OpenAI has spent the past few months offering some large customers a way to pay only after its AI actually finishes a task — a resolved support ticket rather than a seat licence — according to The Information. That puts the company on the same commercial footing as the startups that pushed outcome pricing first: Sierra, which bills only for tasks its AI completes on its own, and Fin, which runs the same scheme and is being bought by Salesforce for $3.6 billion. The change is less a pricing experiment than a statement about who now carries the risk that enterprise AI does not work.
The structure is spreading outward from the youngest vendors. Cognition goes furthest, promising enterprise customers credits of up to $10 million if its software fails to deliver results worth at least what they paid for it. Adobe, HubSpot and Zendesk — older software companies with installed bases and subscription revenue to protect — are moving in the same direction.
The Information's explanation for the timing is cost. Running these products is expensive, and AI has not yet accelerated revenue growth at software vendors. At the same time, customers' IT budgets are under pressure from competing tools, including Anthropic's Claude, which recently shifted further toward usage-based billing. Vendors are being squeezed from both ends: their own unit economics push them to charge for consumption, and their buyers refuse to pay for capacity they cannot measure.
Salesforce has the most developed version of the idea. It already lets companies choose their own terms for Agentforce, signing individual contracts tied to how much the system raises revenue or cuts costs. Speaking at an investor conference, chief executive Marc Benioff said customers want to buy AI and pay for it in different ways, and that the point of the model is not simply to count closed calls and bill two dollars for each one. Companies want payment tied to financial effect: if the software produced $20 or $40 of additional revenue, then the customer pays two dollars.
Benioff's arithmetic is exactly where the model becomes hard. The more common outcome pricing gets, the harder it is to establish what produced the saving or the sale — the software, or the customer's own actions. Stripe has already published guidance on the question, noting that an outcome may trace to product changes, marketing campaigns or seasonality. Without clear rules for attributing a result, customers can dispute a vendor's claim that its AI was the cause.
Adobe's version shows how cautiously the incumbents are entering. Part of its new consolidated CX Enterprise AI suite will be billed according to value created, with one metric being the number of completed advertising campaigns. Adobe named no prices. Its other tools keep subscription and usage-based pricing, and the new scheme does not extend to the existing AI features for editing photos and video.
Read together, these moves look defensive rather than confident. A vendor that believed its product visibly raised customer revenue would price on that belief and collect a subscription; a vendor that has to prove causality per task is answering a buyer who has stopped taking the claim on trust. Outcome pricing transfers the burden of proof from the customer's procurement team to the vendor's measurement stack, and it does so while the vendor is paying inference costs on every attempt, successful or not. Sierra's version is the cleanest because it is also the narrowest: a support ticket either got resolved without a human or it did not. Cognition's is the most exposed, because a $10 million credit is a claim about business value, not task completion.
What is missing from OpenAI's side is the part that would make this legible. There is no disclosed price, no named customer, no indication of how much of its enterprise business is now sold this way, and no definition of what counts as a completed task when the work is not a support ticket. The same gap sits in Adobe's announcement, which names a metric but no number attached to it. Outcome pricing is being described as a philosophy while the contracts that would make it checkable stay private.
The unresolved piece is who adjudicates. Stripe can publish attribution guidance, but guidance is not arbitration, and every dollar of outcome-based revenue is a dollar a customer can contest at renewal. Software companies spent two decades building predictable, recognisable subscription revenue. They are now selling a product whose revenue arrives only after an argument about causation — and the argument runs quarterly, forever.