Unit economics (of AI features)

Traditional SaaS has close to zero marginal cost per user. AI features don't: every request spends tokens, so cost scales with usage. Unit economics measures that relationship at three levels: cost per request, cost per user or account, and cost per outcome, such as a resolved ticket or a published article.

Each level answers a different question. Cost per request shows which features and models are expensive; look at the tail as well as the average, because a few huge requests can dominate. Cost per user shows that usage is never uniform: a small group of heavy users can drive most of the spend, and a flat-price plan can lose money on them. Cost per outcome includes the retries, regenerations, and failed attempts it takes to deliver value, which is what you're really paying for.

None of this is possible without logging tokens and cost for every call, tagged with feature, user, and account. Once you have that, the levers are clear: make it cheaper (smaller models, caching, routing, batching, leaner prompts), change the price (usage tiers or credits), or limit usage (quotas and fair-use caps). Recheck regularly, because model prices, your prompts, and customer behavior all change.