An AI chatbot costs anywhere from nothing to several hundred a month, and the four pricing shapes behind that range matter far more than the number. Per message, per seat, per resolution, or a flat licence where you bring your own model key. The model calls themselves are usually the smallest line: a single answered question is roughly 1,500 tokens in and a few hundred out once you are pasting retrieved passages into every request, which for a few thousand conversations a month lands in tens of euros rather than hundreds.
So the useful question is not what it costs. It is what a bad month costs, and only one of the four shapes gives you a comfortable answer.
What are you actually paying for?
Four things, bundled differently by each vendor. Reading which of the four is doing the work in a given price is most of the comparison.
| Pricing shape | What you pay for | What makes the bill move |
|---|---|---|
| Per message | Each exchange with a visitor | Your traffic, including traffic you did not want |
| Per seat | Human agents with logins | Team size, including the person who logged in once |
| Per resolution | Conversations the bot closed | The vendor’s definition of resolved |
| Flat licence, your own key | The software, once or yearly | Nothing. The model bill is separate and public |
The bottom row is the only one where the software price and the usage price are separated, and separating them is what makes both of them checkable. You pay the vendor for software and the model provider for tokens, at a rate published on a page you can read without a sales call.
The per-resolution row deserves a specific warning. Resolved is not a fact, it is a definition, and the party defining it is the party billing you. Ask exactly what counts: a visitor who closed the window, a conversation with no follow-up within an hour, a thumbs up. Each of those produces a different invoice from identical behaviour.
What does the model itself cost to run?
Less than almost everybody assumes, and it is worth doing the arithmetic once so the number stops being scary.
A retrieval answer is two calls. One small embedding request to turn the question into numbers, which is priced in fractions of a cent, and one completion request carrying the question plus the passages you retrieved. That completion is the expensive half, and retrieval is what makes it expensive: you are sending four passages of your own content with every single question.
Assume 1,500 tokens in and 300 out per answered question, which is a normal shape for a retrieval bot. A thousand conversations a month is then well inside the tens of euros at published rates. Two thousand is still tens of euros. The number only becomes interesting at volumes where you would also be hiring.
Two things move it more than the model choice does. Caching identical questions, since the same four questions arrive constantly and a cached answer costs nothing. And how many passages you paste in: four short ones answer as well as eight long ones and cost half as much.
How do you estimate your own volume?
By counting what you already get, then expecting more, which is the part that catches people out.
Start with the month you have. Count the support emails, the contact form submissions and, if you run live chat, the conversations. That is your current demand and it takes twenty minutes to pull.
Then adjust upward rather than down. A chat widget does not simply move existing questions to a new channel; it lowers the cost of asking, so people ask things they would never have written an email about. Sizing, compatibility, whether something is in stock in a colour. Those questions were always there and were being answered by a visitor guessing or leaving. Your first month with a widget will usually carry more total contacts than your last month without one, and that is the product working rather than failing.
Which is exactly why the shape matters. On a flat licence that increase is free. On per-message pricing you have just made your own bill go up by succeeding.
Two caps are worth setting on day one, before any of this is theoretical. Set a hard spend limit in your model provider’s dashboard, so a runaway loop or a scraper cannot spend a quarter’s budget overnight. And put a rate limit on your own endpoint, per session and per address, which you can only do if requests pass through a server you control rather than going from the browser to the model.
Neither takes long and both are the kind of thing nobody does until the first surprising invoice.
What happens on a bad month?
The shapes separate, which is the entire argument for caring about them.
| Scenario | Per message | Flat licence with your own key |
|---|---|---|
| A quiet month | Cheap, and you notice you are paying for little | Same fixed fee, plus almost no tokens |
| A launch week | Bill rises with attention, not with orders | Same fixed fee, plus a small token bump |
| A scraper finds your widget | Bill rises for traffic that will never buy | Same fixed fee, plus tokens you can rate-limit |
Row three is not hypothetical. A chat endpoint is an open invitation to anything automated, and the defence is a rate limit on your own server, which you only have if the requests pass through a server you control.
The pattern across all three rows is the same. Per message ties your cost to attention, and attention is not revenue. A delivery delay that makes three hundred customers ask the same question is your most expensive day under that model and your least profitable one, at the same time.
None of which makes per-message pricing dishonest. It makes it a risk to accept deliberately, with a cap configured, rather than by default.
