A chatbot that answers from your content, or admits it cannot.
From A$3,500 — ex GST, with the running costs passed through at cost. The useful thing about a support chatbot is that it answers the same forty questions a hundred times a week. The dangerous thing is that it will also answer the forty-first, confidently, from nothing. This one is built so it cannot.
A chatbot is a retrieval problem wearing a conversation. It is only as good as the content it can find: if your returns policy does not say what happens after 30 days, no model can tell a customer what happens after 30 days — it can only guess convincingly. So the build is mostly about what the bot is allowed to see, what it does when it finds nothing, and how fast you learn which questions it could not answer. The conversation part is the easy part.
How it is built.
What is in the build.
Answers from your content only
Your pages, policies, specs and FAQs — retrieved first, then answered from. The model is not asked what it knows; it is asked what the passage says.
A link to the page it used
Every answer cites the page it came from, so a customer can check it and your team can fix the page rather than the bot.
A real “I don’t know”
When nothing relevant is found it says so and offers a person. This is the feature that makes the other answers trustworthy.
The missing-content report
Every question it could not answer, listed. It is the most useful output of the whole thing and most bots throw it away.
Handover to a human
Into the inbox or helpdesk you already use, with the conversation attached, so nobody starts the story again.
A ceiling you set
Per conversation and per month, enforced before the call, priced from the usage the API reports. It refuses rather than overspending.
Three situations where we will talk you out of it.
| If this is you | What we will say instead |
|---|---|
| You have little content | If the answers are not written down anywhere, a chatbot cannot find them. Write the twenty answers first — you will find that solves most of it, and we will say so instead of selling you the bot. |
| The questions need an account lookup | “Where is my order?” is a systems-integration job with a chat window on it. Doable, and priced as an agent rather than as a chatbot, because it has to act. |
| You want it to sell | A bot that pushes a product it cannot explain damages the thing it is selling. We will build one that answers questions well, which sells considerably more. |
What it costs, from.
| What | From | Covers |
|---|---|---|
| AI chatbot | A$3,500 / US$2,300 | running costs at cost |
AUD ex GST for Australian clients, USD everywhere else. Model usage is passed through at cost under a ceiling you set, and hourly work is A$140 / US$95. The cost guide prices a larger build with every hour shown, and its form carries the estimate into the message.
Straight answers.
Will it make things up?
Not from nothing: it may only answer from passages retrieved out of your own content, and it is instructed to say it does not know when retrieval finds nothing good enough. That is a design, not a promise about a model — and it is why the architecture above has the orange path. A bot without that path will answer anyway, every time.
What does it cost to run?
Usage, at cost, under a ceiling you set. A support bot answering a few hundred questions a month on a cheap model is dollars rather than hundreds. We run the same pattern on our own free generator: a US$2 monthly budget, priced from the token usage the API reports, and a refusal that names the limit when it is reached.
Can it do more than one language?
Yes, and the retrieval is the constraint rather than the model: it can only answer in a language your content exists in, unless you want it translating your policies on the fly, which we would advise against for anything legal.
How long does it take?
Two to four weeks, and the long pole is almost always the content audit rather than the build. We will tell you which of your pages are good enough to answer from before we take the job.
Where does it live?
A widget on your site, or in the helpdesk you already use. Self-hosted either way, with no third-party script if it is on our build — the same posture as the rest of our work.
Where to go from here
- AI agents — when the job is to do something rather than answer something
- AI development — the stack, the architecture and the cost controls in full
- The cost guide — price it yourself, with every hour shown