Outcome-Based Pricing for AI Support: Fair Model or Expensive Trap?

When an AI support agent can actually close a ticket, a hard question follows: what should you pay for that? The old answer was simple — buy a seat per human agent. But AI does not occupy a seat, and its work does not map neatly to headcount. So vendors have reached for a new idea: outcome-based pricing, where you pay when the AI resolves an issue rather than for the capacity to try. It sounds fair. It can also get expensive in ways that are easy to miss. Here is how the models actually differ and how to tell which one is working for you.

The models on the table

There is no single way to bill AI support, and the choices shape your costs more than the sticker rate does. The main models:

• Per resolution: you pay only when the agent successfully resolves a request without escalating to a human. Intercom, for example, publicly prices its Fin agent at $0.99 per resolution and charges nothing when it cannot answer.
• Per deflection: you pay when a query is kept away from a human, whether or not the customer's problem was truly solved.
• Per conversation: a flat charge for each interaction, regardless of outcome.
• Per token / usage: billed on the compute consumed, mirroring the raw model cost underneath.
• Per seat: the legacy model, priced on the number of human agents on the platform.
• Hybrid: a platform fee plus usage, common as vendors hedge between predictability and pay-for-performance.

Why "per resolution" feels fair

The appeal of paying per resolution is alignment. As Zendesk puts it, you "pay only for customer requests that were successfully resolved by the AI agent, without any escalation to a human agent." If the bot fails, you do not pay — the vendor carries the risk of quality. That is a genuine shift away from per-seat pricing, which charged for capacity whether or not the AI did useful work. For buyers, the pitch is that cost now tracks business value delivered rather than headcount or volume. When the model works well, you scale spend with results, and a vendor that prices this way has an incentive to make the agent genuinely good rather than merely talkative.

Where the trap hides

The fairness depends entirely on how "resolution" is defined — and the vendor usually defines it. If a resolution counts whenever the customer stops replying, you may be billed for conversations that did not actually fix anything; the customer simply gave up or re-contacted through another channel. Watch for: vague definitions, resolutions counted on follow-up messages, and charges that fire even when the customer later escalates. There is also a volume effect. Per-resolution pricing rewards more resolutions, so at high ticket counts a usage model can quietly cost more than the flat seat license it replaced. The model that looks cheap in a pilot can invert at scale.

Reading your own numbers

Before signing, instrument the outcome you actually care about. Useful checks:

• Re-contact rate: how many "resolved" tickets come back within a few days. A high rate means you are paying for resolutions that were not.
• CSAT on AI-handled tickets: satisfaction, not just closure.
• Escalation handling: confirm you are not billed for cases the AI hands to a human.
• Cost per genuinely solved issue: divide total AI spend by tickets that stayed solved, not by raw "resolutions."

These reframe the vendor's metric into yours. The headline rate matters far less than what counts toward it.

Choosing a model that fits

The right structure depends on your volume and risk tolerance. Per resolution suits teams with spiky or unpredictable ticket loads who want costs to follow results and are comfortable auditing the definition. Per seat or hybrid can be steadier for high, stable volumes where a flat fee is cheaper than per-unit charges. Per token appeals to teams building their own agents who want transparent compute costs and will own quality themselves. Many buyers end up with a hybrid: a base platform fee for predictability plus a usage component tied to outcomes. Whatever you pick, negotiate the definition of the billable event in writing — that clause, not the unit price, decides whether the deal is fair.

Bottom line

Outcome-based pricing for AI support is a real improvement over paying for empty seats — but only when the outcome is defined honestly and you can verify it. Treat the per-resolution rate as the headline, not the answer. The questions that protect you are: what exactly counts as a resolution, am I charged when the AI fails or escalates, and does this model still pencil out at my real ticket volume? Get those in writing and the model is fair. Skip them and the "pay only for success" promise can quietly become the most expensive line on the bill.

Sources and further reading

Sources

Intercom: Announcing Fin — resolution-based pricing
intercom.com
Zendesk: Pricing and automated resolutions
zendesk.com