AI chatbots that answer from your content and know when to stop
A chatbot is only as good as the answers you have already written down. We build assistants that draw on your own material, show where an answer came from, and hand over to a person the moment they are out of their depth.
What is AI chatbots?
AI Chatbots are customer facing assistants that answer questions and qualify enquiries using your own service information, policies and documented answers rather than generic web knowledge. They suit Australian businesses handling repetitive enquiries by chat, email or web form, where most questions already have a written answer and a person can take over when they do not.
Get a fixed written quote- Typical timeline
- 4 to 9 weeks
- What drives cost
- How many topics it must cover, how much content preparation is needed, which systems it connects to, and your conversation volume.
- Best for
- Repetitive enquiries that already have a documented answer
- You own
- The prompts, the knowledge base, the transcripts and the accounts
- Built with
- Retrieval grounded answering, guardrails, CRM handoff
Your handover
What an AI chatbot can honestly answer
A language model does not know your business. Left to itself it will produce a fluent, confident description of a returns policy you have never had, because generating plausible text is what it does. That is hallucination, and it is not a defect waiting to be patched out of existence. The engineering response is retrieval: the assistant searches your approved content, receives the relevant passages, and is instructed to answer only from them and to say it does not know otherwise. That constraint removes most of the practical risk, and it is why we never put a bare model in front of customers.
- 01Retrieval grounded assistant deployed on your site
- 02Content audit naming the authoritative source per topic
- 03Guardrail and refusal policy documented in writing
- 04Evaluation set of real questions with approved answers
- 05Escalation path and live handover to staff
- 06Enquiry capture flowing into your CRM
- Transcript logging with a defined retention period
- Draft privacy collection notice wording
- Admin access to prompts, content and settings
The rest of the answer
Which means the real project is your content. If your freight terms live in three documents that disagree, the assistant will faithfully reproduce the disagreement. Before we build anything we inventory the sources, mark one as authoritative per topic, and put the gaps in front of you. Clients often find that exercise more valuable than the bot, because the same gaps have been costing their phone team an hour a day for years. Where an answer does not exist yet, somebody in your business has to write it, and no model can do that part.
How you know it is working
Deflection on its own is a misleading number. An assistant can close seventy percent of conversations by giving fast wrong answers that customers discover later, which produces an excellent dashboard and a worse business. We track containment alongside what happened next: did the customer come back within a day, did they ring anyway, did the enquiry convert, did a staff member have to correct something.
Early on somebody reads every transcript
So a person stays in the loop, and the shape of that changes over time rather than disappearing. Early on somebody reads every transcript. Later they read the escalations, the low-confidence answers and a random sample of the rest. Anything that turns out to be wrong becomes a new item in the evaluation set, which is the only mechanism that makes the system genuinely improve. Budget for that time in the running cost. An assistant nobody reviews degrades steadily as your business changes and the documents behind it drift out of date.
How the engagement runs
How an AI chatbot build runs
We start narrow on purpose. One topic, one audience, one channel, evaluated properly, then widened as the results justify it. Launching an assistant across every question your business receives on day one is how these projects get quietly switched off in month three, usually after one screenshot circulates internally.
- 01ScopingThe questions worth automating, ranked by volume and by how much damage a wrong answer would do
- 02Content auditSource inventory, one authoritative document per topic, gaps written up for somebody to fill
- 03Retrieval buildContent chunked, indexed and tested so the correct passage is actually retrieved before any answer is generated
- 04GuardrailsRefused topics, escalation triggers, tone rules and a standing instruction to decline rather than guess
- 05Evaluation setReal questions with approved answers, scored automatically on every change
- 06IntegrationLive chat handover, enquiry capture into your CRM and transcript logging
- 07PilotA limited audience or after hours only, with every conversation read by a human for the first few weeks
- 08WidenNew topics added only when the evaluation scores support it
Two decisions on your side that keep the project moving
The stage clients underrate is evaluation. Before launch we assemble a test set of real questions, including the awkward and adversarial ones, paired with the answer your best staff member would give. Every change to the prompts, the retrieval settings or the underlying content is scored against that set, so you can see whether a tweak improved things or simply relocated the failures. Without it you are relying on whoever poked at it most recently and came away feeling optimistic.
Choose the right level
Scripted bot, retrieval assistant or a person
Not every enquiry deserves a model. A scripted flow that looks up an order number and reads back a status is cheaper, faster and structurally incapable of inventing anything, and for a narrow task that makes it the better engineering decision. Most sites we work on end up running two of the three options below rather than choosing one, with the scripted paths catching the predictable questions and the assistant handling everything phrased in a way nobody anticipated.
Approach
01
Scripted flow
Suits
Order status, opening hours, sending a booking link
The honest limitation
Breaks the moment a customer phrases it differently
02
Retrieval grounded assistant
Suits
Service, policy and product questions with documented answers
The honest limitation
Only ever as accurate as the documents behind it
03
Assistant with live handover
Suits
Enquiries that become sales conversations or complaints
The honest limitation
Needs staffed hours and a queue somebody actually watches
04
People only
Suits
Clinical, legal and credit advice, distressed customers
The honest limitation
Costs more per contact and cannot absorb a sudden peak
How we work this out during scoping
Token cost is the line item that surprises people. Every conversation sends your retrieved content plus the running exchange to the model, so a support page backed by long policy documents costs meaningfully more per conversation than a lean one. At a few hundred conversations a month this is noise. At tens of thousands it becomes a genuine operating cost, and it changes real decisions: how much context you send, how aggressively you cache, and which questions should never reach a model at all.
