Short answer: somewhere between one cent and ninety-nine cents, depending almost entirely on how the vendor counts — and only secondarily on which model runs underneath. That is a ~99× spread for products doing broadly the same job on broadly the same class of model.
The spread is not a quality gap. It is a pricing-model artefact. Here is every published rate we could source, what each vendor is actually counting, and the arithmetic behind our own number.
Five meters, and why the sticker price lies
Nobody sells "an AI conversation". They sell five different things and call them all AI pricing:
| Meter | Who uses it | What you are buying |
|---|---|---|
| Per resolution | Intercom Fin $0.99 | A flat fee, but only when the AI closes it alone |
| Per conversation handled | HelpCrunch $0.29 · Tidio $0.65 · Freshdesk $0.49 · Crisp $0.05–0.10 | A flat fee per thread, win or lose |
| Per message | Tawk.to $0.030 | A flat fee per turn — chatty customers cost more |
| Per unique monthly user | SberDevices SaluteBot | Not a conversation meter at all |
| Per token | Chatonio | What the answer actually cost, times a published multiplier |
A flat fee for a variable-cost event means the vendor has to price their worst case and keeps the difference on every easy one. A two-line "what are your opening hours?" costs the vendor a fraction of a cent — and costs you 29¢, or 65¢, or 99¢. A token meter is the only one where a cheap answer is actually cheap.
The counter-argument, in full
Resolution pricing cuts both ways, and mostly in Intercom's favour on the comparison we are about to draw. You pay $0.99 only when Fin closes the conversation without a human; the ones it fails at are free. So the fair per-conversation figure is $0.99 × resolution rate:
| If Fin resolves… | Blended cost per conversation |
|---|---|
| 30% of conversations | $0.30 |
| 50% | $0.50 |
| 70% | $0.69 |
So the honest headline is not "99× cheaper". It is 30–70× cheaper, depending on how good Fin is on your queue — and we are happy to quote the low end.
Note what per-conversation pricing does not get you, though. HelpCrunch's docs define the billable unit as "the total number of conversations handled by selected AI agent" — handled, not resolved. You pay the 29¢ whether the AI solves it or gives up and hands it to a human. "Deflected" and "resolved" exist in their product, but as analytics, not as billing units.
What every vendor charges, side by side
Each step to the right is roughly 10× more expensive — on a linear axis everything under 5¢ collapses into a single pixel.
Vendor pricing as published, September 2026. Hatched bars are derived rather than quoted rates. Intercom bills per resolution, so only conversations its AI resolves are charged at all.
How these were sourced. First-party published rates: Intercom, HelpCrunch, Tidio, Freshdesk, Crisp, Chatling, Voiceflow. First-party arithmetic: Jivo (11,041–12,990 ₽ per 2,000 dialogues) and HelpCrunch ($29 per 100 AI conversations).
What we deliberately left out. Zendesk — the ~$1.50 per automated resolution figure is widely repeated, but we could only source it to third-party write-ups, not Zendesk's own rate card. We are not going to argue against a number a vendor has not published. Tawk.to — priced per message; converting to per-conversation needs a messages-per-conversation assumption we would be inventing.
One number still circulating that is simply out of date: Freshdesk's Freddy at $0.10 per session. That meter was retired. The current published rate is $0.49. If a comparison page quotes you 10¢, it has not been updated.
The part the price tag hides: who decides when a human steps in
Cost per conversation is only half the question. The other half is what happens when the AI is wrong — and that is where the per-conversation vendors are quietly expensive in a way no rate card shows.
In HelpCrunch, the AI decides when a human gets involved — not you. The customer keeps talking to the bot until the bot concludes a person is needed.
That model runs right through the product. Handoff is a routeToTeam call the AI makes on its own judgment, per its instructions. Their analytics has no "an agent took over" outcome at all — every conversation is classified as either deflected or routed to team. No pause-the-AI or claim control appears in any of their operator-facing documentation. And their own cookbook documents a configuration where the AI is instructed never to hand off, "even if the user asks".
