Everyone charges per resolution. Nobody means the same thing.

Zendesk’s legacy bots go into maintenance mode on August 31. That is two weeks away. On December 10 they stop working, and every team still running bot builder or autoreplies has to move to the upgraded AI agent experience. Near the bottom of the end-of-support notice sits one line worth more than the rest of it: switching may change your automated resolution metrics and the costs attached to them.

So a lot of support teams are re-picking their AI this month, and the new bill is metered per resolution.

That is the model now. Support AI has stopped being priced like software and started being priced like piecework. Published rates run from about $0.49 at Featurebase to $2.00 for a Salesforce Agentforce conversation, with Intercom’s Fin at $0.99 as the number everyone quotes back.

Here is the trouble with comparing those numbers. A resolution is not a standard unit. It is whatever each vendor’s definition says it is, and the vendor writes the definition, and the definition is what makes the meter click.

What makes the meter click

Zendesk publishes theirs, which is more than a lot of vendors do. For an email or web form ticket, you are billed after 72 hours of inactivity if the AI gave a generative reply, no human agent answered the ticket, a second model reviewed the reply and confirmed it was relevant, and the customer either said it helped or said nothing at all. Escalated conversations are not evaluated for billing.

Read that again watching the money. Two things trigger a charge: the customer saying yes, and the customer saying nothing. Silence has a lot of reasons behind it. The relevance check is a real gate and it does throw out bad answers, but no model reading a transcript can tell the difference between a person who got what they needed and a person who read the reply, gave up, and went hunting for a phone number.

Other vendors draw the line in other places. Some bill any conversation that ended without a human joining. Some require a positive satisfaction rating, which produces smaller invoices on identical traffic because most customers never answer the survey. Some bill only when the agent completed a logged action, a refund issued or an order actually tracked. A few wait several days to see whether the customer comes back with the same problem.

Four vendors quoting a dollar can be selling four different things. That gap shows up on the invoice, not in the demo.

The definition is a design document

This matters well past procurement. Whatever counts as billable becomes the thing the product gets tuned to produce.

If resolved means no human touched it, then a handoff costs the vendor revenue. Nobody has to be cynical for that to bend a roadmap. Every threshold about when to escalate, every default about how hard to keep trying before letting go, sits next to a number that shrinks when the agent gives up the conversation. And the setting you care about most, the point where the machine stops and a person starts, is the one on the far side of that incentive.

Which is a strange thing to hand over, since escalation is most of the safety story in AI support.

So the useful questions for a vendor are not really about price.

What event bills me, in one sentence. Does a handoff to my own team cost you anything. Can I see per-conversation logs showing what was billed and why. If the same customer comes back tomorrow with the same problem, do I pay twice.

Get the answers in writing. A rate card without a definition is not a price.

Where we stand

We meter per resolution too, and we put the definition on the pricing page instead of in a contract you sign later. One conversation Celeste answers, however many questions are packed into it, billed when the reply goes out. Ask five things at once and it counts once. A live chat bills only if it actually landed, meaning a thumbs up, or the visitor left on a real answer they never marked unhelpful with no teammate stepping in. Go quiet for about a week and open a fresh exchange, and that is a new one.

We do not charge when Celeste hands off to your team, when a draft never sends, or when nothing happens. No per-seat fee. You can set a monthly ceiling on the whole thing.

The free handoff is the part to look at. Escalation should cost you nothing, because the rules about when to escalate belong to you, and they should not be sitting next to somebody else’s revenue.

One last thing about pricing support by the unit. It makes the trade visible. When a resolution costs 60 cents on our $99 plan and the thing it stands in for is an hourly wage, the saving stops being a claim in a deck and becomes a line on an invoice. Gartner expects the generative AI cost per resolution to pass what an offshore human agent costs by 2030, so that comparison gets messier from here. Either way, the number is written down now. That is the number The Dividend Standard runs on. Every year we give back the greater of 5 percent of our revenue or 70 percent of our profit, as ownership, to the workers this automation displaces, and the meter that bills you is the same meter that measures what we owe.