Privacy, the Australian Privacy Principles and where the data goes
Every conversation with a customer facing assistant is a potential collection of personal information, and obligations under the Privacy Act 1988 do not pause because a model is doing the typing. APP 8 is the one that catches people out. Sending a transcript to a model hosted overseas is a cross border disclosure, and you remain accountable for what happens to it. That belongs in your privacy policy and in a short collection notice at the start of the chat, written in plain words rather than buried in a link nobody opens.
Practically, we reduce what is exposed
Practically, we reduce what is exposed. Fields that do not need to reach a model are stripped before the request, transcripts are stored in Australian infrastructure with a retention period you have chosen deliberately, and we check the provider terms covering whether your inputs can be used for training. Where Australian data residency is a hard requirement we use models served from Australian regions, and we tell you plainly what capability you give up in exchange. For medical practices and anyone handling health information that trade is usually worth making, and the surrounding controls sit with our security work.
When an AI chatbot is the wrong fit
If you receive twenty enquiries a week, an assistant will consume more attention than it saves, and a well written FAQ page plus a fast reply will beat it comfortably. If your answers genuinely depend on individual circumstances, as clinical, legal and credit advice do, a general assistant is a liability rather than an asset, and the right build is a triage tool that gathers information and routes it to a qualified human. If your documentation does not exist, fix that first and revisit the bot afterwards.
The customer does not want a conversation
It is also worth naming that a large share of projects arriving as chatbot requests are really automation requests. The customer does not want a conversation. They want to reschedule a delivery, download an invoice or change an appointment, and that is a form and an API call. Those are better served by workflow automation or a self-service area than by a model. When the underlying process is the problem, adding a chatbot in front of it just makes the process faster at being unhelpful, and we will say so during scoping.
How we scope it
Four ways to scope your AI Chatbots project
We do not publish package prices, because the same brief can be a short build or a long one. These are the shapes the work usually takes. Tell us which one sounds like you and you will get a fixed written quote that spells out exactly what it covers.
Proof of value
One use case, evaluated honestly before it goes near a customer
Fixed written quote, agreed before work starts
- Retrieval grounded assistant deployed on your site
- Content audit naming the authoritative source per topic
- Guardrail and refusal policy documented in writing
Production build
In production, with a human approval step and an evaluation set
Fixed written quote, agreed before work starts
- Everything in Proof of value
- Evaluation set of real questions with approved answers
- Escalation path and live handover to staff
- Enquiry capture flowing into your CRM
Embedded platform
Built into the product rather than bolted onto it
Fixed written quote, agreed before work starts
- Everything in Production build
- Transcript logging with a defined retention period
- Draft privacy collection notice wording
- Admin access to prompts, content and settings
AI Chatbots Model care
Monitoring, evaluation and retraining as the inputs drift
Rolling monthly, quoted in writing
- Evaluation set rerun as the model and the inputs change
- Cost and quality reported monthly, not assumed
- Prompt, tool and guardrail changes as the work shifts
- Rolling, cancel with 30 days notice
These are shapes, not menus. Most quotes end up somewhere between two of them, and we will say so when the honest answer is the smallest one. Describe the problem and we will tell you which it is.
Questions buyers usually ask
Frequently asked questions
Ownership and handover
Who owns the prompts, the content and the conversation data?
You do. The knowledge base, the prompt configuration, the evaluation set and every transcript belong to your business, and the model provider account is opened in your name with your billing details. We work with access you can revoke. If you move to another partner, nothing needs to be negotiated because none of it was ever held on our side.
What happens after launch?
Someone reviews transcripts, missed answers become new content, and the evaluation set grows. Most clients keep a small monthly block for that plus content updates. If you want the assistant to start completing tasks rather than answering questions, that is a different build and closer to AI agents, which we would scope separately rather than bolt on.
Detail and edge cases
How long before an AI chatbot is answering real customers?
Usually 4 to 9 weeks. Retrieval and guardrails are quick to build. What sets the pace is your content: how scattered it is, how much contradicts itself and how fast someone can write the answers that turn out to be missing. We shorten the wait by piloting on a narrow topic set while the rest of the content is being sorted out.
What drives the cost of an AI chatbot?
Four things: how many topics it must cover, how much content preparation is needed, which systems it connects to, and your conversation volume. Volume matters because you pay the model provider per token, so long documents and busy channels raise the running cost. We scope all four and send a fixed written quote for the build, with the ongoing platform costs shown separately as they are paid by you.
What stops the chatbot from making things up?
Three layers, and none of them is a promise of perfection. The assistant may only answer from retrieved passages of your approved content. Guardrails refuse topics you have ruled out and escalate on low-confidence. An evaluation set catches regressions before they reach customers. That combination makes invention rare rather than impossible, which is why review of real transcripts stays part of the running arrangement.
Can it answer questions for a regulated profession?
It can answer factual and administrative questions such as fees, locations, referral requirements and appointment processes. It should not give clinical, legal or financial advice, and we build refusals that say so. For an AHPRA regulated practice we also strip testimonial style content out of the knowledge base so the assistant cannot repeat something the advertising rules do not allow.
Related services
Get a fixed written quote for your AI chatbot
Send us the top twenty questions your team answers every week and tell us where the answers currently live. We reply within one business day with an honest read on whether this is worth building.