How to Add a Chatbot to Your Website Without Embarrassing Yourself: 7 Steps

The short answer

To add a chatbot to your website well, decide what it is allowed to answer, give it your real content to answer from, define the exact moments it must hand off to a person, test it with a hundred real questions before launch, and tell visitors it is automated. The widget itself takes an afternoon. The seven steps below are about the other 90% of the work, which is the part that decides whether the bot helps customers or hands your competitors a story.

Before you add a chatbot to your website, decide what it is for

Chatbots fail when they are added because everyone else has one. They succeed when they are assigned a job. Write one sentence: "This assistant answers questions about our services, hours, pricing ranges and process, books consultations, and hands anything else to a person within business hours." If you cannot write the sentence, you are not ready for the widget.

Three jobs suit a small business website:

• Answer the questions your inbox already gets. Hours, location, parking, what you do and do not do, how the process works, rough price ranges, what to bring.
• Capture and qualify enquiries after hours. Take the details, ask two useful questions, book a slot or promise a callback time. Our guide to AI lead qualification covers the questions worth asking.
• Help existing customers with the routine. Order status, rebooking, invoice copies, where to send a document. This needs a connection to your systems and a way to confirm who the customer is.

Virtual agents and chatbots were the third most common AI application among Canadian businesses using AI in the second quarter of 2026, used by 28.2% of them, according to Statistics Canada. They are mainstream. Doing them well is not.

Step 1: choose what it may answer, and what it must not

Draw two lists. The "may answer" list is everything on your website plus any internal documents you are willing to publish, because anything the bot can say, a visitor can screenshot. The "must not" list includes legal, medical or financial advice specific to the visitor, anything about a complaint, refunds outside your written policy, promises about delivery dates or outcomes, and any question it cannot answer from your content.

The reason this matters was spelled out by a BC tribunal in early 2024. In Moffatt v. Air Canada, the airline's chatbot told a customer he could apply for a bereavement fare after travelling; the airline's policy said the opposite. The tribunal rejected the argument that the chatbot was, in the member's words, "a separate legal entity that is responsible for its own actions" and awarded $650.88 in damages plus interest and fees. The amount was small; the principle was not. What your bot says, you said. We draw out the lessons from the Air Canada ruling in a separate post.

Step 2: give it the data, not a personality

A chatbot that answers from your actual pages, policies and price list is useful. One that answers from the model's general knowledge is a liability. The technique is retrieval: the bot looks up the relevant passage in your content and answers from it, ideally with a link. Our explainer on RAG, or how AI answers from your documents, covers how that works.

What to prepare:

• Your website content, cleaned of out-of-date pages.
• A plain-language FAQ written from real emails and calls, including the awkward ones ("do you take insurance", "why is it more expensive than X").
• Policies as they are actually applied: cancellations, refunds, deposits, warranty.
• Hours, holidays, service area, contact routes, and who covers what.
• For customer tasks: a read-only connection to the booking or order system, and a rule for verifying identity before any account detail is shown.

Mark the date on every document. Set a monthly reminder to review the set. Stale content is the most common cause of confident wrong answers.

Step 3: define the handoffs

The handoff list is the bot's most important feature. Write it as triggers and destinations:

Trigger	What the bot does
Visitor asks for a person	Hands off immediately. Never asks "are you sure" or one more question.
Words like complaint, refund, cancel, lawyer, injury, urgent	Collects name and contact, states when a person will reply, creates a ticket flagged high.
Two consecutive answers rated unhelpful, or two "I don't know"s	Offers a person or a callback, and logs the question for the FAQ.
Question needs account or personal data	Verifies identity through the proper channel or routes to staff. Never guesses.
Outside business hours	Says so. Takes the message, promises a time, sends a confirmation email.

Decide where handoffs land: a shared inbox, a CRM task, a text to the owner's phone, or live chat staffed in business hours. A hybrid of bot and staffed chat often pays for itself in business hours. If a handoff lands nowhere, you have built a complaint generator.

