Cost sets the floor; value sets the price
AI pricing trips up founders because two numbers feel like they should be the same and are not. Your cost per action — what it actually costs you in model calls and infrastructure to deliver one unit of value — sets a floor below which you lose money. Your price should be anchored to the value the customer gets, which is usually many times the cost. The mistake is collapsing these into one and pricing just above token cost, which leaves enormous value on the table and trains customers to haggle over your inputs instead of paying for their outcomes. Know your cost per action precisely; then forget it when you talk to customers.
Choosing a pricing model
There are three workhorse models and most products blend them. Per-seat charges per user. It is simple, predictable, and familiar to buyers, and it works when value scales with how many people use the tool and usage is fairly even per person. Its weakness for AI is that a heavy user and a light user pay the same while costing you very differently.
Usage-based charges per unit of work — documents processed, messages, generations. It ties revenue to consumption, which protects margin and feels fair when usage varies wildly between customers. Its weakness is unpredictability: buyers dislike meters they cannot forecast, so cap and tier them into understandable allowances rather than exposing a raw per-token counter.
Outcome-based charges for results — a resolved ticket, a qualified lead, a finished deliverable. It aligns your incentives perfectly with the customer’s and can command the highest prices, but only works when the outcome is cleanly measurable and you trust your quality. Most successful AI products land on a hybrid: a per-seat or flat base for predictable revenue, plus usage allowances that scale with consumption, with outcome pricing reserved for cases where the result is unambiguous.
Protecting margin and selling the value
Two forces threaten AI margins: model cost volatility and runaway usage. The first is mostly benign — hosted model prices have trended down, so a healthy margin today tends to widen. The second is the real danger: one customer with a pathological usage pattern can erase the profit on dozens of others. Defend against it in code with hard spend guardrails and in plans with fair-use caps, and monitor cost per action continuously so a regression in efficiency does not silently eat your margin.
When you communicate price, translate everything into the buyer’s world. They do not think in tokens, latency, or model tiers; they think in jobs done, hours saved, and results delivered. Build named tiers tied to their scale — team size, volume, or outcomes — and hide the mechanics entirely. A buyer should grasp what they get and roughly what it costs without ever learning what a token is. Pricing in their language, anchored to their value, with your costs as a private floor, is what lets an AI product capture the value it creates instead of giving most of it away.