A good website chatbot is a narrow tool: it answers questions you have already settled, takes down the details of a visitor who wants a person, and hands over cleanly. A bad one tries to be an expert on everything.
This guide is about scope and safety. It does not cover cost, which depends on your build and tools.
What jobs suit a website chatbot?
- Settled questions: opening hours, delivery areas, how to book, what a service includes, what to bring.
- Finding the right page: pointing a visitor to the page, form or phone number they need.
- Taking inquiries: collecting a name, a contact detail and a short description, then passing them to your team.
- Booking and status: offering a booking link, or reading a booking or order status from your own system.
- Handoff: moving to a person, with the conversation, at the visitor's request or when the bot is unsure.
Public tool libraries show these patterns. Make's template library lists 'Create a basic chatbot for WhatsApp' and 'ChatGPT-Powered Telegram Bot for Instant Replies', and Zapier describes a customer-facing chatbot on its AI page (both read 12 November 2026). They show what the tools can build. They do not show that a chatbot will help your business, and we claim no result.
What should a chatbot not do?
- Give advice on legal, medical, financial or safety matters. Send those to a qualified person.
- Make promises about prices, discounts, delivery dates, refunds or results that you have not approved.
- Pretend to be a person. Say clearly that it is an automated assistant.
- Collect sensitive data it does not need, such as payment card numbers or health details, in a chat.
- Argue or stonewall. If a visitor is upset or asks twice for a person, hand over.
- Replace your contact page. Keep a phone number and a form where people can find them without the chat.
Why does the scope matter in law?
In February 2024 the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada. The airline's website chatbot told a customer he could claim a bereavement fare after travel, contradicting a page it linked to. The tribunal rejected Air Canada's argument that the chatbot was a separate entity and held it responsible for all the information on its website, awarding about $650. Torkin Manes' summary (7 March 2024) advises training AI systems for accuracy, internal policies and clear warnings about the limits of AI tools. It was a tribunal decision, not a court judgment, and not advice for your case.
In the United States, the FTC said in September 2024 that there is no AI exemption from the laws on the books. So everything your chatbot says about your business should be true, and something you could stand behind.
How do you set it up safely?
- Write the source text first. Collect your real answers to the visitor questions you get most, in plain words, and keep them current. The chatbot answers from that text.
- Tell it what it does not know. It should say so and offer a person, instead of guessing.
- Limit its access. Give it only the data it needs to do its job, and no ability to change records or make payments without a person.
- Label it as automated, say what it can help with, and link to your privacy information.
- Test it with hard questions, rude questions and attempts to make it say something it should not.
- Review it every week by reading a sample of conversations, and fix the source text when it errs.
What about security and privacy?
A chatbot reads whatever a visitor types, and visitors can try to steer it. OWASP's prompt injection page (LLM01:2025) says user prompts can alter a model's behavior in unintended ways, and lists mitigations: restrict the model's access privileges to the minimum necessary, use human-in-the-loop controls for privileged operations, and separate untrusted content. It says it is unclear that there are fool-proof methods of prevention. So a chatbot should not hold secrets or powers you would mind a stranger using.
On privacy, tell visitors what you collect and why, collect the minimum, and check the privacy rules that apply where your visitors live. If a chat tool passes conversations to another company, say so in your privacy information.
How do you know it is working?
Measure what visitors experience: how many conversations end with the question answered, how many are handed to a person, how many complaints mention the chat, how many inquiries it captured that led to a reply and a sale, and what visitors typed that the bot could not answer. Read transcripts. If complaints rise or the same unanswered question keeps appearing, change the source text or narrow the scope. Our guide to AI customer service automation covers the wider support picture.
Do you need one?
Not every site does. If visitors rarely ask questions the page does not already answer, a clearer page and a visible phone number may do more. A chatbot helps when many visitors ask the same settled questions outside office hours. Start with the questions, not the tool. Our AI chatbots page covers how we scope them, and our AI automation page covers connecting a chatbot to your other systems.
Conclusion
A website chatbot works when its job is small, its source text is yours and current, it says it is automated, and a person is one step away. It fails when it tries to advise, promise or improvise.
PromotEdge International is an international strategy, technology, marketing and design company, and our delivery team is in India. See our AI chatbots page.
Facts we rely on: see our company facts for what we state about ourselves and where each fact comes from.