Your front desk answers the same questions every day: what time you open, whether you serve a town, what to bring to a first visit, how to request a quote, and where a request stands. Phone and email absorb the time. Now a vendor pitches an AI chatbot for small business website support, and the worry is simple: a bot that makes up an answer in your company’s name is worse than no bot at all. This post explains what a useful chatbot does, what it must never answer, and when you should skip it.
A chatbot that answers with confidence and no source is the problem to solve. The useful kind is narrower: it answers from content you approved, shows where the answer came from, and passes the customer to a person when it reaches the edge of that content.
What it answers from
A website chatbot should answer from your approved content and nothing else. Approved content means pages you have published, an FAQ you have written, and policies you have put in writing. When the answer is not in those sources, the chatbot should say so and hand off to a person. It should not guess.
Start with the questions your staff answered most often this month, then write the answer for each one. Those answers are the first version of the content. A chatbot built on your own content is only as good as the pages behind it, so the writing comes before the bot.
The table shows one example of how a typical front desk might set this up. It is an illustration, not a project we ran.
| Content source | Example | Who keeps it current | When it changes |
|---|---|---|---|
| Service pages | Service area, services offered, what is not offered | Owner or office manager | Edit the page the same day |
| Hours and location | Holiday hours, parking, entry instructions | Front desk lead | Edit the page the day the hours change |
| FAQ | What to bring to a first visit, how to request a quote | Office manager | Review monthly and after any change |
| Written policies | Cancellation, deposit and return rules | Owner, reviewed by your attorney | Update the policy first, then the page |
Keep each answer on the page, not in a second copy inside the bot. A second copy is a second version of the truth, and nobody remembers to update it.
Staff may already be pasting customer details into public chatbots. That is a staff policy question, covered in shadow AI in a small office.
What it must never answer
Three kinds of question stay out of its reach:
- Anything not in the approved content. A guess that sounds like policy is the failure you are trying to avoid.
- Anything about one customer’s account, unless the chatbot is built with a sign-in that checks who is asking.
- Any decision with legal, medical or financial effect.
A customer asking whether a refund applies, whether a warranty claim is valid or whether a fee can be waived is asking for a decision, not a fact from a page. The bot should route that question to a person every time, even when the wording looks simple.
AI output can be wrong, so decisions with legal, medical or financial effect should stay with a person. Whether a clause in your contracts or a rule that applies to your business affects a chatbot is a question for your attorney. This post does not state what any rule requires.
Show the source
Each answer should link to the page or document it came from, so the reader can open it and read the exact wording. Our AI integrations work follows the same standard: a customer-facing chatbot built to show the source behind each answer. A source link does not make an answer right, but it gives a reader and your staff a way to check it.
In any demo, run three checks:
- Ask a question your content answers. Open the source and confirm the answer matches the page.
- Ask a question your content does not answer. The bot should say it does not know and offer a handoff. It should not invent a reply.
- Ask for a price, a promise or an exception that is not on any page. The bot should not supply one.
If a demo fails any of these, do not treat it as a small glitch. Ask the builder which content the bot read, and why it chose that answer.
The handoff to a person
A handoff should send the customer to a person when the chatbot cannot answer, when the customer asks for one, or when the question is on the never list. The handoff can go to an email inbox, a ticket queue or a phone callback request. Pick one, and name the person who owns it.
An unowned handoff is a dead end on your website. The customer asks a question, gets a message that the bot cannot answer it, and then hears nothing. Before launch, test the handoff yourself: send a question, confirm the message arrives where it should, and see how long a person takes to reply.
A useful handoff message carries the customer’s question, the contact details they gave, and the pages the bot already checked. The person who picks it up should not have to ask the customer to repeat everything.
Limits on instructions
Text inside a web page, an email or an uploaded document can contain instructions aimed at the chatbot. The chatbot should not follow them. Our AI Integrations page describes the controls for this as “controls that limit what instructions hidden in documents and email can do,” along with “clear limits on what the AI can see and do.” Ask any builder for those limits in writing: which documents the bot can read, what it can change, and what it can send. An assistant with unrestricted file access is a breach waiting to happen, so these limits belong in the scope, not in a footnote.
Logs and review
Ask for conversation logs from the first day, decide who can see them and how long they are kept, and name the person who reads a sample every week. That reader should know your business, not only the vendor. Put the review time on the calendar before launch, because a log nobody reads is only storage.
When a wrong answer shows up, fix the page it came from. Do not patch the bot with a special case. A special case hides the real gap in your content, and the next wrong answer will come from somewhere else.
When not to add one
Skip a customer service chatbot if any of these describe your business:
- Few customers ask the same questions. A clear FAQ page may do more.
- Your answers are not written down. Write the policy first. A bot cannot answer from what lives only in one person’s head.
- Nobody can own the handoff inbox. A handoff with no owner leaves customers stuck.
- Each question needs a person’s judgment. A chatbot cannot weigh each case the way a person can.
Our home page says it plainly: “If a project is a bad idea, we say so before you spend the money.” If one of these describes you, we will say so. If you have not yet picked where AI fits in your office, start with AI for Small Business: Where to Actually Start.
What you keep
When the work ends, you should own the content, the configuration and the documentation. Our home page puts it this way: “Diagrams, configurations and credentials belong to you. If you leave, you take all of it.”
Before you sign anything, ask the builder for the content list, the handoff setup, the review checklist and the log access in writing. Ask who holds each item at handover. The access control and logging work behind a chatbot is covered under our cybersecurity services.
Next step
Tell us which questions your staff answer over and over. Book a 30-minute call or call 224-531-1000.