At Zendesk’s Relate conference on May 22, 2026, Shashi Upadhyay, President of Product, Engineering, and AI, didn’t unveil a new model or feature. Instead, he reframed the entire economic logic of AI in business software. “Stop thinking of agents as software… start thinking of them as a unit of labor,” he told TechRadar Pro. That’s not a metaphor. It’s now the foundation of Zendesk’s new outcome-based pricing model.
Key Takeaways
- Zendesk now charges customers only when AI agents successfully resolve support interactions
- Each resolution is independently verified by a separate AI evaluation model
- This is the first large-scale commercial shift to treat AI as labor, not infrastructure
- The move pressures vendors to prove ROI, not just offer access
- Revenue risk shifts from customer to vendor—Zendesk absorbs costs for failed interactions
Outcome-Based Pricing Isn’t New—But AI Makes It Possible
The idea of paying for results, not effort, isn’t radical. Performance-based contracts have existed in consulting, legal, and marketing for decades. But applying it to software—especially enterprise SaaS—has always been impractical. You can’t bill a law firm per winning verdict if the case never goes to trial. Similarly, you can’t pay for software by the resolved ticket if you can’t verify resolution at scale.
That’s where AI changes everything. Zendesk’s new model relies on a dual-agent system: one AI handles the customer interaction, the other evaluates whether it was actually resolved. The company claims this second model is trained to filter out low-value exchanges—like automated replies or partial answers—so billing only triggers on meaningful outcomes.
And that’s the pivot: it’s no longer about how much AI you use, but what it does. For years, SaaS pricing evolved from seat-based to usage-based (think tokens or API calls). Now, for the first time at this scale, a major vendor is moving beyond usage to outcomes.
How the Verification Loop Works
Zendesk’s evaluation model doesn’t just check if a conversation ended. It analyzes sentiment, intent, and whether the user’s original issue was addressed. The company hasn’t disclosed the exact accuracy rate, but it says every billed resolution is independently validated. That means no more paying for bot loops, misrouted queries, or canned responses that don’t solve anything.
This isn’t just a billing tweak. It’s a commitment to quality. If the AI fails, Zendesk eats the cost. That’s a stark contrast to current AI pricing, where customers pay per token or per interaction regardless of result.
- Pricing applies only to AI-handled support interactions
- Human-assisted cases fall under traditional pricing
- Verification is done by a dedicated AI model, not humans
- Low-value or unresolved cases are excluded from billing
- Model is trained on historical resolution signals from Zendesk’s platform
Zendesk Is Now Eating Its Own Dogfood—And That’s Risky
Let’s be clear: this is a gamble. For the first time, Zendesk is exposing itself to direct financial risk based on AI performance. If the models fail to resolve tickets, the company still incurs compute and training costs but can’t recoup them. That’s a reversal of the usual SaaS model, where revenue is predictable and customers absorb inefficiency.
But that’s also what makes it credible. Upadhyay’s statement isn’t just marketing—it’s a revenue structure. And it signals that Zendesk believes its AI is ready for real accountability.
There’s irony here. For years, AI vendors sold their tools as productivity boosters—something that makes your team faster. Now, Zendesk is saying: our AI isn’t a tool. It’s a worker. And we’ll only charge you if it does the job. That’s not a feature. It’s a philosophical shift.
The Labor Analogy Isn’t Perfect—But It’s Getting Closer
Calling AI a “unit of labor” isn’t flawless. Workers have rights, liability, and legal status. AI doesn’t. But economically, the parallel holds. You don’t pay a freelancer for logging in—you pay for deliverables. Same with Zendesk’s AI agents now.
And this isn’t just semantics. It changes how companies budget. Instead of forecasting AI costs based on usage spikes or agent volume, they’ll forecast based on resolution volume. That’s a fundamental rethinking of cost modeling.
What’s more, it forces vendors to optimize for outcomes, not just uptime or speed. If your AI resolves fewer tickets, you make less money. That aligns incentives in a way token-based pricing never could.
This Could Reshape Enterprise AI Economics
Zendesk isn’t the first to experiment with outcome-based models. Some startups have offered pay-per-conversion or pay-per-resolution pilots. But none have the scale or installed base that Zendesk does. With tens of thousands of businesses relying on its platform, this move could set a new standard.
The implications go beyond customer service. If this works, we could see AI pricing tied to:
- Sales conversions (pay per closed deal)
- Code quality (pay per bug fixed)
- HR outcomes (pay per successful hire)
- Legal doc review (pay per clause flagged)
- Marketing ROI (pay per verified lead)
And vendors will have to prove it. No more vague promises of “efficiency gains.” You’ll need auditable, verified results.
But there’s a catch: not all outcomes are equal. A simple password reset isn’t the same as resolving a billing dispute. Zendesk hasn’t said whether it tiers pricing by complexity. If it doesn’t, customers might see uniform pricing for vastly different workloads—which could create new inefficiencies.
Still, the direction is clear. As AI becomes more capable, the market is demanding more accountability. And outcome-based pricing is the most direct way to deliver it.
Competition Will Force Copycats—Or Collapse
Let’s not pretend this is purely altruistic. Zendesk is under pressure. The enterprise AI space is getting crowded. Competitors like Salesforce, Intercom, and Freshworks are all rolling out advanced AI agents. Differentiation is harder than ever.
So Zendesk did something clever: it turned pricing into a product. While others compete on model size or speed, Zendesk is competing on trust. You don’t have to believe their AI is better—you just have to believe it works, because you’re not paying if it doesn’t.
That’s a strong value proposition. And it means others will have to follow. They can’t keep charging for AI access while customers question ROI. Gartner reported in early 2026 that 68% of enterprises were reevaluating AI spend due to unclear returns. Zendesk’s model speaks directly to that fear.
But not every vendor can afford this. Smaller AI companies with thin margins won’t be able to absorb failed interactions. Only players with scale and capital can take this risk. That could accelerate consolidation in the space—winners get bigger, losers get acquired or die.
“Stop thinking of agents as software… start thinking of them as a unit of labor.” — Shashi Upadhyay, President of Product, Engineering, and AI, Zendesk
What This Means For You
If you’re building AI tools, this changes your design priorities. You can’t just optimize for speed or accuracy anymore. You have to optimize for verifiable outcomes. That means investing in evaluation models, audit trails, and transparency—not just the primary agent. Your system needs to prove it worked, not just claim it did.
For developers, this means new APIs for outcome verification, logging resolution signals, and integrating with billing systems based on success metrics. It’s not enough to build a chatbot. You have to build a chatbot that can defend its own performance.
If you’re a buyer, this gives you use. You can demand more accountability from vendors. And if Zendesk pulls this off, others will have to match it. That could finally end the era of blank-check AI spending based on hype.
But there’s a warning: outcome-based pricing only works if outcomes are well-defined. If your business has fuzzy success criteria, this model might not fit. And if vendors start gaming the metrics—like marking every interaction as “resolved” to trigger billing—trust erodes fast.
So while this is a step forward, it’s not a magic fix. It shifts risk, but doesn’t eliminate it.
The real question isn’t whether outcome-based pricing will spread. It’s whether AI can actually deliver consistent, measurable value across complex workflows—or if we’re just repackaging old problems in new economic wrappers.
Sources: TechRadar, original report