What costs are not on the pricing page?
Three, and the third is the one people discover late.
Overage rates. A bundle of messages has a price per message inside it and a different price above it, and the second number is often several times the first. It is usually in the terms rather than on the pricing page. Ask for it.
Seats you forgot. Per-seat pricing accumulates quietly: somebody is added for a busy fortnight and is still billed a year later. This is a calendar reminder rather than a technical problem, but it is real money and nobody owns it.
The cost of leaving. Not a fee. The question is whether you can export your transcripts and your knowledge base in a format usable somewhere else. If you cannot, the true price of the product includes never being able to change your mind, and that is worth more than any monthly difference. Test it with one support email before you commit, not on the day you want to move.
Our own position on all of this is on pricing, and what you are permitted to do with the software is in the licence.
When is the free option the right answer?
More often than the paid pages suggest, and specifically when you want the decisions.
AI Engine is free, mature and sits at 100,000 installs against the plugin API. You bring your own key, which means the per-token cost reaches you unmarked up, and you choose the model, build the knowledge base and style the widget. If those are choices you want to make, this is the best value in the category and nothing paid will beat it on price.
What free costs you is attention, and that is a genuine cost rather than a rhetorical one. The setup is a day or two. The tuning is ongoing. The checking, which is the part that decides whether any of this was worth doing, is yours forever. People who enjoy that get an excellent deal. People who wanted it answering by Friday have bought a project.
Paid products are selling that gap and some of them are worth it. The honest way to judge one is to ask what it does that you would otherwise have to build, and to notice when the answer is mostly hosting.
When should you not be paying for this?
Three situations, and the first is the most common by a distance.
If you get fewer than roughly ten questions a week, every option here costs more in attention than it saves in answering. An email address and a good FAQ page is correct and stays correct longer than most people expect.
If your site does not contain the answers, you are paying to index nothing. Writing the three pages that cover most questions, delivery, returns and sizing, is an afternoon and it improves the site for every visitor who never opens a chat window. Pay for software after that, not before.
And if the cost you are trying to remove is a person, look closely at what that person actually does. Automating the repeated questions is real savings; automating judgement is not, and a product priced on resolutions has an incentive to describe the second as the first.
For the shortlist itself, the market and the review data are in 826,000 WordPress sites and almost no AI, and what this category actually contains is in what an AI chatbot for a website actually is.
Questions people ask before moving
How much does an AI chatbot cost?
Anywhere from nothing to several hundred a month, and the spread is about pricing shape rather than quality. The free route is a plugin where you bring your own model key and pay the provider directly for tokens, which for a small site is usually tens of euros a month at most. The paid route is a product that adds the software, the hosting and the support on top, priced per message, per seat or per resolution.
What is the cheapest way to run a chatbot on my website?
A free plugin plus your own model key. You pay the model provider directly at their published per-token rate with nothing marked up, and the software costs nothing. What it costs instead is your attention: you build the knowledge base, write the prompt, style the widget and decide the refusal behaviour. That is a day or two of work, which is cheap if you enjoy it and expensive if you do not.
Why is per-message pricing risky?
Because your support load does not scale with your revenue. A product that briefly goes viral, a delivery delay that makes three hundred people ask the same question, or a scraper hitting your widget all raise the message count without raising your income. Per message means your worst month and your best month can be the same bill, arriving for opposite reasons.
How much do the AI model calls actually cost?
Usually the smallest line on the invoice, and smaller than people expect. A single answered question is roughly 1,500 tokens going in once you are pasting retrieved passages into the request, and a few hundred coming back. At published rates that puts a few thousand conversations a month in the tens of euros, not the hundreds. The software and the support around it cost more than the model.
Are there hidden costs in chatbot pricing?
Three that appear regularly. Overage rates above your message bundle, which are often several times the in-bundle rate. Per-seat charges for team members you added for one week. And the cost of leaving, which is not money: if you cannot export your transcripts and knowledge base in a usable format, the real price of the product is that you now cannot change your mind.
Is a free AI chatbot actually free?
The software is. The tokens are not, and neither is your time. Free plugins hand you the model key decision, which is genuinely better for cost control because nothing is marked up, but the setup, the tuning and the ongoing checking are yours. Free means you pay in attention instead of money, and that is a real trade rather than a trick.
