one, and recompute any published cost arithmetic yourself.