Token costs are a measure of consumption. They tell you what you spent but not what you got. An AI feature that costs $0.02 per API call and generates $2.00 of customer value has better economics than one that costs $0.005 per call and generates $0.10 of value — but the invoice does not show you this. The invoice shows you two numbers that look like they have the wrong relationship.
The organizations that manage AI economics well have built a layer of analysis between the token cost and the business outcome. They call it different things — unit economics, cost per outcome, AI ROI framework — but the logic is the same: translate token spend into the business metric that actually matters.
Why token costs mislead
Token costs create two specific distortions in decision-making.
The first is the cost minimization trap. When token cost is the primary metric, teams optimize for lower costs per call without considering whether lower costs also mean lower quality. Switching from a large model to a small model reduces token costs but may reduce output quality in ways that harm customer outcomes. If you measure only the cost reduction and not the quality change, you declare victory when you have actually made a worse product decision.
The second is the false equivalence problem. Different AI features with similar token costs may have wildly different business values. A customer support deflection feature that costs $0.03 per interaction and deflects a support ticket worth $15 to handle manually has different economics than a content generation feature that costs $0.03 per interaction and produces content the marketing team would have written anyway. The invoice treats them identically. Unit economics does not.
Building a cost-per-outcome model
A cost-per-outcome model requires three inputs: the token cost per unit of AI activity, the business outcome per unit of AI activity, and the connection between them.
The token cost is the easiest part. Pull the API billing data, divide by the number of API calls or interactions, and you have average cost per call.
The business outcome requires definition. For each AI feature, ask: what is the one business metric this feature is designed to move? For a support deflection feature, it is deflection rate. For a sales AI tool, it may be meeting booking rate or deal velocity. For a developer productivity tool, it may be PR throughput or time-to-merge. The outcome metric should exist in your current reporting. If it doesn't, that is a signal that the feature was not designed with measurement in mind, which is a problem worth addressing separately.
The connection is the model: cost per outcome = (total AI spend on this feature) / (units of the outcome generated). A support deflection feature that costs $3,000 per month and deflects 2,000 tickets costs $1.50 per deflection. If the cost of a handled ticket is $15, the net value per deflection is $13.50 and the ROI is 9x. If the deflection rate changes — say, a model update reduces accuracy and tickets start escalating — the cost per successful deflection rises and the ROI falls, even if the raw token cost stays flat.
The three unit metrics worth tracking
For most enterprise AI programs, three unit metrics cover the majority of economically meaningful use cases.
Cost per task completed applies to automation use cases where AI is replacing a human step in a workflow. Document processing, data classification, report generation. The benchmark is the cost of the human alternative — not just the fully-loaded labor cost, but the realistic marginal cost of adding one more unit of that task given current capacity.
Cost per assisted interaction applies to AI features that augment human work rather than replacing it — a sales rep using an AI briefing tool, a developer using an AI code tool, a support agent using an AI response suggestion. The unit of outcome is an interaction where the AI contributed, and the economics are measured against time saved or quality improved relative to the unassisted baseline.
Cost per customer outcome applies to AI-powered product features where the outcome is a customer action — a purchase, a retention event, an activation step. This is the most direct connection between AI spend and revenue, and it is also the hardest to measure cleanly because the AI contribution has to be isolated from other factors affecting the customer outcome.
How to use unit economics in practice
Build a simple table: one row per significant AI feature or use case, columns for monthly token spend, monthly units of outcome, cost per outcome, and the value of that outcome in business terms. Review it monthly alongside the raw cost data.
The table will immediately surface features with poor economics — high token costs, low outcome volumes, or outcomes that are hard to value. It will also surface features with exceptional economics that are being under-resourced. The point is not to cut the expensive features; it is to ensure that resource allocation follows economic performance, not the reverse.
Token costs are inputs. Outcomes are outputs. Managing AI economics means managing the relationship between them, not just minimizing one side of the equation.