If you have ever watched a qualified lead work through a bot loop while you sat in the inbox unable to step in, you already know what that costs — and it is not 29 cents.
Meanwhile the meter runs. You are paying 29¢ for that conversation from the first AI reply, regardless of how it ends.
And when the quota runs out, HelpCrunch's documentation is unambiguous about what happens next: "the AI Agent block will be ignored in chatbot flows… usage stops until the next billing cycle." The bot goes quiet mid-flow. Not escalated — skipped.
Two more things worth knowing before you budget:
- The AI quota does not scale with your team. HelpCrunch Pro includes 50 AI conversations a month. At one seat, that is 50. At ten seats — $250/month — it is still 50. Scaling to a useful 1,000/month is another $200/month on top.
- Multilingual knowledge costs multiples. Their AI data sources are billed per item, and "one article added in 3 languages counts as 3 data sources."
Our number, and how we got it
Chatonio bills the tokens each answer actually uses, drawn from a prepaid wallet, times a multiplier that is printed on our pricing page. A typical conversation runs about a cent:
| Plan | Markup | Typical conversation |
|---|---|---|
| Free | 2.0× | ≈$0.013 |
| Advanced | 1.8× | ≈$0.011 |
| Pro | 1.6× | ≈$0.010 |
| Ultimate | 1.5× | ≈$0.009 |
Every AI vendor has a markup. Most bury it inside a flat per-resolution price where you cannot see it — and where it silently grows every time the underlying model gets cheaper. Ours is a number you can multiply.
It is a typical conversation, not a cap. A short exchange costs less; a long one with a lot of knowledge-base context, image analysis or several tool calls costs more. Token pricing means you see that variation instead of subsidising it.
And you can stop the AI mid-sentence
Every conversation carries a three-way Auto / On / Off switch. Off is a hard stop, not a preference — and it applies during generation, not just before it. An AI reply can take the better part of a minute to compose; if you switch the AI off in that window, the reply is never sent. It is kept as an internal note only you can see, so you can still reuse the wording if it was good, and the conversation is escalated so a human is paged.
Claiming a conversation does the same thing: the AI stops answering new customer messages the moment it is yours. You never have to watch a bot talk to your lead while you wait for it to decide you are needed.
Where "cheapest" is and is not a fair claim
Two products land at or below our number, and we would rather say so than have you find out: Chatling at about $0.025 on its cheapest model, and Voiceflow at $0.031 on its economy model — and even that is mostly platform fee, since their credit floor alone is about $0.030 a conversation before any model cost. On per-conversation cost alone they are in our range.
They are also not the same kind of product. Both are chatbot builders: you get a bot on your website. You do not get a shared inbox, operator seats, shift routing and assignment, a helpdesk with tickets and a knowledge base, Telegram, WhatsApp, Viber and email as first-class channels, or a handoff where a named human takes over and the AI keeps covering until they arrive. If you already have a helpdesk and only want a bot on the site, a builder is a legitimately cheaper answer and you should use one.
Against support platforms — products doing the whole job — the cheapest published rate we could find is Crisp at $0.05–0.10. We are a fraction of that, and roughly a thirtieth of HelpCrunch's entry rate.
What to actually ask a vendor
- Do I pay when the AI fails? Per-conversation and per-message meters usually mean yes.
- What happens when the quota runs out — does it escalate to a human, or does the bot just stop?
- Can my agent take a conversation off the AI before the AI decides to let go? Ask for a straight yes or no — some vendors will tell you plainly that you cannot.
- What is the markup on the underlying model, and what happens to my price when that model gets cheaper?
- Does the AI allowance scale with my team, or is it flat per workspace?
Vendor pricing above is as published in September 2026, with each figure taken from the vendor's own pricing page or documentation. Rates change without notice — check before you commit.