The three meters of AI customer service

Per AI interaction — sometimes sold as per message or per reply — bills for each individual response the AI sends, with no regard for whether the customer's problem ever gets solved. It is the most granular meter and the cheapest unit on the page, which makes it attractive for high-volume, single-touch queries like order status or balance checks that the AI clears in one turn. The catch is multiplication: a genuinely hard issue that takes eight back-and-forth turns costs eight times a simple one, so the model quietly penalises exactly the complex conversations where automation is most valuable. It also creates a subtle misalignment — a vendor paid per AI interaction has no structural reason to be concise, and a chatty agent earns more than a crisp one.

Because of that dynamic, pure per-interaction pricing is comparatively rare as a headline model; it more often appears bundled (a fixed number of replies folded into a credit) to smooth the multi-turn penalty. Read carefully, several "per-ticket" products are per-interaction underneath.

Per-session — also marketed as per-conversation or per-ticket — charges a flat fee each time the AI engages, no matter how many messages are exchanged or whether the issue is resolved. A session is one interaction window, usually closing after a defined period of inactivity. This is the most forecastable meter: you already know your conversation volume, so you can predict the monthly bill with reasonable accuracy regardless of how the AI performs. That predictability is the whole appeal for finance teams, and it neatly removes the per-interaction multi-turn penalty — a ten-turn conversation costs the same as a two-turn one.

The trade-off is that you pay for failure. Every conversation the AI cannot resolve and escalates to a human still incurs the full charge, so a low resolution rate inflates your true cost per solved problem without ever showing up on the invoice. The session boundary is also vendor-defined and a little arbitrary: a customer who returns three times about one underlying issue can be billed as three sessions, and where the inactivity window falls decides whether one problem is one charge or two.

Per-resolution — or per-outcome — bills only when the AI resolves an issue end-to-end, and it is the model whose incentives line up best with the buyer's: the vendor earns more precisely when the AI gets better, and failures are, in principle, free. It carries the cleanest ROI story and is the easiest to defend to a CFO, which is why it has become the default pitch for AI-native agents. But "resolution" is a vendor-defined term, and that single fact is where the model's honesty lives or dies. Two vendors both quoting $0.99 can produce invoices that differ two- or three-fold once you account for what each one actually counts as a billable resolution, the platform minimums underneath, and the seat or helpdesk fees stacked on top.

Two definitional gaps matter most. First, channel differences: a resolution on live chat is not the same artefact as one on email or voice — they differ in cost to deliver, in handle time, and often in how "resolved" is even detected — yet a flat per-resolution rate charges the same for all of them, so cheap chat deflections quietly cross-subsidise expensive voice outcomes (or vice versa) and a blended price hides the mix. Second, handover counted as resolution: some vendors, when configured or instructed to, count a workflow-driven handoff to a human as a billable outcome. So you can be charged a full "resolution" for a conversation the AI escalated rather than solved — the bill says success while the human queue says otherwise.