OpenAI has begun allowing certain customers to pay only when the AI performs successfully.

      OpenAI has started allowing some of its biggest customers to pay only when its AI successfully completes a task. Kevin McLaughlin and Amir Efrati reported this change to The Information, citing an example of a customer support interaction managed fully from start to finish.

      This new arrangement is limited to a few major clients and has not been publicly announced by OpenAI. TNW has not verified the report independently, and details regarding the terms, customers, and pricing remain undisclosed.

      This approach is referred to as outcome-based pricing, which is particularly appealing to finance directors. Receiving a bill only when a service is successfully executed is significantly easier to justify than one that arrives regardless of outcome.

      Token billing has created challenges in this area. One developer managing a hundred agents concurrently amassed $1.3 million in OpenAI tokens over thirty days, illustrating a broader trend where costs increase based on attempts rather than outcomes.

      The customer support sector has already adopted this pricing model, as reaching a resolution is one of the few measurable outputs of AI. Intercom, for example, charges $0.99 for each conversation that its Fin agent successfully resolves, with no charge for unresolved interactions.

      In May, Zendesk took this further by limiting its charges to what it terms Verified Resolutions, which are confirmed by an LLM evaluation within 72 hours following the conversation. Assisted escalations and contained resolutions are provided free of charge, while billable rates are approximately $1.20 to $1.50 based on committed volume.

      Salesforce has been publicly navigating similar issues in a more complicated manner. Agentforce initially charged $2 for each conversation, billing for every 24-hour session, regardless of resolution, which clients found both costly and difficult to predict.

      The introduction of Flex Credits aimed to resolve this, altering the billing structure from conversations to individual actions at around 10 cents each, starting at $500 for 100,000 credits. This represents consumption pricing rather than outcome pricing, which is significant since failed actions still incur charges.

      Buyers seem to prefer both models, with Futurum Group's May findings indicating that 43% lean towards consumption-based pricing, while 27% prefer outcome-based pricing, and fewer than 20% still opt for a pay-per-user approach.

      Keith Kirkpatrick, the firm’s research director for enterprise software, noted that “outcome-based pricing is becoming a market standard,” with vendors offering only seat-based pricing often being dismissed before evaluations commence.

      For OpenAI, this marks a shift in strategy rather than the introduction of a new product. The company has traditionally sold capacity per token, leasing models and calls. Transitioning to a model where enterprises pay for completed tasks involves accepting the risk of incomplete work.

      That risk must be accounted for, raising the intriguing question of how it will be handled. A vendor confident in its success rate might accommodate this arrangement, while one unsure of its success would need to adjust prices accordingly to ensure profitability, explaining the tendency for per-resolution rates to cluster around a dollar rather than a cent.

      This shift also changes who bears the cost of unsuccessful responses. Under a token billing system, customers incur charges for every failed attempt, but in an outcome-based system, the vendor assumes those costs, notably transferring risk from the buyer to the developer of the model.

      The more challenging aspect lies in defining what constitutes success. While a resolution is clearly identifiable, the agentic tasks that OpenAI has been advancing, with 10 million users relying on its agents, involve multi-step processes where completion can depend on subjective judgment rather than being a straightforward entry in a database.

      Swift resolution of these definitions holds commercial significance. Enterprises unable to predict expenses often choose to run indefinite pilot programs rather than finalize agreements. Outcome pricing could alleviate objections precisely at the point in the sales cycle where negotiations typically stall.

      However, none of these matters have been finalized, and they are not publicly available. What is confirmed is that the leading model vendor has begun, discreetly and selectively, selling results instead of capacity.