“How much does an AI chatbot cost?” has two answers, and almost everybody asks about the wrong one first.
There is the build, which is a number you pay once. And there is the running cost, which continues forever and scales with how successful the thing is. For a small business the second one usually decides whether the project makes sense at all.
The running cost is per conversation, not per month
Commercial language models bill by token — roughly, by the amount of text going in and coming out. That has a consequence people find counter-intuitive: a chatbot that nobody uses is nearly free, and the same chatbot on a busy site is a real line item.
Worse, the cost per conversation is not fixed either. A well-built assistant sends context with every question — your services, your hours, relevant documents — so the input is much larger than the visitor’s sentence. Add conversation history and each successive turn costs more than the last.
None of this is a reason not to do it. It is a reason to model it before building rather than discovering it in month three.
What actually drives the number
- Conversations per month. The headline driver, and the one people underestimate — a chat widget gets used far more than a contact form, because it costs the visitor nothing.
- Turns per conversation. A three-turn exchange costs several times a one-turn answer, because history is re-sent each time.
- How much context you send. Retrieving three relevant paragraphs costs a fraction of sending your entire FAQ with every message. This is an engineering decision with a direct monthly cost.
- Which model. There is often an order of magnitude between tiers, and the cheaper one is frequently sufficient for “answer questions about our services”.
The last two are where the engineering earns its keep. A carelessly built assistant and a carefully built one can differ by 10x in running cost while being indistinguishable to a visitor.
What it breaks
The failure modes are more predictable than people expect, and all of them are designed around rather than eliminated.
- Confident wrong answers. The model will state something false in the same tone it states something true. The mitigation is grounding it in your actual content and having it decline rather than guess — which is a prompt and retrieval decision, not a model choice.
- Answering questions it should escalate. Pricing, availability, anything contractual. These need hard routing to a human, not a good-faith attempt.
- Going stale. The assistant knows what it was given. Change your hours or your services and it will keep confidently reciting the old ones until someone updates its source.
- Being gamed. Public chat interfaces attract people trying to make them say something embarrassing. Worth deciding in advance how much you care.
What it is worth
The honest case is narrower than the marketing suggests, and it is real.
The value is in questions arriving outside working hours from people who will not wait until morning, and in the same forty questions your team answers constantly. If a meaningful share of your enquiries are “do you do X”, “are you open on Sunday”, “how does Y work” — that is genuinely well suited, and every one handled is a message your team did not write.
The value is much weaker if your enquiries are mostly unique, high-value, and already converting. Then the bottleneck is not response volume, and a chatbot is an expensive way to add a step in front of a conversation you wanted to have anyway.
Before you build one, check it is not a form
A good share of “we need a chatbot” conversations end somewhere cheaper.
If the forty questions are always the same forty, a well-written FAQ page answers them for free, ranks in search, and never hallucinates. If people are asking because your site does not say, the fix is to say it. If they are asking for a quote, a structured form gathers better information than a conversation does.
We would rather tell you that than build you something. The general principle is on AI consulting: work out which workflows genuinely benefit, and which are just missing a form.
How to decide
Count the repetitive enquiries you get in a month and how long each takes to answer. Then estimate what a conversation costs to serve and what volume you would actually see. If the arithmetic does not work at your current volume, it will not work at double.
That calculation takes about twenty minutes and saves a lot of regret. Tell us what your inbound looks like and we will do it with you — including the version where the answer is a form.
