An AI chatbot for your website answers visitor questions from your own content, captures leads, and hands the hard conversations to a person. It works when it is built around one clear job, fed accurate content, and wired into the tools you already use. We build these as custom AI integrations, and this guide shows what to decide before anyone writes a line of code.
TL;DR
- A website AI chatbot reads your pages, FAQs and policies, then answers in plain language at any hour.
- Give it one main job first: answer questions, qualify leads or book appointments.
- Always add a human handoff, and an email fallback for when nobody is online.
- Tell visitors they are talking to AI. Several US states are moving to require it.
- Chatbot, live chat or both? Most businesses end up with both: bot first, person second.
What an AI chatbot on a website actually does
A scripted bot follows a fixed menu. An AI chatbot reads what the visitor typed, works out the meaning, and answers in everyday words. It copes with ten ways of asking the same question.
In practice, the jobs fall into a short list:
- Answer common questions about hours, services, pricing pages and policies.
- Ask a few qualifying questions and collect a name and email.
- Book a call or an appointment on your calendar.
- Point visitors to the right page.
- Pass tricky issues to a person, with the chat history attached.
Notice what is missing: judgment calls. Quoting a custom job, calming an angry customer and negotiating all belong to people. A good chatbot knows where its lane ends.
How it works behind the widget
The chat bubble is the small part. Behind it sit your content, a retrieval step, a language model and your business tools.
Most production chatbots use a method called retrieval-augmented generation, or RAG. Your content is split into chunks and stored so it can be searched. When a visitor asks something, the system pulls the few most relevant passages and gives them to the language model, which writes an answer from that text. One technical walkthrough of RAG-based chatbots notes that changing a price or policy in the knowledge base takes effect without retraining the model.
That is why your content matters more than the model. If your pricing page is out of date, the bot will repeat the old price politely and confidently.
The answers should also sound like a person wrote them. If you want to tune the tone, it helps to know how to humanize AI text with clear instructions in the prompt you give the bot.
Set one job, then build around it
Businesses that try to do everything on day one usually get a bot that does nothing well. Pick one primary job.
| Business | First job for the bot |
|---|---|
| Home service company | Collect job details and book a visit |
| Realtor | Qualify buyers and sellers, book showings |
| Property manager | Answer tenant questions, route maintenance requests |
| B2B distributor | Help dealers find products and order status |
| SaaS founder | Answer product questions, book demos |
Once that job works, add the next one. The order matters because each new job brings new content to check and new tools to connect.
A practical setup path
- List your top 20 questions. Pull them from your inbox, call notes and social messages. These are the bot's real curriculum.
- Fix the source content. Write or update an FAQ page. Remove contradictions and old prices. The bot is only as accurate as this material.
- Define success. Choose two or three numbers for the first 90 days, such as questions resolved without a person, leads captured, or appointments booked.
- Write the rules. Decide the bot's scope, tone and the point where it must hand off. Include an "I don't know" answer so it never guesses.
- Connect your tools. Link the calendar, the CRM and your email. Start with read-only access, then allow the bot to write data once it behaves.
- Test with real messages. Run 10 to 15 actual customer questions through it. Check the widget on mobile and in a private browser window. Confirm that test leads land where your team will see them.
- Launch to part of your traffic. Start with a slice of visitors, read the conversations weekly, and fix the content gaps they reveal.
The tool connections in step 5 are where most projects stall. A chatbot that cannot book into your calendar or create a CRM record is just a nicer FAQ. This is the work we do: custom AI integration into the software you already run.
Chatbot or live chat?
The choice is less either/or than it looks. A comparison of live chat, rule-based bots and AI agents lays out the trade-offs:
| Live chat | Rule-based bot | AI chatbot | |
|---|---|---|---|
| Who answers | A person | A fixed script | A language model on your content |
| Typical billing | Per seat | Varies | Per resolution, conversation or credit |
| Main risk | No answer outside staffed hours | Visitors stuck in a loop | Wrong or invented answers |
| Best for | Varied, high-value questions in staffed hours | Repeated, structured tasks | High volume of questions already answered in your content |
The same comparison points to a combined setup as the one most vendors now sell: the bot answers first and hands over to a person when it cannot help.
