OpenAI Tests Charging Only When Its AI Finishes the Job

According to a report by the US trade outlet The Information, OpenAI is offering select enterprise customers a new pricing model: they pay only once an AI agent has successfully completed a task, such as handling a support request. An OpenAI spokesperson declined to comment. The move sounds like a customer-friendly offer, but it touches one of the industry’s hardest questions: when exactly has an AI done its job, and who gets to book the result?

Key takeaways

• According to The Information, OpenAI is testing outcome-based billing with select enterprise customers, charging only when a task is completed successfully.
• The approach breaks with the usual model of charging per user seat or per unit of compute consumed.
• The core problem is attribution: without clear rules, it is hard to prove whether a result is due to the AI or to other factors.
• Salesforce, Adobe, and several startups such as Sierra and Cognition are testing similar models, some with credits if the promised performance falls short.
• The driver is cost pressure: AI agents take on work that used to require several employees, so per-seat pricing no longer fits.

What OpenAI is actually testing

Until now, OpenAI has billed its enterprise customers mainly by tokens consumed, meaning the amount of text processed, or through fixed per-user subscription fees. In the new test, the customer pays for a result instead. The example given is customer service: an agent handles a support request from start to finish, and only the completed case triggers a payment. OpenAI has not disclosed how large the amounts are or by what criteria it counts a case as successful.

The timing is not a coincidence. With Astra, OpenAI is preparing an agent-focused flagship model designed to work through multi-step tasks on its own. A pricing model that charges for exactly those completed tasks fits that product promise. At the same time, the entire industry is under pressure to turn expensive models into sustainable revenue, which is why paths such as advertising inside ChatGPT are being tested alongside new billing schemes.

The attribution problem: when does a task count as done?

The weak spot of the model is assigning credit for success. If a sales team closes more deals, that may be down to an AI agent, but it may just as well be a marketing campaign, seasonal swings, or a stronger economy. Without contractually defined success criteria, every invoice becomes a matter of interpretation. Both sides also have an incentive to move the goalposts their way.

On top of that comes a technical uncertainty: AI agents often judge their own progress too optimistically. Studies show they rate their results markedly better than they are, and that they have no reliable sense of how much effort a task takes. A system that declares itself finished makes a poor basis for an invoice. In practice, a human review step or an after-the-fact check will likely remain necessary, which eats up part of the promised efficiency.

The whole industry is looking for a new price tag

OpenAI is not alone in the attempt. The startup Sierra charges only for the cases its support agent Fin resolves fully automatically. The developer-tools company Cognition promises credits of up to ten million dollars if the promised performance does not materialize. Salesforce negotiates individual contracts for its Agentforce agent product, tied to measurable revenue gains or cost savings. Adobe bills parts of its enterprise offering by completed advertising campaigns, and HubSpot and Zendesk are moving in the same direction.

The shared reason: the classic model of one price per user seat loses its logic when software takes on work that previously required several employees. A provider that keeps charging per head paradoxically earns less the better its AI does the job. Outcome-based billing is meant to resolve that contradiction, but it shifts the risk of inconsistent AI performance from the customer to the provider. For OpenAI, whose models are expensive to run, that can squeeze the margin when an agent often needs several attempts.

The bigger picture

Outcome-based pricing is more than a discount campaign. It is an attempt to make the economic value of AI work directly measurable, rather than estimating it through compute time or user counts. Whether that works depends less on the technology than on the contract work: on precise, verifiable definitions of success that both sides accept. If the model catches on, it shifts the burden of proof toward providers and turns agent quality directly into a business risk. For now it remains a test with a handful of customers, and its results will show whether success can be written into a contract cleanly enough that no one argues at the end of the month.