The bill that makes this argument concrete belongs to a developer who ran a hundred agents in parallel and accumulated $1.3m of OpenAI tokens in thirty days [3]. Divide it out and that is about $43,300 a day, roughly $433 per agent per day [15], on a meter that counts attempts. OpenAI has sold capacity by the token, priced per model and per call [12], and nothing on that invoice records how many of the hundred agents finished anything.
Outcome billing needs a referee, and Zendesk supplies one: an LLM evaluation that confirms a resolution within 72 hours, with assisted escalations and contained resolutions dropped from the bill entirely [5]. In practice, the vendor's own model is grading the vendor's own work on the vendor's own clock. That holds up when the unit is a support conversation, which is one of the few AI outputs anyone can define [4]. It holds up less well for the multi-step agent work OpenAI has 10 million users doing, where completion is a matter of judgement rather than a field in a database [13].
Salesforce is the useful place to separate the thing being pitched from the thing being done. Agentforce started at $2 per 24-hour conversation whether or not anything was resolved, and customers said they could neither afford it nor forecast it [6]. Flex Credits arrived as the answer, metering individual actions at about 10 cents, starting at $500 for 100,000 credits [7]. At those terms a credit costs half a cent, so a 10-cent action burns about 20 of them [16]. It meters more finely, but the promise behind it stays the same, because a failed action still bills [7]. For contrast, if half of those old $2 sessions ended in a resolution, the effective price per resolution was $4, about four times what Intercom charges for a conversation its Fin agent actually resolves [18].
Buyers and the analysts describing buyers diverge here. Futurum Group found in May that 43% of buyers prefer consumption-based models against 27% who prefer outcome-based ones, with fewer than one in five still preferring per-user pricing [8], which puts about 70% on a meter tied to work rather than headcount [17]. Futurum's Keith Kirkpatrick nonetheless wrote that outcome-based pricing "is becoming a market standard", and that vendors offering seats alone are disqualified before the evaluation starts [9]. Both readings survive if you accept that buyers are optimising for a bill they can forecast, not for a bill that is philosophically fair. An enterprise that cannot forecast tends to run a pilot forever instead of signing [14]. And per-success prices cluster around a dollar rather than a cent [11] because the miss rate is priced into the hit, so the buyer keeps funding failures, just averaged.
For your own product or procurement, this comes down to two questions. Can you define the completed unit in a sentence a support lead and a finance analyst read the same way. Can you verify it without taking the vendor's word. If the unit is both definable and verifiable, outcome pricing is genuinely cheaper to defend than a token bill, because failed attempts move onto the vendor's side of the ledger [10]. Definable but not verifiable, and you have bought an audit obligation, so ask for the evaluation records and a dispute window of the kind Zendesk's 72 hours implies [5]. Not definable but measurable in usage, and consumption with a hard cap is the honest instrument. Neither, and staying on tokens while you instrument your own completion rate is the cheaper mistake, because whoever gets to measure completion ends up setting the price.