For a small team, that usually means this. The bot handles the repeatable questions at 2 a.m. A person takes over when the visitor is valuable, upset or asking something messy. Pass the whole conversation along, so the visitor never repeats themselves.
What happens when nobody is online
A handoff only works if someone is there to take it. Outside business hours, the bot should say so and offer a form or an email address. Check this early, because it is the moment visitors judge you.
Customer support use: where the line sits
An AI support chatbot is strong on routine questions: order status, hours, how-to steps, policies. It is weak on exceptions. Set confidence thresholds so that low-certainty answers go to a person instead of to the visitor. Our guide on integrating and governing support automation covers how to keep people in charge of it.
If your visitors call rather than type, an AI phone agent follows the same principles by voice.
Privacy and disclosure rules to know
This is not legal advice, so have a lawyer review your final setup. Here is the landscape for US businesses.
- Tell people it is AI. A 2025 review of state AI bills found that six of the seven chatbot bills it tracked require a clear and conspicuous notice that the chatbot is not human.
- Some rules already apply. According to a law firm's analysis of chatbot risks, Utah requires disclosure when a user clearly asks whether they are talking to AI. California's BOTS Act covers platforms with at least 10 million unique monthly US visitors and bars deceiving people about a bot's identity to prompt a purchase.
- California's automated decision rules start in 2027. The same analysis says the CCPA regulations on automated decision-making technology begin January 1, 2027. They require notice and opt-out rights when a chatbot qualifies.
Practical habits cover most of it. Label the bot as AI in its first message. Mention the chat tool and data use in your privacy policy. Never ask for card numbers or passwords in chat. Limit who can read transcripts with personal details.
Serving Canadian visitors too? We cover that in PIPEDA compliance for AI agents.
What it costs
Prices depend on who builds it and how it is billed. Off-the-shelf tools charge per seat, per resolution or per conversation, so your bill grows with traffic. A custom build is a project fee, scoped to your tools and content. We do not publish a price, because the scope sets it. For a Canadian view of ranges, see what an AI agent costs in 2026.
Our view
Most chatbot projects fail for boring reasons. The content is stale. The bot has no job. Nobody reads the conversations after launch.
We start with the questions your customers already ask and the tools your team already uses. Then we build the smallest bot that does one job well, and we keep you in the loop from idea to launch.
A chatbot should take repetitive work off your team, not add a new thing to babysit. That is why we wire it into your calendar, your inbox and your CRM, and why a person always sits behind it.
Loic
Want a chatbot built around your business?
We design and build AI integrations and workflow automation for business owners in Canada and the US. We work with you from first scoping call through launch, and we explain each decision in plain words. See our AI and automation service or look at the full range on our services page.
FAQ
Is an AI chatbot better than live chat?
They do different jobs. The bot answers repeatable questions at any hour, and live chat gives you judgment and trust on valuable or messy requests. Most businesses get the best result by running the bot first and handing over to a person when needed.
Do I have to tell visitors it is a bot?
Yes, do it as a habit. Some states already require disclosure on request, and more are moving toward a clear notice. A short first message that says it is an AI assistant covers you and sets honest expectations.
Can a chatbot give wrong answers?
It can. Language models sometimes write fluent answers that are incorrect, especially when the source content is old. Keep your content current, give the bot an "I don't know" fallback, and read conversations weekly.
How long does it take to launch?
The code is the short part. Cleaning up your content, mapping the top questions and testing handoffs take most of the time. A narrow first job gets you live sooner.
Sources
- Live chat vs chatbot vs AI agent: what's the difference?
- The State of State AI 2025
- AI Chatbots: How to Address Five Key Legal Risks
- How to Build a Chatbot for Your Website and Train It



