Salesforce posted $2 per resolved AI agent issue. The definition of "resolved" is where the actual cost lives.
Resolved means the agent completed the job start to finish without a human stepping in and without the customer walking away unhappy. No escalation. No abandonment. No charge if either happens.
To price an outcome you first have to make it deterministic. Salesforce ran 4.3 million inquiries through its own help portal before announcing Agentforce. They calibrated the definition against real queue data instead of guessing.
Token metering ignores what happened. It just counts compute consumed. Outcome pricing requires logging, tracing, escalation detection, and abandonment signals. The platform has to track whether a transaction actually succeeded.
ServiceNow charges by assists per action. Assist counts vary with every interaction, making bills unpredictable. JPMorgan called Action Fabric's external agent charges a tax on customers. SAP routes everything through Joule without publishing an outcome price.
Salesforce is the first at its scale tying a published price to a defined result. Watch whether that definition holds at scale. Edge cases usually force exceptions and consumption floors that quietly rebuild the token meter under the promise.
Pega made the same bet since June. Infinity 26 ships Q3 with flat per-case pricing and AI reasoning shifted to design time. That is the first real comparison point for whether outcome-based architecture actually delivers the cost predictability it promises.
[link] [comments]