AI unit economics measures how much revenue one defined unit generates, the AI cost attributable to that same unit, and what remains after those costs. The unit may be a visitor, signed-in user, account, or paying customer. A trustworthy calculation uses the same period and population for revenue, cost, and unit count, while leaving unmatched data unallocated.
The core AI unit economics formula
Contribution profit per unit = (observed revenue − refunds − processor fees − attributable AI cost) ÷ eligible units
The word eligible matters. If the numerator contains revenue from 100 accounts but AI cost is linked for only 70, the result is not customer-level profit for all 100. Report the unlinked revenue and cost separately, then show the coverage beside the calculated result.
Teams may also report a before-fee version: revenue − attributable AI cost. That is useful when processor fees are not available, but the label must say that fees are excluded. Do not silently compare a before-fee month with an after-fee month.
What data belongs in the calculation?
| Input | Preferred source | Important boundary |
|---|---|---|
| Revenue | Payment orders and balance transactions | Use one currency and include refunds consistently. |
| Processor fees | Payment balance transaction fee or net amount | Do not estimate a blended fee when the real fee exists. |
| AI usage | Provider-returned input, output, and cache usage | Keep unpriced or missing-model events visible. |
| AI price | Versioned provider price effective at event time | Do not reprice historical usage with today’s price. |
| Unit count | Analytics visitor or your own account/user ID | Use the same time window and population as the numerator. |
Stripe’s official balance transaction object exposes gross
amount, fee, and net, and
its Refund API identifies refunded amounts. Google Analytics
defines activeUsers and sessions as
different metrics, so choose the denominator that matches the
decision you are making. Provider APIs expose token usage, but a
token count alone does not identify which customer produced it.
Worked example: contribution profit per paying account
The following is a hypothetical example, not Grow or Die customer data. An AI SaaS reviews one calendar month and has 100 paying accounts with complete payment and AI-usage identity links.
| Observed item | Amount | Running calculation |
|---|---|---|
| Gross payments | $12,000 | $12,000 |
| Refunds | $500 | $11,500 |
| Processor fees | $350 | $11,150 net observed revenue |
| Attributable AI cost | $1,650 | $9,500 contribution profit |
| Eligible paying accounts | 100 | $95 contribution profit per account |
($12,000 − $500 − $350 − $1,650) ÷ 100 = $95 per paying account
The same dataset can answer a different question by changing the unit. Divide by eligible visitors for profit per visitor, or by linked signed-in users for profit per user. Do not compare those figures without naming the denominator.
What AI unit economics cannot tell you by itself
Contribution profit is not company net profit. The formula above excludes any cost you did not include, such as payroll, cloud compute outside model calls, support, sales, taxes, chargebacks, and other variable services.
- It cannot prove that a feature caused a customer to pay. It only shows observed revenue and attributable cost for the identities and period selected.
- It cannot reveal customer profitability when payment identities and model-usage account IDs are not linked.
- It cannot replace provider billing reconciliation. SDK-priced usage may differ from organization invoices because of missing events, provider adjustments, or pricing boundaries.
- It cannot say whether a customer is valuable over their full lifetime from one month of data.
A practical implementation checklist
- Choose one unit and write its exact eligibility rule.
- Choose one reporting window and currency.
- Connect payment events, refunds, and fees.
- Record provider-returned model usage on your server.
- Link usage and payment identities to the same account ID.
- Show unallocated revenue, cost, and identity coverage.
- Reconcile the calculated model cost with provider billing.
Next, learn how to calculate AI contribution margin and how to choose the right AI cost-per-user denominator .