Train the AI on your own business knowledge

How do you train an AI on your own business data?

In stipt the AI assistant is called dot. Training here does not mean building or tuning a model yourself: you fill a knowledge base with your own sources, and dot searches it the moment a question or email comes in. Everything lives in one place, in Studio under AI agent.

Every item in the knowledge base is a source. There are four kinds: a text, an FAQ, a URL (your website) and a file (PDF). Internal sources stay internal; only sources you mark as customer sources may be shown to customers. You organise sources into groups, and for each AI tool you choose which groups it may read. That way the customer-facing WhatsApp bot can read different knowledge than the tools for your own team.

Besides the knowledge base you give dot business context via About your company: a short text about what you do, what you offer and what dot should keep in mind. That text is used by every tool. If you first want to know what AI can and cannot do well for small businesses, read the guide on AI for small businesses.

How do you have the AI read your website?

In the knowledge base, choose a URL source, paste your website address and click Fetch. You immediately see a preview of the first page, so you know the site is readable. After saving, dot crawls the whole site in the background, up to 25 pages per source, and refreshes that knowledge weekly. Update your website and the changes flow in by themselves.

Every source shows a status: Queued, Indexing, Indexed or Error, plus the number of fragments the content was split into. Those fragments are the pieces dot retrieves per question.

If your website is larger, dot indexes up to the limit of 500 fragments per source and the source shows how many pages are not indexed yet. In the edit window you add those remaining pages as a new source with one click. Split sites never index the same page twice.

What documents can you add?

You upload PDFs as a file source: drag, drop, done, up to 14 MB and 100 pages per file. dot reads the text including photos and scans, so a scanned brochure or a datasheet full of tables comes through as well. A very long PDF is sometimes not read in full; it then gets the label Truncated, so you can see it. If an upload or indexing fails, the source shows the concrete reason, for example too large, too many pages or password protected, and stipt automatically retries stalled processing.

For knowledge that is not written down anywhere, use a text source. Put a clear heading above each topic, so dot can retrieve per topic instead of the whole text at once. For loose question-and-answer pairs, such as frequently asked customer questions, there is the FAQ source.

How do product datasheets end up in the knowledge base?

Attach a datasheet PDF to a product and it automatically lands in the knowledge base, in the fixed group Product datasheets. You manage datasheets on the product itself, not in the knowledge base; which tools may read them is controlled through groups, just like other sources.

If several products use the same file, stipt recognises that by content and indexes it once, shared across those products. Remove the datasheet from the last product and the source disappears by itself. Your product knowledge stays in sync with your catalog, with nothing to maintain twice. How the product catalog itself works is covered in the quotes guide.

Where does the AI use that knowledge?

In three places, mainly. In the inbox, dot drafts a reply to an email on request, using real customer data and knowledge from the knowledge base. The draft always comes to you for review first; no email ever goes out automatically.

In chat, dot answers questions from your team: about a customer, about how you work, or simply something from the knowledge base. Answers come with clickable source references; a click opens the source used, so you can check where something came from.

On WhatsApp, dot can answer on its own. The internal bot is for your own staff, the customer bot answers customer questions from the customer-safe knowledge you connect. Both are off by default; no bot ever answers a real customer until you deliberately switch it on. With its first answer in a conversation, the customer bot also tells the customer they are chatting with an AI assistant.

How do you steer the AI with feedback?

You can give feedback anywhere dot answers. In chat you use the thumbs-down. The internal WhatsApp bot says so itself when an answer partly comes from general knowledge instead of your knowledge base, and adds the buttons This is right and This is wrong under its answer. A typed message like this is wrong counts as feedback too.

All feedback lands in one log in Studio. Admins see who gave the feedback, on which answer, with a click-through to the source that was used. Through Answer questions you work through open questions dot had no source for in bulk: you write an answer per question, each answer is saved as a knowledge base item and the feedback is marked resolved. The knowledge base grows exactly where it fell short.

How do you keep the AI away from certain topics?

You set boundaries per tool. You pick from standard boundaries: never promise prices or discounts, no legal or medical advice, do not discuss competitors, and complaints always go to a human. You can add your own as well. On top of that you give each tool instructions in plain language, for example keep it short or always mention the delivery time, and for customer-facing tools you choose a tone of voice and whether the AI addresses customers formally or informally.

The customer WhatsApp bot has extra guardrails. It only answers from the connected knowledge base: no knowledge, no answer, and it makes no hard promises. Under Always involve a human you list topics that go straight to a colleague, such as complaints or cancellations. With Hand over when unsure the bot judges how confident it is before sending; below your threshold it hands the conversation to a human instead of sending a shaky answer.

The internal bot has hard rules too: prices, competitors, specifications of a specific brand or model, and legal or tax advice never come from general knowledge, not even with a disclaimer. Those only come from your own knowledge base or retrieved data.

How do you test whether the AI is trained well?

Nearly every tool has a Test button: you try it on an example or a recent real case and see the result, with no customer involved. The WhatsApp bots have a built-in test chat where you talk the way a customer or colleague would; it never touches a real WhatsApp.

Whether your knowledge base is in good shape shows on the dot page in Studio. The brain shows per source how often it was consulted in the past 30 days; sources that are never consulted appear dimmed. The dot log also shows every AI run with its status and the linked record. For the customer bot there is a performance panel covering the past 30 days: how many conversations were resolved, how often a human took over and how confident the bot was on average.

Frequently asked questions