Outcome-based pricing charges for the result the customer wanted, not for seats, access, or consumption. It is the pricing model AI agents make possible, because for the first time the software itself completes the work, so the work's completion can be metered.
The mechanics, using Sierra as the reference case
Sierra's version, in Bret Taylor's words: "If the AI agent resolves the case, no human intervention, there's a pre-negotiated rate for that. If we do have to escalate to a person, that's free." Reported support-context rates are around $1.50 per resolution, inside custom enterprise contracts that blend volume pricing for routine interactions with pay-per-resolution for complex ones.
The free escalation is the load-bearing detail. Every failure costs Sierra the revenue and the compute. The vendor is financially exposed to its own product's reliability, which is the cleanest possible answer to who owns the landing: the P&L does. It also creates what Taylor calls vertical alignment: cutting token cost for the same outcome is Sierra's margin problem, not the customer's bill, so efficiency work has a direct owner.
Outcomes are not usage
The two get conflated constantly and are different animals. Taylor's test: an agent that produced a tenth of the sales while using a hundredth of the tokens would be worse for the customer, not better, so tokens cannot be the price. Usage pricing, like Snowflake's credits, forces landing, because revenue only arrives on use. Outcome pricing goes one level deeper and forces landing plus quality, because revenue only arrives on success. The Sierra playbook shows how this one choice cascades into implementation, engineering, and go-to-market.
The honest boundary: where outcomes are fuzzy or gameable, use usage. Taylor concedes there is no great outcome metric for every agent type. Pick the countable workflow, put a slice of one contract on it, and let the negotiation teach you what your product's reliability actually is.