The most common AI pricing mistake isn't a bad price. It's pricing the wrong thing. Seats price the input — a person who has access. Tokens price the mechanism — the compute consumed doing the job. Outcome pricing prices the job itself. The first two are comfortable because they're models you already know. The third is the one that actually matches what you're selling, and it's the one almost nobody builds.
Seats: the comfort food that stops making sense
Seat pricing is the default for a reason. It's predictable for both sides, it's easy to model, and it scales with headcount, which your investors understand. It's also systematically wrong for most AI products, because an AI product's value doesn't scale with the people who can open it — it scales with the jobs completed.
Take a contract-review tool used by a team of twenty-five. A seat license says "twenty-five people may touch this." But the value the team gets has nothing to do with how many people have login access; it has to do with how many contracts got reviewed correctly this month. When the number of users stays flat but the number of reviews doubles, you're delivering twice the value for the same price — a great deal for the customer, and a quiet cap on your own upside.
Worse, seat pricing sets up the exact conversation you don't want: your champion has to defend licenses for people who "barely use the tool." The accountant eventually notices that five people open the product once a day, and your price morphs from "value" into "tax." You lose either way — on price, or on the utilization story.
Tokens: pricing the mechanism instead of the job
Usage pricing feels prudent: your revenue scales with your cost of compute, so you can't lose money on a heavy user. It feels modern — metered, pay-per-use, like a utility. And it's the mirror-image mistake: it prices the mechanism instead of the job, and it punishes the one behavior you're trying to earn.
Every metered product I've audited has the same pattern. Adoption spikes, then users start moderating: "is this query worth two cents?" They batch their usage, stash prompt templates, and try to move the work back to the spreadsheet. They modernize, in other words, in the dumbest possible way — and your usage graphs flatten precisely because the product works.
Token pricing also hands the customer a scary surface: an unreadable invoice line that moves up and down each month. Finance teams hate variance they can't explain, and a cost-per-invoice line that spikes in November invites a renegotiation, not a renewal.
Usage pricing is defensible only when there is truly nothing else you can measure — and even then, cap the downside: per-seat access at a floor price, metered overage on top, so nobody ever gets a surprise bill that teaches them to hate your product.
Outcomes: the only pricing that makes the customer root for you
The pricing structures I've seen actually work in applied AI share one property: the customer gets charged when value demonstrably arrives. Per contract fully reviewed. Per claim dispositioned. Per shipment disposition recorded correctly. Not "percentage of revenue," which mostly fails on trust and auditability — a flat, observable fee per outcome, with a clear definition of what counts.
Worked example, using the contract tool again. The team reviews twelve thousand contracts a year at roughly six minutes each. That's twelve hundred hours of lawyer time, which the company already values somewhere around six figures. Whether you charge per review, per seat, or per outcome, there's a number that maps to this — and the conversation is completely different depending on which one you use.
- Per seat, the buyer asks "why do I pay for five licenses nobody opens?"
- Per token, the buyer asks "what does this cost me if adoption doubles?"
- Per review, the buyer asks "what does a correct review save me, and how do we prove it?"
Only the last question is about value. The customer defends a price tied to an outcome they already believe in. That's the difference between a vendor negotiation and a budget line item.
The price is a specification of what you believe
I've come to think of pricing as the most honest document a company produces. It says what you believe your product is worth, to whom, measured in what currency of the customer's own work. When the pricing experiments live in your weekly loop along with everything else — one change at a time, watched against activation — you learn faster about your business than any feature roadmap teaches you.
The model is the toy. The pricing is the business.