How is an AI agent trained on your business information?

«Training» is not exactly what happens

The word "training" suggests the model learns your business the way an employee learns a trade: through repetition, until it sticks. That is not what happens. At Vyrkon, your information does not modify the model — it is stored in a knowledge base belonging to your agent, and the system consults it on every question to build the answer.

The distinction is not a technicality, because it changes what you can expect. If your information trained the model, correcting a fact would require retraining it, and deleting something would be nearly impossible. Because it is a lookup, the knowledge is edited the way a file is edited: you load the new version, remove the old one, and the next conversation already answers with what is current.

It also changes how much material is needed. Real training demands volume; a knowledge base demands precision. You do not need thousands of examples: you need what you load to be true, complete and free of contradictions.

Where the material comes from

From three places, and it is worth telling them apart because they are maintained differently. The first is the rules: the text defining who the agent is, what tone it uses, which topics it does not touch and who each type of case goes to. That is defined once and changes little.

The second is the documents you upload: the price list, the hours, the requirements, the return policy, the PDF you currently send everybody over WhatsApp. It is the bulk of the knowledge, and also what ages fastest.

The third is connected sources. Your website can be read directly, and if the catalog lives in Shopify, WooCommerce, HubSpot, Odoo or Clientify, it is connected once and refreshes itself several times a day. Anything that can be connected instead of uploaded, connect it: an uploaded file is a snapshot of the day it was uploaded.

Which material works best

Here is the real work, and it is honest to say so: uploading the files takes minutes; deciding what is true in your business takes longer. The material that works best is not the brochure — it is what your team answers every day, in the customer's words, not the org chart's.

The brochure fails for a concrete reason: it is written to persuade, not to answer. It says «personalized service» and the customer asks «are you open Saturday at seven?». An agent loaded only with marketing material answers beautifully and answers nothing.

• Your team's real answers: go through a month of chats and pull out the repeated questions, with the answer actually given.
• One single version of each thing: two price lists coexist badly, and the old one always finds its way out.
• The edge cases in writing: what is said when they ask for a discount, when they complain, when they ask about something you do not offer.
• The customer's words: if your team says «restoration» and the customer says «filling», the knowledge has to say both.

What happens with what you did not load

The short answer: it does not know, and a well-built agent says so. When the question has no backing in the knowledge, the right move is to admit it and pass the conversation to a person with the case already summarized — not to improvise an answer that sounds plausible.

This is worth being clear about from day one, because it defines how the agent is tested. The test is not only asking it the questions it can answer and checking that it answers them well: it is asking the ones it cannot and verifying that it does not answer them. An agent that answers everything with confidence is not better trained — it is worse supervised.

And there is a middle case that deserves its own article: the fact that is loaded but went stale. There the agent is not inventing — it is repeating something that used to be true, with precision. That is fixed by removing the document, not by layering instructions on top, and the detail is in the article on making things up.

Training does not end on day one

An agent is not trained once: it is maintained, because knowledge ages at the pace of the business. A price changes, a service is added, a promotion ends — and each of those is a document to replace, or a connected source that already brings it in on its own.

That is what having every conversation recorded in the Vyrkon portal is for: maintenance is not done by imagining what people will ask, it is done by reading what they actually asked. The conversations where the agent handed off are the to-do list — each one points at a piece of knowledge that is missing or in the way.

The routine that works is short: review conversations regularly, correct in the knowledge —not in the instructions— and test again with the exact question that failed. It is less work than it sounds, and it is the difference between an agent that improves and one that quietly degrades.