AI & Automation

The Unit Economics of an AI Feature, From Prompt to Gross Margin

Build the cost-per-use model for an AI feature — every token, retry, retrieval and human review that sits between a prompt and a margin — and price it so that your best customers are not your least profitable ones.

  • Advanced
  • 14 min total
  • 15 chapters

What decision this helps you make: Whether an AI feature can be sold at your current price, at what usage ceiling, and which of the four cost levers to pull before you touch the price at all.

What this topic is

The unit economics of an AI feature is the profit on ONE use of it: the revenue you can attribute to that use, minus every variable cost that use causes. Those costs are unusual for software, because they are genuinely per-use — input and output tokens, retries, retrieval and embedding calls, guardrail passes, and any human who checks the output. A feature can be loved, well-adopted and quietly negative on every heavy account.

Why it matters

Software was priced for forty years on the assumption that serving one more customer cost roughly nothing. An AI feature breaks that assumption in the one place that matters most: the cost rises with usage, and usage rises with how much a customer likes the product. Left unmodelled, that turns your most engaged accounts into your worst accounts, and it shows up first as a gross margin that drifts down a point a quarter with no obvious cause.

Who should learn it

Anyone shipping an AI feature inside a priced product — founders, product leads, and finance people being asked to sign off on a plan whose cost of goods is now a usage curve rather than a fixed hosting bill.

What you will understand

  • The full cost stack behind a single user action, including the fan-out most models miss
  • How to compute cost per active user per month and turn it into a gross margin
  • Which four levers move that number, ranked by how much they cost you to pull
  • How to set a fair-use ceiling that protects margin without punishing the users you want

Prerequisites

Common misconception

"Inference prices keep falling, so the cost problem solves itself." Prices per token have indeed fallen fast, and they will probably keep falling. What has not fallen is the cost of an AI feature, because falling prices are spent immediately: cheaper tokens buy longer context, more retrieval, a verification pass, and an agent loop that calls the model eight times instead of once. This is the ordinary Jevons pattern — a cheaper input raises total consumption of it — and it means the per-call price is the wrong thing to watch. Watch calls per user action, and watch it monthly, because that is the number your own roadmap keeps increasing.