An AI chatbot works on a website when it answers common questions immediately, stays inside what the business actually knows, says plainly that it is an assistant, and hands the conversation to a real person in the same window. It annoys people when it blocks the way to a human, guesses, or repeats itself.
We built a website chat called Ask Tamou for a criminal defense firm. A visitor types a question and gets an immediate, careful answer from an assistant that knows the firm's practice areas. When someone at the firm wants to step in, they reply from a messaging app on their phone, and the visitor sees that reply in the same chat window.
It replaced a third-party chat widget and its monthly bill. Building it taught us that the model is the smaller half of the job. The larger half is the moment a person takes over.
What makes an AI chatbot useful on a website?
An AI chatbot is useful on a website when it gives a visitor a correct, short answer faster than they could find it on the page, then makes the next step easy. For most businesses the next step is talking to a person, so the chat should lead there and never stand in the way.
Visitors arrive with a question and limited patience. A chat that answers the first question well earns a second one. A chat that opens with a form, or replies with three paragraphs of hedging, gets closed.
Speed matters too. Our chat streams its reply as it is written, so the visitor sees words appear right away and is not left watching a spinner.
Why is the human handoff the most important feature?
The human handoff is the most important feature because the chat exists to start a conversation with the business, and a model cannot finish that conversation. When a person can step into the same thread without the visitor repeating anything, the chat becomes a front door instead of a wall.
Ask Tamou handles this through a messaging app on the staff member's phone. The moment a person at the firm wants to step in, they type a reply there, and it appears in the visitor's chat window as part of the same thread.
Two things we would aim for in any handoff:
- No new tool for staff. People answer from a phone they already carry, not from a dashboard they have to remember to open.
- Full context. Whoever steps in should be able to read the exchange so far, so the visitor never has to start over.
If you build only one thing beyond the basic chat, build this.
What should an AI chatbot do and not do?
An AI chatbot should answer from the business's own information, admit when it does not know, identify itself as an assistant, and offer a person early. It should not invent details, quote prices or promises it was never given, pretend to be human, or trap visitors in a loop.
| Do | Do not |
|---|---|
| Say it is an AI assistant at the start | Give it a human name and photo and let people assume |
| Answer from the business's own pages and facts | Let it improvise about services, prices, or outcomes |
| Keep replies short, then ask a useful follow-up | Send long generic essays |
| Offer a person at any point | Hide the human option behind several failed answers |
| Say "I do not know, let me get someone" | Guess to keep the conversation going |
| Pass the full conversation to staff | Make the visitor repeat everything |
| Stay on the topics the business handles | Answer anything a visitor types |
| Tell visitors what happens with their messages | Collect sensitive details without saying why |
What annoys people about website chatbots?
People are annoyed by chatbots that pop up before they have read anything, demand an email address before answering, give confident wrong answers, repeat the same canned reply, and make it hard to reach a person. Each of these treats the chat as a gate when visitors wanted a shortcut.
The most damaging of these is the confident wrong answer. A language model will produce a fluent reply whether or not it has the facts, so the chat has to be built to stay inside what it has been given. We cover the methods in how to reduce AI hallucinations.
For a law firm, an assistant answering questions is not a lawyer giving advice, and visitors should be told that. If your business is regulated, check your own rules on what an automated assistant may say.
How do you set up an AI chatbot that works?
Setting up an AI chatbot that works means deciding what it may talk about, giving it the business's own information, building the handoff to staff, and reviewing real conversations after launch. The model choice matters less than those four decisions.
- Define the scope. List the topics the chat should handle and the ones it should pass to a person immediately.
- Give it your facts. Services, areas served, hours, and how to get in touch, written plainly.
- Write the limits. What it must never state, such as prices, guarantees, or advice specific to someone's situation.
- Build the handoff. Route conversations to a channel staff already check, with the full thread attached.
- Decide the fallback. Say what the visitor sees when nobody is available, such as a phone number or a callback request.
- Read the transcripts. The first weeks of real conversations show what the chat gets wrong and which questions your pages fail to answer.
A chat is not always the right call. If your visitors mostly need a phone number and a form, a clear contact page does the job with less to maintain. For how a chat fits with the rest of the product, see how to add AI to your business software.
Frequently asked questions
Will an AI chatbot replace my front desk or intake staff?
No. A chat answers first questions and keeps a visitor engaged until a person is free. Someone still has to qualify the inquiry, make decisions, and build the relationship. The chat makes sure fewer inquiries are lost before that happens.
Should the chatbot tell visitors it is an AI?
Yes. Visitors usually work it out anyway, and finding out later costs trust. Saying so at the start also sets expectations, so offering a person later feels like a step up and not an admission.
Can a chatbot work outside business hours?
Yes, the assistant can answer at any hour. What changes is the handoff. Decide what happens when no one is available: a callback request, a phone number, or a note about when someone will reply.
Is a custom chatbot better than a chat widget subscription?
Not always. A subscription widget is quicker to install and fine for simple needs. Building your own makes sense when you want control over what the assistant says, how staff are notified, and where the conversations are stored.
What does it cost to run an AI chat?
The model is billed by the volume of text processed, so cost follows the number and length of conversations. We explain the drivers in what it costs to add AI to an app.
If you want a chat that ends with a person on the other side, tell us how your inquiries come in today.