Chatbot pricing spans more than an order of magnitude, and unlike some markets the spread is honest — the cheap ones and the expensive ones are genuinely different things.
The ranges
From about $2,000 for a basic build. Typically a general model pointed at your website content. It will answer everything, confidently, and it will sometimes be wrong.
$6,000 to $18,000 for a single-purpose agent — one that qualifies leads, or answers support questions from your own material. This is where most small and medium businesses land, because most have one job they actually want done.
$5,000 to $50,000 for a custom build, depending on how much it has to know and what it connects to. The top of that range is an assistant wired into systems you already run.
$300 to $800 a month to keep one running, covering the hosting and the model it answers with.
For comparison: an administration assistant in Australia averages about $70,185 a year, and a customer service administrator sits between $65,000 and $75,000 — roughly $7,600 a month for one administrator once on-costs are counted. That is the comparison most businesses are actually making, even when they describe it as a technology decision.
What you are actually paying for
Not the conversation. Models are good at conversation now, and that part is close to free.
You are paying for the work of deciding what the assistant may say — and that turns out to be most of the project.
Consider a question as ordinary as "can you do this by the end of the month?" A general model will answer it. It has no way of knowing whether that is true, and a yes said in your name is a commitment your business now has to meet or retract. Multiply that by prices, availability, whether you cover a suburb, whether a material is in stock, and what your warranty actually says.
Every one of those needs a decision before launch: does the assistant answer, or hand over? The decisions are the deliverable. The chat window is the easy part.
What it costs to run each month
$300 to $800 a month is the common range, and it is the half of the price most quotes leave until last. Three things sit inside it.
Hosting. Small, fixed, and the least interesting part of the number.
The model it answers with. Usage-based, so it moves with how many conversations you actually have. A busy site costs more than a quiet one, which sounds obvious and is worth asking about directly: a monthly fee quoted as a flat figure either has a message limit in it somewhere, or the supplier is carrying the variation for you. Both are fine. Not knowing which is not.
Keeping it current. This is the one that decides whether the monthly is worth paying at all, and it is the one that never appears on a spreadsheet. Your services change, your prices change, you stop covering a suburb. An assistant nobody updates does not fail loudly — it keeps answering, fluently, about a business you no longer run.
Which means there are two quite different things being sold as a monthly fee, and the word hides it — the same confusion as a website care plan next to a development subscription. One keeps the assistant switched on. The other keeps it accurate. The first is worth a few hundred dollars and the second contains somebody's time, so if two quotes are far apart here, that is usually the difference rather than a margin.
The practical consequence: a build price with no monthly attached is an incomplete quote. Twelve months of running costs sits in the same order of magnitude as a mid-range build, so a supplier who has not raised it has either absorbed it or has not thought about it yet, and it is worth finding out which before you sign.
The expensive mistake, and it is not the build cost
An assistant that answers everything looks better in a demonstration and is worse in production.
The failure is not obvious, which is what makes it expensive. A confidently wrong answer about a price, a lead time or a safety procedure reads exactly like a right one — the person on the other end has no way to tell, and acts on it. You find out later, from a customer who is annoyed rather than curious.
So the property worth paying for is the opposite of the one that demonstrates well: an assistant that will say it does not know. Which comes back to what it has been given — an assistant is only as good as what it knows. That takes more work to build, because somebody has to define the boundary, and it takes more nerve to sign off, because it will visibly decline questions in front of you.
We build ours that way and would again. Our internal staff assistant refuses to answer anything it cannot cite a document for — on a question nobody has written up yet, you get "I do not have an answer for that", which is exactly what you should get.
Questions worth asking a supplier
- What will it refuse to answer, and how did you decide that? A supplier without a considered answer has not thought about the part that matters.
- Is the guard in the code, or in the instructions? Instructions to a model are guidance and can be outweighed by the rest of the prompt. A check applied to the answer before it is sent is not.
- What happens when it should hand over to a person? And how quickly does that person find out.
- Who keeps it current? Your services will change within the year. An assistant nobody updates becomes a liability on a schedule.
- Where does our data go, and on what terms? Ask specifically whether conversations can be used to train the provider's models.
What we charge
Our AI chatbots are an add-on to a Growth build, and only ever on a website we build. A chatbot is quoted after a conversation, because most of its cost is how much your business has already written down. The automation behind it — follow-ups when an enquiry goes quiet, routine replies — is a second add-on, quoted the same way, because it depends heavily on what it actually has to do.
Bought on its own, without a website build, our internal staff assistant — answering your team's questions about your own processes, with the source document attached to every answer — generally starts from $4,000 to set up and $600 a month after that. Most of the setup is writing and sourcing the answers it will be allowed to give, which is the same point as the rest of this article: that work is the product.