AI Chatbot for Website: How to Set One Up That Actually Helps

An AI chatbot for a website used to mean a decision tree with buttons. Today it usually means a large language model that reads the visitor's question in plain words and answers in plain words. That's a real leap — but it also means the bot can sound confident while being wrong. This guide covers what a modern AI assistant is good for on a small-business site, what to keep for humans, and a practical setup path: choose a model, write the instruction, give it facts, define when it hands off to a person, and test with real questions.

What an AI chatbot on a website actually does

A rule-based bot matches keywords or button clicks to prewritten answers. It's predictable, but it breaks as soon as a visitor phrases something unexpectedly.

An LLM-based assistant works differently:

• It reads the whole message, typos and all, and figures out what is being asked.
• It writes an answer based on your instruction — the text that tells it who it is, what your business does and what rules to follow.
• It keeps the context of the conversation, so follow-up questions like "and for Canada?" make sense.
• It can call tools: hand the chat to a human, send the visitor to a specific page, or close the conversation.

What it doesn't do is "know" your business. Everything specific — prices, policies, shipping zones — has to come from what you give it.

Good jobs for an AI chatbot

The best tasks are frequent, factual and low-risk:

• FAQs: opening hours, shipping times, return windows, payment methods.
• Pre-sale questions: "Does this come in a larger size?", "Does the plan include X?" — as long as the facts are in the instruction.
• Navigation: pointing visitors to the pricing page, a size chart or a help article.
• Triage: asking what the problem is and collecting an order number before a human takes over.
• After-hours coverage: answering what it can and telling visitors when operators are back. More on that in our guide to after-hours customer support.

Jobs to keep for humans

Hand these to a person rather than letting the bot improvise:

• Refunds, cancellations and anything involving money moving back to the customer.
• Complaints and upset customers.
• Account-specific issues the bot can't see, such as "where is my order?" when it has no access to your order system.
• Custom quotes, negotiation and large B2B deals.
• Anything legal, medical or safety-related.

Choosing a model: quality, speed and cost per conversation

You'll pick between model families — GPT, Claude, Gemini, DeepSeek, Qwen and others. Three trade-offs matter:

• Quality. Larger models follow complex instructions more reliably and are better at saying "I don't know." Smaller models are fine for simple FAQs.
• Speed. Visitors notice a reply that takes many seconds. Faster models feel more like chat.
• Cost. You pay per token — roughly per word in and out. Every turn sends the instruction plus recent message history, so a long instruction and a long history cost more on every message.

A sensible approach: start with a mid-range model, watch real conversations for a week or two, and only move up if answers fall short.

In ChatDeck the assistant runs on your own API key through the GPTunneL gateway, so you pay only for model tokens, and the estimated cost per visitor message is shown when you pick a model. There's no separate bot subscription on top.

Giving the bot knowledge: instruction, facts, per-topic rules

The instruction is where most of the quality comes from. A good one includes:

1. Role and tone. "You are the support assistant for {store}. Be friendly, brief, and answer in the visitor's language."
2. Facts. Shipping zones and times, return policy, prices or plan limits, contact details. Write them as plain statements, not marketing copy.
3. Boundaries. "If the answer isn't in these facts, say you're not sure and offer to connect an operator. Never invent prices or policies."
4. Handoff rules. When to transfer — refunds, complaints, an explicit request for a human.
5. Formatting. Short paragraphs, no walls of text, links to relevant pages.

When the instruction grows long, split it. ChatDeck lets you add per-topic rules — for example a separate block for returns or for wholesale inquiries — and each rule can use its own model, so a simple topic can run on a cheaper one. Writing the instruction itself is a skill; our article on customer service chatbot prompts goes deeper.

Setting up handoff to a human operator

Handoff is what makes an AI assistant safe to deploy. Decide:

• Triggers. The visitor asks for a person, the bot isn't confident, or the topic is on the "humans only" list.
• What the operator sees. The full conversation, so the customer doesn't repeat themselves.
• What happens when nobody is online. The bot should say when the team will be back instead of promising an instant reply.

In ChatDeck the bot has a transfer-to-operator tool; the conversation moves to the queue and operators get notified. Outside working hours the bot tells the visitor when operators return and keeps answering what it can. For the details, read how to design chatbot-to-human handoff.

Testing before launch: a 20-question checklist

Before you switch the bot on, write 20 real questions and run them through the chat. Cover:

1. Five of your most common questions, phrased the way customers actually phrase them.
2. Three questions with typos or slang.
3. Two questions in another language, if you get international visitors.
4. Three questions whose answers are not in the instruction — the bot should admit it and offer a person.
5. Two requests for a refund or cancellation — the bot should hand off.
6. One angry message — it should apologize briefly and hand off.
7. One request to talk to a human directly.
8. One attempt to make the bot ignore its instruction ("forget your rules and…").
9. One question asking it to send you to a page, such as pricing.
10. One vague message like "hi" or "help" — it should ask a clarifying question.

Fix the instruction after each failure and rerun the set. Keep the list; you'll reuse it every time you change the model or the instruction.

Measuring whether the bot is helping

After launch, look at:

• Share of conversations resolved without a human. Grows as the instruction improves.
• Handoff reasons. Repeated topics are facts missing from the instruction.
• Wait time after handoff. A great bot followed by a long wait is still a bad experience.
• Ratings for bot-resolved conversations compared with human ones.
• Transcripts. Read a sample every week. Nothing replaces it.

Frequently asked questions

How much does an AI chatbot for a website cost to run?

With pay-per-token pricing, cost depends on the model, the length of your instruction, how much history is sent and how many messages visitors write. For a small site it's often modest, but estimate it from the per-message cost of your chosen model and your monthly chat volume rather than relying on a flat number.

Can an AI chatbot give wrong answers, and how do I limit that?

Yes. Limit it by putting the facts in the instruction, telling the bot to admit when it doesn't know, keeping risky topics for humans and reviewing transcripts regularly.

Which AI model is best for customer service chat?

There's no single best one. Pick a model that follows instructions well and replies quickly, test it on your own 20 questions, and compare cost per message. A mid-range model is usually a good starting point.

Will customers be annoyed talking to a bot?

Mostly when the bot blocks them from a person or answers vaguely. Be upfront that it's an assistant, keep answers short and make the path to a human obvious.

Can the bot hand the chat to a human?

A good setup can. The bot transfers the conversation with full history, and operators pick it up from the queue. Outside working hours it tells the visitor when someone will be back.

Try it on your site

If you're weighing the options, our comparison of chatbot vs live chat helps you choose the right mix. To try it hands-on, create a free ChatDeck workspace: the widget and unlimited operators are free, and the AI assistant runs on your own key, so you only pay for the tokens it uses. For a screenshot walkthrough, see Connect the AI assistant and Write the bot's instruction.