AI unit economics is the practice of measuring what a unit of business value costs to produce with AI, rather than measuring how many tokens you burned to produce it. Token counts and monthly bills tell you how much you are spending. They tell you nothing about whether that spending is worth it. The number that matters to a CFO is cost per outcome: cost per resolved support ticket, per qualified lead, per shipped code change, per closed contract.

This shift in denominator is the entire discipline. A team can cut its token usage in half and still be running an economically broken product, if the outcomes fell faster than the tokens did. A different team can triple its token spend and be printing money, if each additional dollar returns three. You cannot tell these two apart from a usage dashboard. You can only tell them apart with unit economics.

Why tokens are the wrong unit

Tokens are an input, not a result. Optimizing for tokens is like a nineteenth century railroad measuring success by coal burned rather than freight delivered. The coal is a cost. The freight is the business. A railroad that obsesses over coal while ignoring tonnage will make locally sensible cuts that destroy the actual enterprise.

AI has the same trap. The frontier model that costs more per token may resolve a customer issue in one pass, while the cheaper model needs three attempts and a human escalation. Measured per token, the cheap model wins. Measured per resolved ticket, the expensive model may be far cheaper. Only the outcome view exposes this.

Building the outcome denominator

The hard part of unit economics is not the cost side. Allocated cost is knowable. The hard part is defining and counting the outcome. A workable approach:

  • Pick outcomes that the business already values and already counts, so the metric connects to numbers leadership trusts.
  • Attribute AI cost to each outcome through the same tagging that drives allocation, so cost and result share a key.
  • Include the full cost of an outcome, not just the winning call. Retries, failed attempts, and human escalation all belong in the denominator's cost.
  • Track the trend, not just the level. A cost per outcome that is falling as volume grows is the signal of a healthy AI product.

What unit economics changes

Once a team knows its cost per outcome, model selection stops being an engineering preference and becomes an economic decision. You keep the expensive model exactly where its higher success rate lowers total cost per outcome, and you route everything else to something cheaper. That is a defensible allocation of spend, and it is one you can put in front of a board.

Unit economics also reframes the AI investment conversation. A cost per outcome that sits below the value of the outcome is a machine you should feed more capital, not less. A cost per outcome that sits above the value is a product to fix or retire, regardless of how impressive the demo was. This is the same logic that separates a good acquisition from a bad one, applied to AI at the level of a single transaction.

From spend tracking to advantage

Most AI cost programs stop at spend tracking, which answers only how much. Unit economics answers whether it is worth it, and that is the question that turns AI from a cost center into a competitive advantage. The companies that will win with AI are not the ones that spend the least. They are the ones that know, transaction by transaction, that each dollar of AI spend returns more than a dollar of value, and can prove it.