Try Omnifox free 7 days free. No credit card required.
Guides

AI Website Chat That Answers From Your Knowledge Base: Documents, FAQ and Catalog

AI website chat only works when it answers with your data: load documents, FAQ and catalog, test in the simulator, set guardrails and hand off to agents.

September 06, 2026
AI Website Chat That Answers From Your Knowledge Base: Documents, FAQ and Catalog
In this article

    AI website chat has a reputation problem, and it earned it. Most visitors have met a bot that answers "Great question! Our team is here to help" and nothing else, or worse, one that confidently invents a return policy the company does not have. The difference between that and an AI agent people trust is not the model. It is whether the agent answers from your knowledge base (documents, FAQ, catalog with real prices and stock) and knows when to stop and hand off to a person.

    This post is about building that second kind: what to load first, how to keep it from contradicting itself, how to test before a visitor sees it, and which guardrails keep it honest.

    Generic chatbot vs AI agent grounded in your data

    A generic chatbot, rule tree or language model with no context, has two failure modes: it does not know, so it deflects; or it does not know, so it makes something up. A grounded AI agent has the third option you actually want: it does not know, so it says so and gets a human. The same question, answered both ways, for a small HVAC company in Phoenix:

    Visitor: Do you do same-day AC repair in Chandler and what does the call-out cost?

    Bot without a knowledge base: Yes, we offer fast and reliable AC repair services in your area! Contact us today for a free quote.

    AI agent with a knowledge base: Yes, Chandler is in our service area. Same-day appointments are available Monday to Saturday if you book before 1 pm. The diagnostic visit is $89 and it's waived if you go ahead with the repair. Want me to check today's slots? I'll need your ZIP and a phone number to confirm.

    The second answer came from a service-area document, a pricing document (the fee is illustrative) and the booking tool. Every sentence traces back to something you wrote.

    What to load first: the 20 real questions from your inbox

    Do not start by uploading the entire website. Open your shared inbox (email, WhatsApp, existing website chat, phone notes) and pull the last few weeks of first messages. A short list covers most of the volume: hours, service area, price of the common thing, delivery time, returns, how to book, how to cancel, which insurance or payment methods you take. A practical order:

    1. The 20 most common questions, each with the answer your best agent gives, written as an FAQ entry.
    2. The pricing page or price list, as a document, with the date it was last updated in the text.
    3. Policies: returns, cancellations, warranty, what is and is not covered.
    4. The catalog, if you sell products: name, variant, price, availability, and the one-line description a salesperson would use.

    Test after each step and stop adding when the agent handles a week of real conversations without a wrong answer. Conflicting documents are worse than missing ones.

    One source of truth per topic

    The most common cause of a wrong answer is two documents that disagree: last year's price list and this year's, a shipping FAQ that says 3 to 5 days and a product page that says 2. The model cannot know which one you meant.

    Topic Keep one source Retire or merge
    Prices The current price list with a "valid from" date Old brochures, blog posts that mention prices
    Hours and locations One "locations and hours" document Footer text copied into three places
    Returns and cancellations The policy page, verbatim Support macros that paraphrase it
    Product details The catalog (with price and stock) Marketing PDFs with outdated specs
    Insurance, payment methods, compliance One "what we accept" document Scattered FAQ entries

    When a topic changes, update the one source and re-sync. If you maintain a help center, use it for how-to content and keep the FAQ for short, high-volume answers.

    Test in the simulator before anyone sees it

    Publishing an AI agent without testing is putting a new hire on the phones with no shadowing. Use the simulator in the dashboard against your real question list before you set the reply strategy to anything but "never":

    • Ask the 20 questions exactly as visitors write them, typos and all. Note every answer that is wrong, vague or too long.
    • Ask five questions the knowledge base does not cover. The correct response is a handoff, not a guess.
    • Ask two questions that bait a promise ("can you confirm my refund is approved?"). The correct response is to take the details and pass to a person.
    • Ask in a second language if you serve one; the agent replies in the visitor's language, so check the meaning survived.
    • Re-run the list after every knowledge base change. Ten minutes, and it catches regressions.

    Guardrails: do not invent, hand off

    The agent's instructions should say plainly what it may and may not do. Three rules cover most of it:

    1. Only answer from the knowledge base. If it is not there, say so and offer a human. No "typically" guesses about prices, availability or policy.
    2. Do not commit the business. No confirming refunds, discounts, delivery dates or medical, legal or financial specifics. Collect the details and hand off.
    3. Hand off on request, on frustration, and on anything with money attached beyond what the price list states.

    The handoff should be smooth: the agent tags the conversation, assigns it to the right team (sales, support, billing) using that team's distribution strategy, and tells the visitor what happens next. Outside business hours it keeps answering from the knowledge base, collects name and email, and lets the conversation wait in the shared inbox. The routing side is covered in departments in website chat. The line it uses at the edge of what it knows matters more than most people think:

    I don't have that detail in front of me and I'd rather not guess. I'll pass this to Priya on the support team; she usually replies within the hour during business hours. Can I get an email or phone number so she can reach you if you close this tab?

    The last sentence captures a way to continue by email, SMS or WhatsApp if the visitor leaves.

    Reply strategy: always, first message only, or never

    The agent's reply strategy decides how much of the conversation it owns.

    • Never: the agent is off; useful while you test in the simulator.
    • First message only: the agent answers the opening question, then a human takes the thread. The safest live setting for a sales-heavy site: the visitor gets an instant reply and the human sees the context.
    • Always: the agent handles the whole conversation and hands off only when the guardrails say so. Right for support-heavy sites with a mature knowledge base.

    Start at "first message only" and move to "always" once a few weeks of transcripts show the agent is not being corrected. Either way it uses the AI credits included in your plan, expandable as volume grows; see the pricing page.

    Frequently asked questions

    What is the difference between a chatbot and an AI agent with a knowledge base?

    A chatbot follows scripted rules or general training data; an AI agent with a knowledge base answers from your own documents, FAQ and catalog and hands off to a human when the answer is not there. Only the second can quote your real prices and policies.

    What should I upload to the knowledge base first?

    The twenty questions your inbox receives most often, each with the answer your best agent gives. Add the price list, policies and catalog after that, one topic per document, and test in the simulator after each addition.

    How do I stop the AI from making things up?

    Instruct it to answer only from the knowledge base, say when it does not know, and hand off for anything involving money, medical, legal or commitments. Then test with questions it cannot answer and confirm it hands off instead of guessing.

    Start with twenty questions, not a hundred documents

    A grounded AI agent is a list of real questions, one clean source per topic, a simulator run and a handoff rule. Do those four things and your website chat answers like your best person on their best day, and steps aside when it should.

    Try it with a 7-day free trial: create the website chat channel, upload your FAQ, test in the simulator, then switch to "first message only". The AI agent, triggers and knowledge base are described on the website chat page.

    Ready to try it in your business?

    Bring WhatsApp, Instagram, Messenger, email and calls into one inbox, and let AI handle the first reply. Set it up today.

    7 days free. No credit card required.

    Comments (0)

    No comments yet. Be the first to share your thoughts.

    Leave a comment

    Your email is never published. Comments are moderated before appearing.

    Markdown supported. HTML is stripped.

    Read more

    Message us on WhatsApp