What an AI support assistant costs, what the other tools charge for, and how to train one properly — with calculators you can put your own numbers into.
Fin charges $0.99 when nobody asks a follow-up question. Eight alternatives, compared on the event that triggers a charge rather than the headline price — because that is what decides the invoice.
pricingcompareintercom
1,200
14%
What counts as resolved
Billed under this rule
$974
Intercom Fin, $0.99 per outcome
Zinx Chat, per reply
$79
Pro plan — every answer counts the same, resolved or not
Counted as resolved984Billable events under this rule
Silent168Of which nobody confirmed anything
Escalated to a person216Assumed, not measured
168 of those outcomes are someone leavingFin's pricing page defines a resolution as one where the customer confirms it, or simply does not ask for more help. A 24-hour quiet period turns silence into an assumed resolution.
What it costs
What an assistant costs, worked out rather than asserted.
3,000
Model
On our key
$199
Business plan
On your own key
$80
Pro plan, plus $1.40 billed to you by the provider
Replies spent per answer0.40Platform only, on your key
On our key0.87Platform plus the model
Model cost per answer$0.00047At the provider's list price
Your own key saves about $119 a monthPast a few thousand answers the model is most of the bill, and buying it directly is cheaper than buying it bundled. Assumes 2,560 tokens in and 420 out per answer. A longer knowledge base does not change this; more retrieved passages do.
Past a few thousand answers a month the model is most of the bill, and buying it directly beats buying it bundled. Work out your own break-even, and see where the tokens in one reply actually go.
600 a month
3 answers
Zinx Chat
$79 / mo
Pro plan
Chatbase
$120 / mo
1,800 messages at $40 per 700
Gorgias AI Agent
$590 / mo
600 × $0.90, and each one is a ticket too
Intercom Fin
$594 / mo
600 outcomes × $0.99
Zendesk AI agents
$955 / mo
600 resolutions × $1.50, plus one $55 seat
Zinx Chat is cheapest here, at $79 a monthThe spread is $876 a month between cheapest and dearest for the same 600 conversations. Drag the answers-per-conversation slider: the per-message plans climb with it and the per-outcome plans do not.
Worked out rather than asserted: what a thousand support conversations cost under per-resolution, per-message and per-reply pricing, with a calculator you can put your own numbers into.
Compared
What the other tools charge for, in their own units.
1,200
55%
70%
Help centre, actually resolved
92
Free per visit, and the only channel nobody thanks you for
Assistant
776
568 of them bounced back out and had to be answered again
A person
332
Everything with judgement in it
Monthly cost$1,255Assistant plus agent time
If people answered all of it$4,200At $3.50 a ticket
A help-centre page$45Writing one good page, once
About $2,945 a month of the differenceNotice what the middle slider does not do: raising help-centre coverage barely moves the total, because Gartner found only 14% of issues are fully resolved in self-service. A page that exists is not a page that answered anybody — the assistant is largely reading the pages your customers could not find.
A help-centre page costs nothing per visit and resolves far less than you think. Work out which of the three channels should take each question, and what the mix actually costs you per month.
Their own published prices, read on 2026-10-07. Entry prices are the cheapest paid way in, which for several of them is billed annually. Follow the link on each name for what it says today.
Intercom Fin, Zendesk, Gorgias, Tidio, Chatbase, SiteGPT, CustomGPT and Crisp bill for different events. Here is what each one charges for, what that does to your invoice, and when each model is the right one.
Guides
How to set one up, and what it does with what you give it.
1,000
82%
23%
9%
“AI handled”
82.0%
Every conversation it replied in. The number on the slide.
“Resolved” by silence
63.1%
Nobody followed up, so it is counted as solved. Includes everyone who gave up.
Plausibly resolved
55.7%
Answered, no follow-up, and no sign they abandoned it.
A 26.3% spread, from one inbox74 people left without a word and 189 asked again. The middle bar counts the first group as a success, which is why a vendor whose billing depends on silence will quote it.
A deflection rate is a definition before it is a number, and the flattering definition is always available. The four metrics worth tracking, and the containment trade-off a randomised trial actually measured.
What the model was given
A shopper asks: “Do you ship to Norway, and how long does it take?”
Yes, we ship to Norway. Delivery usually takes 3–5 business days and shipping is free over $50.
No source to show.
Not grounded in anythingFluent, specific, and invented. Nothing in it came from your store — the model is completing a sentence that sounds like a shipping policy.
Retrieval cuts invented answers by more than half and does not eliminate them. What the measured numbers are, why a confident wrong answer is the dangerous kind, and what to do about the remainder.
900
18%
4 min
Order-status questions16218% of 900
Answerable from the order record13885% of them; a lost parcel still needs a person
Time back each month9 hAt $25 an hour, fully loaded
About $230 a month of someone's timeThis is the honest part of the pitch, and the dull one. It is a lookup with a date in it — no judgement, no tone, no risk of inventing a refund policy. The default share here, 18%, is Gorgias's figure across its own ecommerce customers.
Order status is 18% of ecommerce questions and answering it is a lookup. Here is what an assistant can read from your catalogue, what it must never read, and what the dull questions are worth.
What happens when it cannot answer
Where is my order? It said delivered but nothing arrived.
I’m sorry, I don’t have information about that.
A dead endNothing is wrong with the sentence. What is wrong is that it is the last one: the visitor has no next move except to leave, and they will not try the chat again.
87% of customers say a company using AI for support must still offer a human. What that means in practice: the three endings a stuck conversation can have, and what each one costs you.
How long does delivery take? — nothing to answer from
Can I return it if it doesn’t fit? — nothing to answer from
Is it machine washable? — nothing to answer from
Which size am I if I’m 5'9"? — nothing to answer from
What do you actually sell?
Do you ship to Spain, and what’s the duty? — nothing to answer from
My order says delivered but isn’t. — nothing to answer from
1 of 7 answerable — 14%Marketing pages are written to persuade someone who has not decided. Almost nothing a buyer asks after deciding is on them.
An assistant is not trained on your site in the way people assume. Here is what retrieval actually hands the model, which pages earn their place, and why the dull documents beat the marketing ones.