Zendesk Now Bills Only When Its AI Resolves a Ticket

Let me catch you up on something that is about to hit your renewal talks and your pricing page at the same time.

For twenty years, software sold access. You bought seats. You added seats. The vendor got paid when your headcount grew. That math is breaking. Agents now do the work a person used to do, and they do not log in like a person, buy like a person, or stay loyal like a person. The whole bill is moving from "who has access" to "what work got done." Here is where we are.

The deep cut

• Pricing tells your team what the product is. Measure actions and your agents will perform more actions, as Salesforce's own architecture guidance warns.
• A better agent can shrink its own bill. Zendesk charges per resolution, so a smarter agent means fewer billable actions, not more.
• The model is cheap now, so defend the work. GPT-6.1 Sol runs 13x cheaper than Grok 4.7 at similar scores, and a thin wrapper has nowhere to hide.

The seat was always a stand-in

A seat was never the thing you valued. It was a clean proxy. Give a person Figma or Jira and you gave them a capability, so The Product Journey puts the old equation plainly: value roughly equals users times price per user. Predictable, easy for procurement, stable revenue.

Now one employee runs an agent that does the work of several people. Headcount stays flat while software-driven work climbs. In a support tool where an agent handles 40% of cases on its own, you still have 100 seats and far more work getting done. The seat stops tracking value. The unit you charge for has to move with it.

The spectrum nobody has settled

Vendors are testing the whole range at once: seat, usage, action, task, outcome. Microsoft split the two. The flat Copilot subscription stays per user, while Copilot Cowork bills through credits priced on model use, context, tool calls, and runtime. Zendesk went furthest, charging only on verified resolutions. Zero resolution, zero charge.

The catch is that customers do not want a software bill that swings like a utility bill. Push toward usage and you get complaints about unpredictable cost, and some vendors are already handing out discounts to stop customers from walking. Vendors want pricing that tracks their cost. Customers want pricing that tracks their value. Those two numbers are rarely the same.

The success definition becomes a product spec

Before you can charge for an outcome, you have to define the outcome. In support, is a ticket resolved when the agent says so, when the customer goes quiet, or when an independent check confirms it? Each answer prices differently. Zendesk treating verification as part of billing is the tell: the definition of success is now a product decision, and product and pricing stop being separate conversations.

There is a trap hiding here. Make your agent dramatically better and, under action pricing, your customer's bill drops. You improved the product and cut the thing you charge for. That is why swapping seats for any old meter is not enough. The real question is the highest-value unit you can measure without gaming.

The model is getting cheap, so own the work

The model layer is commoditizing fast. Paweł Huryn's bug-hunt tests show GPT-6.1 Sol running 13x cheaper than Grok 4.7 at similar scores, and Opus 5.5 nearly matching a pricier rival at two-thirds the cost. When raw capability keeps halving in price, charging for tokens means selling the one thing racing toward zero.

That is why investors are backing the opposite bet. BAG Ventures says the experimental phase is over and enterprises are "dialing in heavily on the unit economics," paying for completed jobs, not open-ended chatbots. A thin wrapper on a frontier API "gets wiped out." Own the intent layer and the workflow, or the next model release eats you.

Prove the value before finance asks

Steady, business-critical demand turns consumption pricing into a line item nobody can forecast. MIT Technology Review's HPE piece frames the real move: once usage is predictable and large, find the crossover point where owning capacity beats buying it one request at a time, then keep that capacity productive. Capital alone does not create value. Adoption and governance do.

And the hidden costs are real. Ascerta, formerly Pay-i, found a customer whose agent averaged $0.40 per run but spiked to $70 hundreds of times a day. Meanwhile MIT's 2025 finding, cited by Timothé Frin, still stings: 95% of enterprise GenAI pilots showed no measurable P&L impact. If you cannot prove the work moved the business, your budget is the first thing cut.

Three questions for your team

• What is the smallest unit of value our product creates, in one sentence? If you cannot write it, you do not have a pricing problem yet, you have a product problem.
• If our agent got twice as good tomorrow, does the customer's bill go up or down? If a better product means a smaller invoice, your meter is pointed at the wrong thing.
• Can we name the business outcome our AI moved last quarter, with a number? If the answer sounds like the 95% of pilots that proved nothing, fix that before the model you rely on gets cheaper than your margin.