Hybrid pricing was always a stopping point on the way to somewhere else. Charging a per-seat fee and metering AI is what you do when you don't yet trust your AI to carry the weight on its own.
But what happens when someone else takes the other side of the bet?
At Relate this year, Zendesk did just that. They dropped hybrid pricing from their AI Automation offering and went all-in on outcome-based pricing. It changes the pricing model the whole category has been using - and it’s the biggest pricing decision a major service vendor has made in years.
What they did
The old pricing model came in at roughly fifty dollars per agent, per month, plus per outcome resolution charge on top: two revenue streams, one guaranteed and one variable. Zendesk is taking the guaranteed stream off the table for autonomous work; so instead of paying for the seat and then paying again for what the AI burns through, you pay only when the AI Automation produces an outcome.
The first consequence is mathematical: the per-agent fee was the floor, the money that arrived whether or not the AI did anything useful. By removing it, Zendesk has pulled out a chunk of guaranteed revenue and replaced it with revenue that only exists if usage grows. The only way the math works is if they grow the pie, with more interactions running through the system. They're betting the volume of AI-handled work expands fast enough to more than cover the floor they kicked away; but they’re also removing barriers. Previously, committing to AI required a minimum bet of purchased AI agents - now the door is wide open.
In plain terms, they've taken a haircut. They're choosing to make less in the near term on the conviction that the long-term base gets much (much) bigger. And I respect the nerve of it. But I also think it's going to get messy before it gets clean; existing customers who already prepaid for licenses bought seats under the old model, and Zendesk is now deprecating the very model that justified the cheque. There's no graceful way to tell a customer who paid in advance that the thing they bought is no longer how you sell; I’d expect a run of back-end deals and renegotiations. But that’s not a reason to avoid a better pricing paradigm. It’s the cost that comes with doing something - anything - differently.
Zendesk went first, but they’re not the only ones eyeing the prize. Everyone in this category has been circling outcome pricing and waiting for someone else to find the landmines.
What we love
First, they've cleaned up a space that was a conceptual swamp. For a year now the industry has been acting as though "AI in service" is one single thing you can wrap a single price around, and it isn't. But there's AI Automation that does the job end to end with nobody in the loop, and there's AI that works next to a human and makes them faster. Those are different products with different economics, and pricing them the same way was always going to produce nonsense.
Zendesk split the two. Fully autonomous work, where the AI resolves the ticket itself, moves to outcome-based pricing. AI that assists a human stays on seat pricing; because in that mode you're buying an efficiency gain for a person you already employ, and a seat is the right unit for that. It's an efficiency play, it's arguably fairer, and it gives buyers a coherent story to tell their own finance teams. At Cloudset, we'd like to see the rest of the category adopt the distinction even if they don't swallow the outcome model whole.
Second - the hard part. Zendesk has taken a real swing at the blocker that has stalled outcome pricing since the idea first existed: who decides whether the outcome was achieved? If the vendor marks its own homework, no buyer with a functioning legal team is going to sign off; and if the buyer has sole say, no vendor is going to agree. The whole model has been frozen on that one question.
Zendesk answer is a separate adversarial model. It analyses the interaction and produces semi-independent evidence that an outcome happened - a different system with a different job, arguing the other side of the case before the bill is calculated. Fail that gate and the customer pays nothing. Is it perfect independence? No, and we'll get to that. But it's at least an attempt at a real answer to the hardest problem in the category, and in a market this full of hand-waving, building something that works is a win.
What you'd want to think about
Outcome pricing is a different animal from value pricing. You pay the same whether the AI handled a one-line password reset or untangled a gnarly multi-step billing dispute that would have eaten a senior agent's afternoon. Zendesk bills for the moment of resolution, ignoring how much it was actually worth. On simple tickets the buyer overpays; on hard ones they get a steal. Across a realistic ticket mix it should wash out, but buyers have to model their own distribution and usage before assuming it does, because a buyer whose queue skews simple is only subsidising the vendor.
Second - and this one is a little cheeky - Zendesk is now telling customers not to do deflection: the old reflex of pointing a user at the knowledge base and scoring the unasked question as a win. There's a legitimate experience argument here, since deflection frequently amounts to friction with a metric attached. But notice who benefits - every question the AI answers instead of deflecting is a billable resolution. "Stop deflecting" is decent advice but it just so happens to route more volume onto the meter.
Even with the adversarial verifier, Zendesk is still both poacher and gamekeeper. They built the model that does the work and the model that judges whether the work counts, and they absolutely bank the result when the judgment comes back positive - the house always wins. A truly independent arbiter, whether a third party or a standard a buyer could audit, would be the ideal scenario. Zendesk’s in-house adversary is a step closer to independence; but even if it beats self-certification, it still falls a meaningful distance short of true independence. They're taking the trust problem seriously but they haven’t solved it. Anyone signing a large commitment should ask exactly how they can inspect and overrule that verifier, and then get the answer in writing.
Where this leaves us
Zendesk has made the cleanest case in the category, autonomous work priced on outcomes and assistive work priced on seats, and backed it with a working mechanism for the verification problem. Those are the wins.
They've also accepted a near-term haircut, lined up an awkward stretch with prepaid customers, built incentives that favour more billable resolution, and handed the refereeing to a verification regime they run themselves. Those are the worries.
But their entire strategy now rests on a single question no keynote has an answer to yet: does autonomous volume grow fast enough to cover the floor Zendesk just removed? If it does, they’ve reset the category's pricing language and the rest of the field will spend at least the next year catching up. If it doesn't, this becomes the most expensive lesson in support software since the seat itself.
Either way, the bet is in the open.
.png)