Step 4: pick the tool, or build on your stack

Three paths, in rising order of control:

1. The chat module inside a tool you already pay for (your helpdesk, booking platform or e-commerce platform). Cheapest, limited to that tool's data.
2. A dedicated AI support product. Intercom's Fin, for example, lists US$0.99 per resolved conversation plus US$29 per helpdesk seat a month as of October 2026, with a 50-outcome monthly minimum if you run it on another helpdesk. Outcome pricing is attractive because you pay for answers, not seats, but check what counts as an outcome.
3. A custom assistant on your own data and systems. Right when the bot must read your inventory, bookings or customer records, or when the rules are yours. Our AI agent development work usually starts here; our cost guide for AI chatbots compares all three tiers.

Step 5: test it like a hostile customer

Before launch, collect 100 real questions from your inbox, call notes and reviews. Add 20 you hope nobody asks. Run them all and grade each answer on three points: correct against your source, says "I don't know" when it should, hands off when it should. Then try to break it: ask it to ignore its instructions, ask for a discount it was not given, ask about a competitor, type in French, paste a long angry message. Fix the content gaps and the handoff rules, then run the set again. Keep the set; it becomes your monthly regression test when you change the content.

Step 6: disclose, and handle data properly

Say it is automated in the first message, and keep that label visible. Several jurisdictions now require this, and a visitor who was told up front rarely holds a wrong answer against you the way a visitor who was fooled does; see which chatbot disclosure laws apply. Collect only what the job needs. If the bot gathers names and emails for follow-up, remember that marketing messages fall under Canada's anti-spam law: under CASL, an enquiry gives implied consent for six months, every commercial message must identify you and include a working unsubscribe, and unsubscribe requests must be honoured within ten business days. Tell visitors what you keep and for how long. This is general information, not legal advice.

Step 7: launch small and watch the transcripts

Put the widget on two or three pages first, not the whole site. Read every transcript for the first two weeks. You will find questions you never thought to answer, a phrasing that confuses the bot, and a page on your site that contradicts another. Fix those, then widen. Track three numbers: share of conversations resolved without a person, share handed off, and share where the visitor gave up. The last one is the one to watch; the AI chatbot glossary entry lists the common metrics.

Key takeaways

• Assign the bot a job in one sentence, and give it your real content to answer from.
• Write the handoff triggers before launch and make sure each one lands with a person.
• Test with 100 real questions, then try to break it. Keep the set for retesting.
• Disclose that it is automated, collect only what you need, and read the transcripts.

Frequently asked questions

How long does it take to add a chatbot to a website?

Installing a widget takes an hour. Preparing clean content, writing the FAQ and policies, defining handoffs and testing properly takes one to three weeks for a small business, longer if the bot needs to connect to booking or order systems. Launching before that work is done is how chatbots earn their bad reputation.

What should a website chatbot never answer?

Advice specific to the visitor's legal, medical or financial situation, anything about an active complaint, refunds outside your written policy, promises about dates or outcomes, and any question it cannot answer from your approved content. In those cases it should say so and hand off to a person with a stated reply time.

Do I have to tell visitors they are talking to a bot?

In several jurisdictions, yes, and everywhere it is the sensible choice. A one-line disclosure in the first message costs nothing and protects you when the bot gets something wrong. Our post on chatbot disclosure laws covers which rules apply where. This is general information, not legal advice.

Sources

1. Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026 — Statistics Canada, 11 June 2026
2. Airline Ordered To Compensate A B.C. Man Because Its Chatbot Provided Inaccurate Information — Dentons via Mondaq, 23 February 2024 (on Moffatt v. Air Canada, 2024 BCCRT 149)
3. Fin Pricing — Intercom, accessed October 2026
4. An Act to promote the efficiency and adaptability of the Canadian economy (S.C. 2010, c. 23), full text — Justice Laws Website, Government of Canada

How Flowmatix can help

We put this into practice for businesses in Toronto and around the world through our AI Agents & AI Solutions, ai receptionist: chat & phone services. Want to know what it would take for your business? Get a personalised quote in about 3 minutes.