AI Chatbot Pricing, Explained for Busy People

Here is a fun exercise. Open the pricing pages of four AI support tools in four tabs and try to answer one simple question: "What will this cost me per month?"

You can't. Not directly. One page wants to know how many seats you need. Another charges per "resolution," whatever that means. A third sells you a bucket of credits. The fourth just says "Growth: contact us."

This is not an accident. Every vendor picks the pricing model that makes their product look cheapest for their ideal customer, and the models are deliberately hard to compare. I am going to decode all four, show you the hidden multipliers, and then give you a five-minute method for computing your own real cost across any of them.

The unit of pricing tells you what the vendor thinks you are buying. Keep that sentence in mind; it explains almost everything that follows.

The four pricing models, decoded

Nearly every tool in this market uses one of four models, or a hybrid of two.

Per-seat: you pay for humans

This is the classic help desk model. You pay a monthly fee per agent (per human, that is) who logs into the tool. The AI features are bundled in, or bolted on as an add-on per seat.

Per-seat pricing made sense when software was a tool humans used, like a hammer. You paid for hammers. But notice the strange incentive when AI enters the picture: the vendor earns more when you employ more people, and the whole point of AI support is to need fewer of them. If the AI does its job brilliantly, the vendor's revenue from you shrinks. Some vendors resolve this tension by adding a second meter (usage, resolutions, credits) on top of seats, which means you are effectively paying twice, in two currencies.

Per-seat is predictable and easy to budget, which is genuinely valuable. It suits teams where humans do most of the answering and AI assists at the margins.

Per-resolution: you pay when the bot claims success

This is the model Intercom popularized with Fin, and several competitors have followed. You pay a fee each time the AI "resolves" a conversation.

On paper this is beautiful alignment: no resolution, no charge. You pay for outcomes, not activity. In good months it really can work out that way.

The entire model, though, lives or dies on one definition: what counts as a resolution? Read the fine print of any per-resolution plan and you will usually find something like "the customer confirmed the answer helped, or the customer left without asking for a human." That second clause matters. A customer who gets a wrong answer and gives up in disgust often counts as resolved. They did not ask for a human; they just left. You get billed for the privilege of annoying them.

I am not saying vendors are acting in bad faith here; measuring true resolution is genuinely hard. I am saying that before you sign anything per-resolution, ask the vendor, in writing, exactly which conversation endings trigger a charge, and whether you can audit and dispute them.

Message credits: you pay per AI reply

The third model gives you a monthly allowance of credits, and AI activity consumes them. Simple in concept, wildly variable in detail, because everything depends on what consumes a credit.

Questions to ask on any credit-based plan:

• Does every AI reply cost a credit, or every customer message, or the whole conversation?
• If a customer asks three follow-ups in one conversation, is that one credit or four?
• Do canned answers, FAQ-style responses, or scripted flows consume credits, or only true AI generation?
• Do more advanced models or longer answers consume multiple credits per reply?
• What happens when you run out mid-month: does the bot stop, degrade, or auto-bill overage?

Two plans with the same headline credit number can differ in real cost by three or four times, purely on these definitions.

Flat tiers: you pay for a bracket

Finally, the simplest model: a flat monthly price for a usage bracket, like "up to N conversations." Predictable, easy to budget, no meter anxiety. The trade-off is that you pay for the bracket whether you use it or not, and crossing the bracket boundary forces a jump to the next tier, so the marginal cost of your busiest month can be steep.

Flat tiers are underrated for small teams, precisely because the bill never surprises you.

How do you compute your real cost?

Now the practical part. Headline prices are useless; what you want is your effective cost per conversation, computed from your own numbers. Here is the whole method:

1. Estimate your monthly conversation volume. Count support emails, chats, and DMs from the last three months and take the average. If you have no data, estimate from orders or signups: for many ecommerce and SaaS teams, somewhere between 5 and 15 percent of monthly customers get in touch.

2. Estimate messages per conversation. Read twenty real conversations and count the back-and-forth. Most support conversations run 3 to 6 customer messages. This number is the silent killer on credit-based plans.

3. Estimate your automatable share. What fraction of conversations could a bot plausibly handle end to end? Repetitive inboxes (order status, shipping, returns) often sit at 50 to 70 percent; consultative inboxes much lower.

