Can an AI agent make things up about my business?

Where the answers actually come from

A Vyrkon agent answers with two things: its instructions —the text that defines who it is, what tone it uses and what it must not do— and its knowledge, which is the documents, prices, hours and catalogs the business loaded. When someone asks a question, the system searches that knowledge and hands the model the fragments that look relevant.

That matters because it changes where to look when something goes wrong. If the agent gives a wrong price, it almost never imagined it: it read it. The price is sitting in one of the files that were uploaded, in an old version of the catalog, in a PDF nobody has opened since the day it was loaded.

Knowledge beats instructions

This is the lesson that cost us the most, and it is counterintuitive. When an agent repeats an old fact, the natural reaction is to add an instruction: "do not mention the previous catalog", "the correct prices are these". It almost never works.

The reason is that the instruction and the document are not competing on equal terms. The document arrives as concrete context, carrying the specific fact the person asked about; the instruction is a general rule the model has to remember to apply. When the knowledge says one thing and the prompt says another, the knowledge usually wins.

The repair is not writing a better instruction: it is taking the fact out of where it lives. Delete the old document, replace it with the current one, and verify by asking the agent exactly what a customer would ask. If the fact is no longer loaded, there is nowhere to pull it from.

Denying is also asserting

There is a sibling of the error above, and it goes far more unnoticed: the agent denies something without ever having checked. "No, we do not offer that service", "we have no availability" — said with confidence, and without looking.

It gets missed because when you review conversations you go looking for odd data: a price that does not add up, an impossible date. A denial has no numbers to check, so it reads like a cautious answer. But for the customer who asked, the effect is identical to a false fact: they leave with wrong information, and convinced of it.

A well-built agent distinguishes three things: what it knows, what it does not know, and what it knows does not exist. Only the third one licenses a denial.

What can be done so it does not happen

The first thing is being able to see it. In the Vyrkon portal every conversation is recorded, so a wrong answer is caught by reading what the agent actually replied, not what it is supposed to reply.

The second is that the knowledge updates itself wherever it can. If the catalog lives in Shopify, WooCommerce, HubSpot, Odoo or Clientify, it is connected once and refreshed every 6 hours. A new price no longer depends on somebody remembering to upload a file again, which is exactly where stale data gets in.

• Load one single version of each thing: two price lists coexist badly.
• Connect the source instead of uploading files, whenever the system allows it.
• Test with the customer's words, not the business's words.
• Review real conversations regularly, especially the ones that ended in a denial.