OpenAI is experimenting with outcome-based pricing that could shift part of the financial risk of failed AI work from the customer to the AI provider.
WHAT’S HAPPENING
OpenAI has begun giving some major customers the option to pay when its AI successfully completes defined tasks, according to reporting by The Information. One example is handling a customer-support interaction from beginning to end. The arrangements are limited, and OpenAI has not publicly disclosed the customers, prices or contract terms.
The broader direction is no longer just private reporting. OpenAI CFO Sarah Friar has publicly identified outcome-based pricing as part of the company’s evolving business model. OpenAI has also begun measuring AI economics around concepts such as cost per successful task, rather than compute alone.
WHY IT MATTERS
Most AI pricing still charges for access or consumption — subscriptions, tokens, API calls or other units of usage.
That means a company can pay even when an AI system takes multiple attempts to complete a task or fails altogether.
Outcome-based pricing changes the equation.
Instead of simply asking how much AI was used, the commercial question becomes:
Did the AI actually accomplish the job?
That could move part of the financial risk of unreliable AI from the customer toward the provider.
WHO BENEFITS
Businesses could gain more predictable economics because spending can be tied more closely to completed work.
That could also make AI easier to evaluate during pilots. Instead of measuring tokens consumed or hours saved indirectly, companies can negotiate around defined business outcomes.
The model is already being used elsewhere. Intercom prices its Fin AI Agent from $0.99 per outcome and says unsuccessful attempts are not billed as outcomes.
WHO LOSES
AI providers take on more performance risk.
If a system repeatedly attempts a task but does not produce the agreed result, the provider may still absorb computing costs without generating corresponding revenue.
Customers could face another challenge: defining exactly what counts as success.
Resolving a support ticket can be measured. Research quality, analysis, sales assistance, financial work or other judgment-heavy tasks can be much harder to score objectively.
WHAT HAPPENS NEXT
The biggest fight may not be over the price of AI.
It may be over who defines a successful outcome.
As AI agents move from answering questions to performing complete workflows, pricing systems may increasingly require their own verification layer — determining whether work was completed, whether it met agreed standards and whether the customer should be charged.
If outcome pricing expands beyond select enterprise arrangements, AI economics could begin shifting from paying for intelligence consumed toward paying for work accomplished.
And that would change more than the invoice.
It would change who carries the cost when AI fails.
