How a chatbot that only answers from your own knowledge works

Most chatbots are built to sound sure. That is the problem: a confident wrong answer about a price, a delivery date or a refund is worse than no answer. The AI Chatbot is built around the opposite idea. It answers only from material you gave it, and when that material does not hold the answer, it stops.

It starts with your material

You give the assistant text you paste in, files you upload and pages you import from your own website. That is its knowledge base. Nothing else counts: it does not fill gaps from general knowledge, and it does not browse the web on the customer's behalf.

Small knowledge bases, such as a typical FAQ, are handed to the model whole, so the question does not have to share exact words with the text. Larger ones are searched first, and only the relevant pieces go to the model. See how knowledge works for the details.

No knowledge, no AI answer

If a question finds nothing relevant in your material, the model is not asked to improvise. The conversation goes straight to a hand-off. This removes the most common way a bot invents things: being asked something it has no source for and answering anyway.

When the model is asked, it is told to use only the excerpts it was given and to hand over when they do not contain the answer. Output that comes back malformed is treated as a hand-off too, never as a guess.

Sources shown, people in charge

Each answer records the sources it rested on, so you can check an answer against the text it came from. And a person always outranks the bot:

• Once someone on your team replies in a conversation, the assistant goes quiet.
• A customer who asks for a person gets one, and the bot does not argue.
• A customer who writes "stop" or something like "please stop texting me" is never answered by the bot, and a person is told.

The hand-off itself is not silent. The customer gets a sentence you wrote, the conversation lands in your Inbox with the question attached, and the people you chose are alerted. Review and hand-off explains the flow.

Shadow mode: read before it speaks

On WhatsApp and e-mail the assistant starts in shadow mode. It writes a draft instead of a reply, and a person approves it, edits it or rejects it with one click. Nothing reaches the customer until then.

Shadow mode is also how you learn. After you have reviewed at least ten drafts, you can switch the channel to send by itself, and you see how many drafts were sent as written, edited or rejected. If you edit most of them, that is a signal about your knowledge, not about the customers.

What this does not promise

A grounded bot is only as good as the material behind it. If your knowledge is outdated, the answers will be too, only clearly sourced. If a question is phrased in a way the model reads differently than you would, a draft can still be wrong, which is exactly why a person reads the first ones.

We also have not measured this chatbot on real customers, so we will not tell you what share of questions it will settle. What we can describe is the mechanism: sources you control, a stop when there is none, and a human in the loop until you decide otherwise.

Where to go next

Add your first knowledge from your website, try questions in the test console, and keep the shadow mode on until the drafts look right.