4. Translate into each vendor's currency. Seats: how many humans still need logins? Resolutions: volume times automatable share. Credits: volume times automatable share times messages per conversation. Tiers: which bracket does your volume land in, and how close to the edge?

5. Divide the monthly bill by total conversations. Now every vendor, whatever their model, collapses to one comparable number: cost per conversation. Compute it for a normal month and for your busiest month of the year.

Here is the counterintuitive part: when teams actually run this exercise, the plan with the lowest headline price is frequently the most expensive one for them. A cheap-looking credit plan with per-message billing gets brutal when your average conversation runs five messages. A per-resolution plan looks pricey per unit but can beat it when your conversations are long and your resolution rate is honest. The ranking depends on the shape of your inbox, which is exactly why vendors cannot tell you the answer and headline prices cannot either.

One deliberate omission: I have not quoted a single competitor's actual price in this article, and that is on purpose. Prices, limits, and definitions in this market change every few months. Any number I print today would mislead you by summer. Learn the models, run your own math, and check the current pricing pages the week you decide.

The hidden multipliers

Three fine-print details move real-world bills more than the headline price does.

Do deterministic answers consume credits? A big slice of any support inbox is the same twenty questions on repeat, and those do not need generative AI at all; a pre-approved canned answer or a scripted flow handles them perfectly. Some platforms charge these the same as AI-generated replies. Others meter them separately or not at all. If your inbox is repetitive (and if you are wondering whether it is, this breakdown of how AI reduces repetitive support work will help you see the pattern), this single line item can dominate your bill.

Do follow-ups count? "One conversation" is not one message. If each customer message in a thread bills separately, multiply everything by your messages-per-conversation number from step 2. A plan that looked half the price just doubled.

What do spikes cost? Black Friday, a shipping delay, a viral post. Your busiest month can run at two or three times normal volume. On flat tiers you might jump a bracket; on credits you hit overage rates, which are often priced well above the in-plan rate; on per-resolution you simply pay linearly more. Model your peak month explicitly before signing an annual deal.

There is also a structural way to shrink the whole equation, regardless of vendor: route fewer conversations through the expensive path in the first place. Not every question deserves a generative AI answer; many deserve a canned answer or a structured flow, and some deserve a human straight away. The triage framework in not every customer question deserves AI is worth reading before you size any plan, because it changes your step 3 number.

Since I work on Fetchply, one honest disclosure about how we fit into this picture: on Fetchply, Instant Answers (pre-approved canned responses) and Guided Flows (structured step-by-step processes) consume zero AI messages; only genuinely AI-generated replies count against your allowance, and there is a free plan with 200 AI messages per month to test your real consumption. For repeat-heavy inboxes that changes the math meaningfully, because the repetitive majority of traffic becomes free to serve. Whether that makes it the right choice for you still depends on your own numbers; run the five steps above against us too.

• Four models dominate AI support pricing: per-seat, per-resolution, message credits, and flat tiers. Each makes a different inbox look cheap.
• Never compare headline prices. Compute your effective cost per conversation from your own volume, messages per conversation, and automatable share.
• The fine print moves bills more than the price: whether canned answers consume credits, whether follow-ups each count, and how "resolution" is defined.
• Model your busiest month, not your average month, before signing anything annual.
• Reducing how many conversations need generative AI at all (canned answers, structured flows, sensible triage) shrinks the bill on every model.

Which pricing model is cheapest?

None of them, universally. Per-seat favors human-heavy teams, per-resolution favors long conversations with honest resolution definitions, credits favor short conversations and low volume, and flat tiers favor predictable volume. Compute your own cost per conversation across all four; the ranking depends on your inbox.

What should I ask a vendor before signing a per-resolution plan?

Ask for the written definition of a resolution, whether a customer who silently abandons counts as resolved, whether you can audit and dispute charged resolutions, and what a resolution costs beyond your included allotment.

How do I estimate volume with no support history?

Work backwards from customers. Take monthly orders or active users, assume 5 to 15 percent generate a contact, and pick the higher end if your product involves shipping, sizing, or billing. Then start on a free or monthly plan and replace the estimate with real data after 60 days.

Are free plans actually usable?

Often, yes, for validation. A free tier will not carry a busy inbox forever, but it answers the two questions that matter before you spend: what your real conversation volume is, and what share the bot can genuinely handle. Those two numbers are exactly what you need to price every paid plan accurately.