How something is priced reveals how defensible it is. In AI, pricing is in the middle of a quiet but consequential migration — climbing from the raw cost of computation toward the value of outcomes. Where any given AI business sits on that ladder is one of the best available signals of how durable its economics are.

The pricing ladder

At the bottom of the ladder is token and compute pricing — paying per unit of model output or per hour of hardware. This is how foundation-model access is sold, and it's a commodity game by nature. Multiple providers offer comparable capability, the unit being sold is fungible, and the long-run direction of price is down, sometimes dramatically, as models get more efficient and competition intensifies. Selling at this layer means living on volume and thin margins, perpetually undercut. It's a perfectly real business — at hyperscaler scale — but it's a brutal one.

In the middle sits per-seat and usage-based pricing, the familiar SaaS model adapted for AI: pay per user per month, or per workflow run, or per document processed. This is a step up in defensibility because it's tied to a product experience and a set of workflows rather than to a raw commodity. But it carries a peculiar tension unique to AI: the vendor's costs scale with usage (more model calls cost more), while seat pricing doesn't, so heavy users can be unprofitable and pure usage pricing can punish the customer for adopting the product more. Getting the blend right is genuinely hard, and a lot of AI products are quietly losing money on their most engaged users.

At the top is outcome-based pricing — charging for a result rather than for access or activity. Pay per resolved support ticket, per qualified lead, per successfully completed task, per percentage point of cost reduced. This is the most defensible position on the ladder, and also the hardest to reach.

Why the top of the ladder is defensible

Outcome-based pricing is defensible for a structural reason: it aligns the vendor with a result the customer is buying anyway and can't easily source elsewhere. When you sell tokens, you're competing on price against everyone selling tokens. When you sell a resolved ticket or a closed deal, you're competing against the customer's alternative ways of getting that outcome — which are usually slower, more expensive, or impossible to staff. The comparison shifts from "is your model cheaper" to "is this result worth what you're charging," and the answer is often yes by a wide margin.

It also creates the right alignment. Under outcome pricing, the vendor only wins when the customer wins, which builds trust and makes the relationship sticky. And it captures value in proportion to impact rather than to activity, so the vendor isn't penalised for becoming more efficient — falling model costs become margin rather than a price war.

Why most vendors can't get there yet

If outcome pricing is so much better, why isn't everyone doing it? Because it's hard to execute. You have to be able to measure the outcome cleanly, attribute it to your system rather than to everything else, and absorb the risk that the outcome doesn't materialise. It requires confidence in your own reliability, instrumentation to prove impact, and a willingness to tie your revenue to results you don't fully control.

That difficulty is exactly why it's defensible. The vendors that can credibly sell outcomes have, by definition, solved the measurement and reliability problems that most can't — and those solutions are themselves a moat.

Reading the ladder

For anyone evaluating an AI business — as a buyer, a builder, or an investor — the pricing model is a fast read on durability. A company selling raw tokens is in a commodity knife-fight. A company selling seats and usage has a product but a margin tension to manage. A company credibly selling outcomes has done the hardest work and earned the most defensible position on the ladder. Pricing isn't just how the money comes in — it's a map of where the power sits.