Charging for an AI service when your own cost changes every month

AI services have a cost structure that traditional software does not: a meaningful variable cost per unit of work, on top of the fixed cost of building the thing. That breaks the habits borrowed from software pricing. Per-seat pricing is simple to sell and disconnects revenue from cost. A single heavy user can consume more than their seat pays for. It works when usage per person is predictable and fails when it is not. Per-use pricing tracks cost directly and is the safest for margin, but it punishes the customer for adopting the product — exactly the behaviour that makes them stay. Most implementations soften it with an included allowance and overage above it. Outcome pricing — per document processed, per lead qualified, per case resolved — is the easiest to justify to a buyer because it maps to something they already value. It requires being able to measure the outcome and being confident about the cost per outcome. Whatever the model, three things need to be in place before the first contract. Know the cost per unit of work, measured rather than estimated. Put a usage cap or a rate limit in the product, because the first runaway loop is not a question of if. And keep the right to change prices with notice in the terms, since the underlying costs have moved substantially and in both directions. A margin that only works at the current provider price is a margin with a dependency, not a business model. This article is general information, not legal, tax or financial advice. Rules change and every deal is different — check your own case with a licensed professional. Alberto Zaltzberg — Adonait · adonait.com