What is a custom AI chatbot trained on your own content?
A custom AI chatbot is a chat window on your website that answers visitors by looking up your own approved material and then using a language model to write a reply from it. It is not a model retrained from scratch; it is a search system and a writer working together.
The technical name is retrieval-augmented generation, usually shortened to RAG. Your content is split into passages and stored in a searchable index. When a visitor asks a question, the system finds the few passages most likely to contain the answer and hands them to the model with strict instructions: answer only from these passages, cite them, and say "I don't know" when they do not cover the question.
That design matters for a Canadian business for a simple reason. A general chatbot answers from everything it absorbed during training, which includes other companies' prices, other provinces' rules and outdated information. A grounded bot answers from your return policy, your service areas and your hours, and nothing else.
When people search for an AI chatbot development company in Canada, most want exactly this: a bot that sounds like the business, knows the business and does not make things up. The rest of this guide explains how to get there, what it costs and where the risks sit.
Does your business actually need an AI chatbot?
You need one when visitors ask the same questions repeatedly, arrive outside business hours, or leave without contacting you because finding the answer takes too long. You probably do not need one if you get a handful of enquiries a week and answer them all personally.
Useful decision rules we apply on discovery calls:
- Choose a chatbot when staff answer the same ten questions every day, or when enquiries peak in evenings and weekends.
- Choose a better FAQ page first when the questions are few and stable. It is cheaper and also helps search rankings.
- Choose a phone assistant instead when your customers call rather than type, as many clinic and trades customers do. See our AI receptionist page.
- Choose a person when every enquiry is complex and high-value, such as bespoke commercial projects. A bot can still collect details before the call.
If you are unsure, ask us to review a month of your enquiries. The pattern usually makes the answer obvious.
Training on your content: retrieval versus fine-tuning
For almost every Canadian SMB website, retrieval is the right approach and fine-tuning is unnecessary. Retrieval keeps answers tied to documents you can edit; fine-tuning bakes knowledge into a model that is harder to update and audit.
Retrieval (what we build by default)
Your content lives in an index. Update a policy page and the bot's answers change the same day. Every reply can show the passage it came from, which makes checking and correcting easy.
Fine-tuning (rarely needed)
A model is further trained on examples to adopt a style or format. Useful for very specific output formats, but it does not reliably teach facts, and changing a price means retraining.
Prompt rules (always needed)
A written instruction set defines tone, languages, forbidden topics and when to hand off. It is short, versioned and reviewed with you like any other business document.
This is also why ownership of the knowledge base matters. The index and the source documents should sit in your accounts, so you can move to a different model provider later without starting again.
What content should a chatbot be trained on?
Feed the bot only what you would happily have a new employee quote to a customer: current service pages, policies, FAQs, hours, service areas and product documentation. Leave out internal notes, draft pricing and anything you have not reviewed recently.
The biggest source of wrong answers is not the model; it is contradictory content. If your website says one return window and a PDF from 2021 says another, the bot will sometimes pick the old one. Part of every build is a content audit where we list conflicts and ask you to resolve them before indexing.
A practical content set for a service business looks like this:
- Service and product pages from your live website.
- A written FAQ of the questions staff answer most, in your words.
- Policies: cancellations, refunds, warranties, privacy.
- Service areas, hours, holiday closures and contact routes.
- A short "things we don't do" list so the bot can say no politely.
- French versions of all of the above, supplied or approved by you or your translator.
If your website content itself is thin, fixing it improves both the bot and your search visibility. Our website redesign guide covers that route.
Building a bilingual English/French AI chatbot for Canada
A bilingual chatbot should detect the visitor's language, search content in that language first and answer in it, rather than translating English answers on the fly. That keeps terminology consistent with your French website and avoids awkward phrasing.
We work in English. For French, you or your translator supply or approve the French source content and the fixed phrases the bot uses: greetings, hand-off messages, privacy notice, refusal lines. The model then writes answers from French passages. We test with French-speaking reviewers you nominate, using a list of real questions in both languages.
There are three practical design points. The language switch must be obvious, because some visitors in Montreal or Moncton will start in one language and move to the other. Links in answers should point to the French version of a page when the conversation is in French. And lead notifications to staff should record which language the visitor used, so the follow-up call starts in the right one.
For businesses in Quebec, the Charter of the French language shapes how customer-facing content is presented. We build the technical side; your advisor confirms what applies. Our pages on Bill 96 website requirements and web development for Montreal businesses go further.
Chatbot lead capture into your CRM
A chatbot earns its cost when it turns anonymous visitors into named leads with clear needs. The trick is asking for contact details at the right moment, after the bot has been useful, not as a gate before the first answer.
We design the capture flow around the questions your best salesperson asks. For a renovation firm that might be project type, rough timeline, postal code and budget band. For a law office it might be practice area and urgency. The bot asks these conversationally, then offers a callback or booking.
