What an AI support assistant actually costs per conversation

Every AI support tool publishes a price, and almost none of them publish the same unit. One charges per resolution, one per message, one per conversation, one per seat and then per resolution on top. The figures below take your own traffic and price it under each, so the comparison happens in money rather than in adjectives.

The numbers for Zinx Chat come straight from the table the billing code uses, so they cannot drift from what you would actually be charged. Everyone else's come from their own public pricing pages, with the date they were read printed under the table.

The unit is the whole argument

A pricing page that says $0.99 per resolution sounds precise. It is, right up until you ask what a resolution is. Intercom counts an outcome once per conversation and says so plainly — a handoff to a human is also an outcome, and so is a disqualification. Zendesk charges per automated resolution, on top of a per-agent seat. Gorgias charges per interaction and bills that same interaction as a helpdesk ticket as well, so one AI conversation lands on the invoice twice.

Per-message plans behave in the opposite direction. Chatbase sells credits; SiteGPT sells messages and charges ten of them for one answer from its stronger model. Under those plans a visitor who asks four follow-ups costs four times a visitor who asks one — which is to say, a curious customer costs more than an incurious one.

The question is never “what does it cost”. It is “what has to happen before I am charged”.

What the alternative costs

The case for any of this rests on a comparison with a person answering instead. Gartner (2019) put a live-channel contact at $8.01 against roughly $0.10 for self-service. That figure is old enough to be quoted carefully — it predates all of this — but the shape of it has not changed.

The honest version is less flattering than the brochure. Deflection is not free money: the conversations the assistant cannot answer still reach a person, and those people have now waited twice. Gartner found (2024) that only 14% of service issues are fully resolved in self-service, and that in 43% of the failures the customer simply could not find the right content. That is a content problem wearing a technology problem's clothes, which is why what you train it on decides more than which model answers.

The direction of travel

Two Gartner predictions are worth holding together, because they point opposite ways. The first is that agentic AI will resolve 80% of common service issues without a human by 2029. The second is that the cost of a generative-AI resolution will pass $3 by 2030, overtaking an offshore human agent.

Both can be true. Volume moves to the machine and the price per machine-handled conversation goes up, because model costs are a floor under every per-resolution plan. If you are buying on a per-resolution contract today, that is the line item that moves against you later. A plan with a fixed monthly allowance moves the other way: you know the bill in advance, and the vendor absorbs the model price.

How to decide

• Count conversations, not messages. Your analytics already has the number. Multiply by the vendor's unit, not by their headline.
• Price the bad month. A product launch or an outage triples support volume. Under per-resolution pricing that triples the bill on exactly the month you can least afford a surprise.
• Count the seats. Several AI prices assume you are also paying for the helpdesk underneath. Fin without Intercom's helpdesk has a 50-outcome monthly minimum; inside it, add a seat.
• Check what happens at zero. When the allowance runs out, does the assistant stop, or does it keep answering and keep charging?

We publish ours at chat.zinx.app/pricing: $29 a month for 600 replies, $79 for 2,100, $199 for 5,500, and 25 free replies to try without a card. Packs top up a plan rather than replacing it, and they do not expire. The comparison post sets those side by side with everyone else's.