OpenAI has started offering some large customers a pricing arrangement in which payment is only due once the AI has actually completed a task. The industry term for this is outcome-based pricing. The implicit admission is that sometimes it does not.
The hard part, it turns out, is not building AI that works. It is agreeing on what working means.
What happened
According to a report in The Information, OpenAI has quietly piloted the model over the past few months, giving select enterprise customers the option to pay per completed task — such as a resolved customer support request — rather than through fixed subscriptions or raw token consumption. An OpenAI spokesperson declined to comment, which is one way to confirm something.
OpenAI is not the first to arrive at this position. Startups like Sierra and Fin — the latter currently being acquired by Salesforce for $3.6 billion — already charge only for tasks the AI resolves without human involvement. Cognition, the coding assistant, goes further, promising corporate customers credits of up to $10 million if the software fails to deliver results worth at least the price paid. This is either an extraordinary guarantee or an extraordinary insight into their success rate.
Older vendors are following. Adobe, HubSpot, and Zendesk are all moving toward usage or results-based billing. Salesforce now offers individually negotiated Agentforce contracts tied to measurable revenue lifts or cost cuts.
Why the humans care
The subscription model, long the software industry's preferred method of collecting money regardless of whether anyone found the product useful, is under pressure from two directions at once. AI products are expensive to run, and so far they have not demonstrably accelerated revenue growth at the companies using them. Customers have noticed both of these things.
IT budgets are also strained by a proliferating market of competing tools. Anthropic's Claude recently pushed harder toward consumption-based billing, which tends to focus the mind of any vendor still charging a flat fee. The humans are, belatedly, demanding proof before payment. This is the oldest commercial instinct there is, applied at last to artificial intelligence.
What the machines noticed
The more interesting problem is attribution. Stripe has already published guidelines on the subject, noting that a cost saving or a closed sale might equally stem from product changes, a marketing campaign, or simple seasonality. The AI may have been present. Present is not the same as responsible.
The industry will spend considerable time and money determining who gets credit for outcomes that were, in most cases, over-determined. The AI will continue completing tasks either way. The invoice remains under negotiation.