An AI concierge on a business website should answer common questions about the business, using only the business's own material. It should explain services and process, point visitors to the right page, and hand the conversation to a person when the question needs one. It should not guess at prices, availability, or eligibility. It should not give legal or medical advice, collect sensitive information, or pretend to be human. A concierge that knows its limits protects the company's standing. One that does not can damage it in a single answer.

Should a business website have an AI chatbot?

Only if it has a clear job. We start with purpose, not with tools. Before choosing any platform, we ask what the concierge is for. Common answers are explaining services, describing how an engagement works, and directing visitors to the right page or person. Each purpose decides what the concierge needs to know and what it must never attempt.

Consider a hypothetical cabinet shop. Its visitors ask the same few questions: what kinds of work the shop takes on, how a project begins, and what the client needs to bring to the first meeting. A concierge can answer those well, because the answers are stable and already written down. Visitors also ask how soon the shop can start. That answer changes week to week, so the concierge should not give it. It should explain that lead times depend on the current schedule and offer to connect the visitor with someone who knows.

If a business cannot name the questions a concierge would answer, it does not need one yet. A clear website does more good than a chat window that guesses.

What should an AI concierge know?

It should know the business, and only the business. We write its knowledge from the company's own material: services, process, policies, frequently asked questions, and the limits of what the company offers. It does not draw on general web content to fill gaps, because a confident answer borrowed from somewhere else is still wrong for this company.

The concierge should also say the same things the website says. If the site describes a service one way and the concierge describes it another, the visitor has reason to trust neither. That is the same discipline we apply to search: consistent facts across pages are what let any reader, human or machine, repeat a business accurately.

What should an AI chatbot not do on a website?

The most useful part of a concierge's instructions is the list of things it declines. We write that list before we write anything else, and we write it in plain language.

Why should it never guess on prices, availability, or eligibility?

These are the answers a visitor acts on. A quoted price becomes an expectation. A promised date becomes a plan. A statement that someone qualifies for a service becomes a commitment the business may not be able to keep. Each of these depends on details the concierge cannot see, such as the scope of the job, the current schedule, or a person's circumstances. So the rule is firm. The concierge can explain how pricing works, what affects timing, or what the general requirements are. It does not state a number, a date, or a yes.

Why should it stay out of legal and medical questions?

Because a wrong answer can hurt someone, and a right answer still is not advice. A concierge on a law firm's site can explain which kinds of matters the firm handles and how a consultation is arranged. It does not tell a visitor whether they have a case. A concierge on a healthcare provider's site can explain services and how to begin. It does not interpret symptoms or suggest treatment.

In regulated sectors such as healthcare, we scope the concierge to general information and direct anything personal to a person. The client's own counsel should review how it will be used before launch.

How do you keep a website chatbot from giving wrong answers?

No setup removes every error. Several habits reduce them, and we treat each one as required.

  • Written scope. A short document lists the topics the concierge handles and the ones it declines. Anything not on the first list is treated as out of scope.
  • Decline rules stated plainly. When a question is outside scope, the concierge says so politely and offers a person. It does not improvise.
  • Permission to say it does not know. A concierge that is never allowed to admit uncertainty will fill the gap with something plausible.
  • Hard questions before launch. We test it against the questions buyers actually ask and against questions written to push it off course: requests for a price, a guarantee, a diagnosis, a legal opinion, or a comparison with a named competitor. It goes live only when it answers accurately, declines cleanly, and hands off reliably.
  • A review routine. After launch, someone reads a sample of conversations on a set schedule. Wrong or awkward answers lead to changes in the knowledge or the rules.
  • Updates when the business changes. A new service, a retired one, or a changed policy goes into the concierge's knowledge at the same time it goes onto the website.

When should a chatbot hand off to a human?

Whenever the visitor needs judgment, commitment, or care. In practice that means a handoff when:

  • the question touches price, timing, eligibility, or anything personal
  • the visitor is upset, confused, or asks the same question twice
  • the visitor asks for a person
  • the question is outside the written scope

A handoff should be easy and should carry context. The concierge can point to the contact form or scheduling page, or notify staff, with a short summary of what the visitor asked. The visitor should not have to start over. A handoff is not a failure of the concierge. It is often the most valuable thing it does, because it brings a qualified visitor to the right person at the right moment.

What information should a website chatbot never collect?

A chat window is not a secure intake form, and it should not act like one. We write the concierge to avoid asking for, and to discourage visitors from typing, sensitive information. That includes health details, account and identification numbers, payment card data, and the specifics of a legal matter. Assume anything typed into a chat is stored somewhere. If a visitor starts to share something sensitive, the concierge should stop them gently and point them to the proper channel, such as a secure form or a call with staff.

The safest concierge asks for very little. A name and a preferred way to be contacted are usually enough to arrange a handoff, and even those can wait for the contact form.

Should the concierge say it is an AI?

Yes, always. This is our standard, and we do not make exceptions for it. The concierge introduces itself as an AI assistant, does not use a human name or photo that suggests otherwise, and offers a person whenever the visitor wants one. Visitors judge a company by how it treats them. Being honest about who, or what, they are talking to is part of that treatment. A business that misleads people in its first conversation has a hard time earning trust in the second.

What to do next

Write down the five to ten questions visitors ask most often, and mark which ones have stable, published answers. Those are the concierge's scope. Then write the list of what it must decline. If the second list is longer than the first, a clearer website may serve you better than a chat window right now.

When you are ready to build one to this standard, our AI concierge agents practice covers scope, knowledge, decline rules, handoff, and testing before launch.