Why Per-Resolution AI Pricing Is Broken (And How to Fix the Incentives)

Per-resolution pricing for AI support sounds like the most honest model in the market: you pay when the machine does the work, not for seats or volume. The pitch aligns the vendor's incentive with your outcome. In practice, the model is only as honest as the definition of "resolution" — and most vendors keep that definition vague on purpose. When the unit of billing is defined by the party collecting the money, the incentive doesn't align with your outcome at all. It aligns with billable events.

The first failure is counting deflection as resolution. A customer asks a question, the bot answers something, the customer closes the chat or simply stops replying, and the meter ticks. But "the conversation ended" is not the same as "the problem was solved." People abandon chats out of frustration constantly. If your AI vendor counts an abandoned session as a win, you are paying a premium specifically for the cases where you most failed the customer.

The second failure is billing the handoff. A meaningful share of AI conversations should end with the AI recognizing its limits and routing to a human. That is the system working correctly. Yet many pricing models charge for the AI's portion regardless, so you pay the machine for the easy first turns and then pay a human agent to actually close the ticket. You get double-billed on exactly the hard cases, and the vendor has a quiet incentive to delay the handoff to rack up AI turns first.

The third failure is the audit gap. Even where a vendor's definition is reasonable, you usually cannot inspect any individual charge. The invoice is a number. There is no receipt that says: here is the customer's question, here is the answer we gave, here is the evidence it was grounded, here is the signal the customer was satisfied. Without that, you cannot reconcile spend, you cannot dispute an error, and you cannot build trust in the meter. You are asked to take billing on faith — which is precisely the thing software is supposed to remove.

We think the fix is structural, not rhetorical. First, never bill for the absence of work: handoffs are $0, and silence or abandonment is $0. The AI should only earn revenue when it does something the customer can confirm was useful. Second, require evidence on every billable event — a verdict backed by what was asked, what was answered, and why the system believes the issue was actually resolved. Third, make every charge disputable. If you can't challenge a line item and have it adjudicated against the evidence, the meter isn't auditable; it's just opaque billing with a better story.

There is a fair counterargument: evidence and disputes add overhead, and a looser model is cheaper to run, so it can be cheaper to buy. That's true, and it's a real trade-off. But the cheaper model externalizes its cost onto your finance and support teams, who spend their time reverse-engineering an invoice they can't see into. Auditability isn't a tax on per-resolution pricing — it's the thing that makes per-resolution pricing safe to adopt at all.

The deeper point is about incentives over time. A pricing model is a contract about what the vendor is optimizing for. If "resolution" is defined loosely, the product will, over many small decisions, drift toward generating more billable events rather than solving more problems. If "resolution" is defined as a verifiable, disputable, evidence-backed outcome — and only that — the product is forced to optimize for genuinely resolving issues, because that is the only thing that pays. We are early and supervised in our own deployment, and we don't have years of data to wave around. But we'd rather be honest about the mechanism than impressive about the metrics: get the definition right, expose the evidence, and let customers dispute it. The numbers can come later, and they'll mean something when they do.