Outcome-Based AI Pricing Sounds Fair Until You Ask Who Defines Success

In June 2026, Salesforce made a move it avoided for decades: its customer service agent, Agentforce, announced it would charge $2 only for a "successfully and autonomously resolved customer issue." If the customer escalates to a human, leaves a negative rating, or abandons the chat midway, Salesforce charges nothing. Just three weeks earlier, the company spent $3.6 billion to acquire Fin, an AI customer service startup that charges per interaction.

During that same week, software billing researchers highlighted a striking disparity: the public price for an AI agent to resolve a single customer ticket ranges from $0.50 to $2 across vendors—a fourfold difference. That gap does not stem from model capabilities. It exists because vendors define "resolution" in completely different ways.

Enterprise software is moving away from charging by the seat toward charging by the completed task. Anyone buying AI today needs to look closely at what that actually means.

Moving from Selling Seats to Selling Work

For twenty years, the SaaS business model relied on headcount. Software value was tied directly to how many employees logged in. Vendors charged an annual fee per seat: one user, one license.

When software starts doing the work itself, that logic falls apart. According to venture capital firm Bessemer's 2026 AI pricing report, the share of SaaS companies charging purely by the seat dropped from 21% to 15% in a single year. HubSpot charges $0.50 per successful resolution for its customer service agent. Zendesk charges between $1.50 and $2 per automated resolution. Independent agent platforms like Sierra are moving in the same direction. Vendors know that the more capable the software becomes, the fewer human seats a buyer needs. If they cannot charge for people, they must charge for the work the software completes.

On paper, this looks like a win for enterprise buyers: you stop paying for unused features and only write a check when a problem is actually solved.

The Trap: The Vendor Controls the Definition of Success

The real risk lies in the fine print. Salesforce defines a billable "successful resolution" using three criteria: at least two conversational turns, no explicit negative feedback from the user, and no escalation to a human agent. If any of these conditions are missed, the interaction is free.

It sounds reasonable. But consider what happens in practice: a customer asks three questions, the AI answers, and the customer simply stops responding. Does that count as resolved? Does silence mean satisfaction? Who makes that determination?

When different vendors evaluate the same ticket, their standards diverge sharply. Some count a ticket only if the customer submits a positive rating and no human ever touches the record. Others bill as long as the AI participated in the thread. Run the same batch of support tickets through four vendor pricing models, and you will get four wildly different invoices. In almost every case, the vendor decides which interactions count.

Outcome-based pricing appears to shift performance risk to the vendor, but it actually gives the vendor control over what counts as an outcome. If a buyer does not define that term in the contract, they accept the vendor's version of completion by default. Procurement teams are already wary: an enterprise buyer survey in the first half of 2026 found that only 27% of buyers prefer pure outcome-based pricing. Most still favor consumption-based models because usage metrics are transparent and predictable.

The Contrast: Paying for Execution Versus Paying for Results

In Asian markets, most enterprise AI pricing still follows a different path: charging by token consumption or API calls. Alibaba Cloud recently commercialized its Agentic Search service, billing strictly for tokens consumed and web search queries performed. Coze (ByteDance's bot-building platform), Bailian (Alibaba Cloud's AI platform), and standard model APIs operate on the same logic.

The underlying incentives could not be more different:

• Consumption-based pricing charges for execution. You pay for every token the model burns, even if the run fails to deliver what you needed. The vendor has an incentive to maximize volume, with no commitment to the outcome.
• Outcome-based pricing absorbs the cost of failed runs. However, the vendor keeps full control over the criteria for success.

Both models present risks. You either pay for failed attempts, or you pay based on the vendor's self-serving definition of success. Neither guarantees you pay only for the exact business result you set out to achieve.

Three Terms to Clarify Before Signing an AI Contract

The shift in pricing models is still in its early stages. Unit prices are declining rapidly—HubSpot cut its per-ticket price by half within a year—so fixating on unit rates during vendor negotiations is rarely the best use of time. Instead, focus on three contractual terms that price drops will not fix:

1. Write the definition of completion into the contract. Specify the required conversation turns, clear criteria for success and failure, and rules against duplicate billing if a customer contacts support again on the same issue within a set window. Do not rely on vendor default metrics.
2. Examine bundled base fees. Many "outcome-based" proposals still include platform seat fees, data licensing costs, or mandatory implementation charges. The outcome component often covers only a fraction of the total spend.
3. Require auditable, itemized invoices. Insist on an interaction log that your team can verify line by line, rather than accepting an aggregate monthly figure.

In our work deploying enterprise agents at Starbox AI, we find that an AI agent's business value never depends on its pricing model. It depends entirely on whether its work can be inspected and verified. Breaking complex tasks into verifiable units of work gives both sides clear standards and an audit trail for what was actually accomplished. That verification right is the most critical asset buyers must protect.