ai-unit-economicslisted

# AI unit economics — every feature use costs money Classic SaaS has near-zero marginal cost: once the code is written, the millionth user costs almost nothing to serve. **AI products break this assumption.** Every query, every feature use, every agent loop spends real tokens that scale linearly (or worse) with usage. Margin is no longer managed per-account at signup — it is managed **per-query, every single time the feature fires.** If a PM cannot state the fully-loaded cost of one AI action and the revenue that covers it, the product does not have a business model — it has a demo with a credit card attached to an LLM API. This skill is the AI extension of `cost-model` (general initiative cost) and `pricing-models` (how you charge). It models the thing those two do not: **per-query variable cost and its margin consequences.** ## 1. The core equation ``` gross margin per query = revenue per query − fully-loaded cost per query ``` A product is viable only if this is positive **across the realistic usage distribution**, not just at the median. Power users sit in the right tail and can be margin-negative while the average looks fine. ## 2. Cost-per-action math (the part teams get wrong) The naive estimate uses only raw token price: ``` naive cost = (input_tokens × input_price) + (output_tokens × output_price) ``` The honest estimate adds **orchestration overhead — typically 30–60%** on top — because a real AI action is rarely one clean call: ``` fully-loaded cost per