On submission, the chatbot creates or updates a contact in your CRM, attaches the full transcript, tags the lead source as website chat, and creates a follow-up task for the right person. If your CRM is HubSpot, Zoho, Pipedrive or Salesforce, we use its API. If you run a spreadsheet, we write to that, and suggest when it is time to move up.
Consent is captured at the same moment. The bot asks whether the visitor wants marketing emails separately from the service reply, and stores the answer with a timestamp, so your follow-up sequences stay on the right side of Canada's anti-spam rules. For the automation that happens after capture, see AI automation for Canadian SMBs.
How human hand-off should work
Every business chatbot needs a clear exit to a person, triggered by the visitor asking, by low confidence in an answer, or by a restricted topic coming up. A bot that traps people in loops does more damage than having no bot.
We set up hand-off in one of four ways depending on your staffing: a live chat inbox for teams who can answer in real time; a WhatsApp message to the on-duty person; an email with the transcript; or a CRM task with a promised callback window. Outside hours, the bot says honestly when someone will reply instead of pretending a human is available.
The hand-off message carries context so the visitor never repeats themselves. Your staff see the language, the questions asked, the passages the bot used and why it handed off. That last part is useful for improving the bot: if the same topic keeps triggering hand-offs, the fix is usually a new FAQ entry.
We also recommend a weekly review of a sample of conversations for the first two months. It takes twenty minutes and catches problems early.
Guardrails: stopping wrong answers about pricing and legal advice
Guardrails are the rules and tests that stop a chatbot from promising prices, outcomes or advice your business would not stand behind. In Canada this is not theoretical: in Moffatt v. Air Canada (2024 BCCRT 149), British Columbia's Civil Resolution Tribunal held the airline responsible for incorrect refund information its website chatbot gave a customer, rejecting the argument that the bot was responsible for its own statements.
The lesson for any business is that visitors will treat what your chatbot says as what you said. So we design guardrails in layers:
- Source limits. The bot may only answer from indexed content and must cite it.
- Restricted topics. Prices, discounts, refunds outside policy, legal, immigration, medical and tax advice trigger a fixed response and a hand-off.
- Wording rules. No guarantees, no "definitely", no outcomes promised on your behalf.
- Test set. Before launch, the bot answers a list of tricky questions, including attempts to talk it into discounts, and you sign off the results.
- Logging. Conversations are stored in your account so you can see exactly what was said.
Regulated professions need stricter lists. An immigration consultant's bot, for example, should never assess eligibility; see our guide to immigration consultant websites for why.
Privacy disclosures for chatbots under PIPEDA and Law 25
A chatbot collects personal information the moment a visitor types a name, an email or a description of their problem, so it needs a privacy notice at the start of the chat and a clear policy behind it. Your organisation remains accountable for that data even when an outside provider processes it.
The Office of the Privacy Commissioner's guidance on cross-border processing says PIPEDA allows transfers to providers in other countries, but organisations must protect the information by contract and tell people it may be processed abroad. In December 2023, Canada's privacy commissioners jointly published principles for generative AI that stress openness about AI use and limiting collection to what is needed.
In practice our chatbot builds include: a short notice in the chat window that the visitor is talking to an AI assistant and how their messages are used; a link to your privacy policy; an instruction that tells visitors not to share sensitive details such as health card or passport numbers; retention settings for transcripts; and a list of processors (model provider, hosting, CRM) for your policy.
If you serve Quebec residents, Law 25 adds stricter consent and transparency expectations. We build the mechanisms; your counsel decides the wording and confirms compliance. For the rest of the site, see Law 25 website compliance.
How much does AI chatbot development cost in Canada?
Quotes vary widely between chatbot subscriptions, Canadian development companies and remote teams, mostly because "chatbot" covers everything from a canned FAQ widget to a multi-system assistant. With BtechWaleTech, a grounded, bilingual chatbot with lead capture starts at US$600.
Think of the total cost in three parts. The build: indexing content, prompt rules, integrations, guardrail testing. The running cost: language model usage per conversation and hosting for the index, both billed to your account. The care: updating content, reviewing conversations, adjusting rules, which is free for two months and then starts at US$120/mo.
What pushes the build price up: hundreds of PDFs or scanned documents to clean, several integrations such as booking plus CRM plus help desk, strict industries that need a long test set, and a custom dashboard for staff to review and annotate conversations.
What keeps it down: a tidy website, a written FAQ, one CRM, and a clear list of what the bot must never discuss. Doing that homework before the first call can take a meaningful share off the quote. See our starting prices for everything else we build.
Choosing an AI chatbot development company in Canada: a vetting test
The fastest way to judge any AI chatbot development company in Canada, or any freelancer, is to ask them to show you how their bot fails. Everyone demos the happy path; the useful information is in what happens with awkward questions.
Put these questions to every candidate, and ask for the answers in writing:
- What does the bot do when the answer is not in our content?
- Show a test where a visitor tries to get a discount the policy does not offer.
- How is French handled: translated on the fly or answered from French content?
- Where are transcripts stored, for how long, and who can read them?
