Why does nobody publish a price per conversation?
Almost every help desk and AI support tool prices on the unit that suits its own billing system, not the unit you care about. Some charge per seat, so your bill tracks headcount. Some charge per AI resolution, so it tracks how often the bot finishes a job without a human. Some charge per ticket or per contact, so it tracks volume but counts it in a way that may not match how you count a conversation internally. Open source tools charge nothing for the software and a fair amount in engineering attention instead.
Because the billing units differ, two vendors with similar looking sticker prices can produce very different costs per conversation for the same business, and neither vendor is being dishonest about it. Cost per conversation is a derived number, not a listed one. You get it by taking a quote, applying it to your own traffic pattern, and doing the division yourself. Anyone who hands you a universal per-conversation figure has quietly assumed a volume, a team size and a deflection pattern that are probably not yours.
What pricing models will you actually meet?
Per seat, per month. This is the classic help desk shape used by Zendesk, Freshdesk and, in its own variation, Help Scout. You pay for each person who can log in and answer, tiered by feature set, with AI capability usually sold as a separate line rather than included in the base. Cost per conversation falls as each agent handles more, and rises the moment you add occasional helpers, a seasonal hire or a manager who only needs to look.
Per resolution or per AI answer. Intercom's Fin popularized this, and Zendesk's AI agents work in a related way. You pay when the AI actually closes something. The logic is attractive because cost follows outcome, and the per-unit number is easy to reason about. The catch is that your monthly total is now a function of success: the better the AI gets at handling volume, the more line items it generates. Budgeting becomes a forecast rather than a fact.
Per ticket volume. Gorgias plans are sized by a monthly ticket allowance, with overages beyond it and automation priced as its own layer. This fits ecommerce, where ticket count is the natural unit, but it punishes the spiky months that ecommerce specializes in.
Per contact. Billing on the number of distinct people you talk to rather than the number of exchanges. Lower cost per conversation for chatty customers, higher for a wide audience that each asks one quick question.
Conversation packs. Tidio and its Lyro AI sell blocks of AI conversations. Simple to understand, and a decent fit at modest volume, though running out mid-month is a real operational event you have to plan around.
Self-hosted open source. Chatwoot gives you the software for free and leaves you the rest: hosting, upgrades, monitoring, model API usage and the hours of whoever maintains it. The software cost per conversation is zero. The total cost per conversation is whatever your engineering time is worth.
Flat rate. One monthly fee regardless of seats or volume. Corebee sits here at $99 per month with unlimited teammates and unlimited AI conversations, no per-seat charge and no per-resolution fee.
How do you calculate your own cost per conversation?
Start by defining the unit before you look at any price. Decide whether a conversation means a ticket, an inbound contact, a chat session or a thread, then stick to that definition for every vendor you compare. Mismatched units are the single biggest source of bad comparisons.
Next, pull a representative month of volume from whatever you use today, even if that is a shared mailbox and a phone log. Then build the full monthly bill for each vendor, not just the headline plan: seats for everyone who needs access, the AI add-on, channel extras such as voice or messaging, platform fees, overage charges, and any onboarding or implementation cost spread over a year. Divide the total by your volume.
Then do it twice more. Once for your quietest month and once for your worst: a product launch, a shipping delay, a billing error that reaches everyone at the same time. The gap between those two numbers tells you more than the midpoint does, because the expensive month is exactly the month you cannot postpone.
Which costs do people forget?
The four that reliably surprise owners are seats, escalations, channels and knowledge upkeep.
Seats creep. A per-seat plan quietly taxes the warehouse lead who answers three questions a week and the founder who wants read access. Teams respond by sharing logins, which costs them visibility into who answered what.
Escalations can bill twice. On a per-resolution model, a conversation the AI handles halfway and then passes to a person may count as a billable AI unit while still consuming human time. Ask each vendor precisely what counts as a resolution and what happens when the handoff is partial.
Channels have their own economics. Voice minutes and WhatsApp messaging fees are usually separate from the support platform, and the carrier side does not disappear just because your software is unmetered. Budget them as their own line.
Knowledge upkeep is labor, not software. Every AI support tool answers from whatever you give it, so someone has to keep your knowledge base and site content accurate. That hour is real whichever vendor you pick.
Why does a flat rate change the math?
Metered pricing has a structural quirk: it charges you more for the two things you were trying to achieve. More AI deflection means more billable resolutions. More people watching the inbox means more seats. You end up making operational decisions, like whether to invite a part-timer or let the AI take a whole channel, partly for billing reasons.
A flat fee removes that tension. At $99 per month with unlimited seats and unlimited AI conversations, every extra conversation the AI handles drives your cost per conversation down rather than holding it flat, and inviting another teammate costs nothing. Setup takes about ten minutes and needs no developer, with web chat, email, voice (an AI receptionist that tells callers it is an AI), WhatsApp, Slack, Facebook, Instagram, SMS and Telegram available, plus meeting booking through Cal.com or Calendly and syncing into HubSpot, Salesforce, Pipedrive, monday or Zoho.
Flat rate is not automatically the cheapest option, and it would be silly to pretend otherwise. If you handle a trickle of conversations with one person, a free tier or an entry-level per-seat plan can beat any flat fee on absolute dollars. Flat rate wins on predictability and on the cost curve as volume and team size grow, which is why it tends to look better the more your support load varies.
Which model fits which kind of business?
If your volume is genuinely small and stable and one person handles everything, optimize for the lowest absolute monthly price and accept the per-seat shape. If your volume swings with seasons, promotions or shipping cycles, be suspicious of anything that meters the swing, because the overage lands in your busiest week. If you have many occasional responders, per-seat pricing will cost you either money or visibility. If you have engineering capacity and strong opinions about data residency, self-hosting is a legitimate choice as long as you price the maintenance honestly. And if voice or WhatsApp is a primary channel rather than an afterthought, compare what each vendor includes before comparing anything else, since channel add-ons often outweigh the base plan. A side-by-side on the per-seat incumbents is a reasonable next step once you know which shape you want.
How should you test a vendor so the numbers mean something?
Run a trial on real traffic, not a sandbox. Point one live channel at the tool for a couple of weeks, keep your existing setup running behind it, and watch three things: how often the AI answers without help, how often it hands off cleanly when a customer asks for a person or when it is unsure, and what it does with the questions your documentation does not cover.
While that runs, read the billing clauses. Ask what the billable unit is, what happens on overage, whether AI capability is included or added, whether unused allowance rolls over, and whether the quoted price assumes an annual commitment. Then recompute cost per conversation using the trial's actual volume rather than your estimate, and check what a shared inbox view costs once everyone who needs access actually has it. A vendor that answers those questions plainly is telling you something useful about how the relationship will go.