- Whose account holds the model API key and the knowledge base?
- What does a month of running costs look like at our visitor numbers?
- How do we update an answer ourselves when a policy changes?
Walk away from anyone who says their bot "never makes mistakes", who keeps the knowledge base on their own platform with no export, or who cannot tell you which model provider handles your visitors' messages.
Pick a widget platform when your needs are standard and English-only; pick a custom build when you need French answered from French content, specific CRM behaviour, strict refusal rules or ownership of everything.
Widget platforms are quick. You paste a script, point it at your site, and it starts answering. Their limits show up later: French often comes from automatic translation, hand-off options are fixed, per-conversation pricing climbs with traffic, and moving away means rebuilding.
A custom build takes two to four weeks and gives you control over every behaviour. It runs as a small service in your own cloud account, with the chat widget embedded on your site. You choose the language model, and can switch providers later because the index and prompts are yours.
There is also a middle path we sometimes recommend: a reputable help-desk product with an AI add-on, configured properly, with our work limited to content cleanup, rules and CRM wiring. It is cheaper than custom and better than a raw widget. We will say which route fits after seeing your enquiries.
Whichever route you choose, the website underneath matters. A slow or confusing site sends visitors to the chat for things a clear page should answer. Our WordPress website design page covers sites built for easy editing.
How we test a chatbot before it talks to your customers
We test against a written list of real questions and adversarial prompts, and you approve the results before launch. Testing is where a chatbot becomes trustworthy, and it is the part cheap builds skip.
The test set is built with you from three sources: questions your staff actually receive, questions pulled from your emails and reviews, and deliberately difficult ones. The difficult ones include pushing for a discount, asking for legal or medical opinions, asking about a competitor, switching language mid-conversation, and trying to get the bot to ignore its instructions.
Each answer is scored as correct, acceptable, wrong or should-have-handed-off. We fix wrong answers at the source: missing content, conflicting documents or an unclear rule. Then the whole set runs again, because fixing one answer can shift another.
After launch, the test set becomes a regression check. Whenever content changes or the model provider updates a model, we run it again. That is part of the two months of free maintenance, and part of the care plan afterwards.
How a chatbot project runs between Canada and India
You talk to us in your morning; we build during your night. India is 9.5 hours ahead of Eastern time in summer and 10.5 in winter, so a 9 a.m. Toronto call is 6:30 p.m. for us, and the adjustments you ask for are usually ready to test when your next day starts.
Week one is content and rules: we collect your pages and documents, flag contradictions, and draft the refusal list and hand-off routes with you. Week two produces a working bot on a private test page, in both languages, with a shared spreadsheet where you mark answers right or wrong. Weeks three and four cover integrations, the guardrail test set and launch.
Communication runs on WhatsApp seven days a week plus a scheduled video call each week. Quotes are in USD; you pay by Wise, bank wire or PayPal, and Wise lets you pay from CAD. The written quote sets milestones and ownership, and nothing is billed until you approve it. For more on remote engagement, see outsourcing web development to India.
We do not visit sites or run in-person training. Staff training is a recorded screen-share session they can rewatch.
Worked example: a hypothetical bilingual physiotherapy clinic in Ottawa
Say a two-location physiotherapy clinic in Ottawa gets a steady stream of evening questions about direct billing, whether a referral is needed, parking and which therapist treats sports injuries, in both English and French. Front-desk staff answer the same questions every morning. This is an illustrative scenario, not a client.
The scope would be a bot grounded in the clinic's service pages, an insurance FAQ, location pages and therapist bios, all in both languages. Restricted topics would include diagnosis, treatment recommendations and exact out-of-pocket costs, each answered with a polite fixed line and an offer to book or call.
The bot would link to the clinic's online booking tool rather than booking inside the chat, keeping the integration light. Visitors who want a callback would leave a name, phone number and preferred language, creating a task for the front desk each morning. The chat notice would tell visitors not to share health details in the chat window.
A realistic plan: three weeks to launch, a starting quote from US$600, and a two-month review period where the clinic manager checks twenty conversations a week. For the clinic's website itself, see physiotherapy clinic website design.
Keeping an AI chatbot accurate after launch
A chatbot stays accurate only if its content stays current, so plan for a monthly half hour of review and an easy way to update source documents. Most accuracy problems after launch come from a policy changing on paper but not in the knowledge base.
We give you a simple admin route to update content: edit the website page or replace a document in a shared folder, and the index refreshes on a schedule. You also get a monthly summary listing the top questions, questions the bot could not answer, hand-offs by topic and conversations that ended with a lead.
That summary is a free content plan. Questions the bot could not answer are often questions your website should answer, which helps search rankings and AI search visibility too. Nobody can guarantee rankings, but clear answers on your pages are what search engines and AI assistants quote. Our local SEO services pick up from there.
The first two months of maintenance are free. After that, care starts at US$120/mo and covers model updates, re-running the test set, content refreshes and small rule changes.