Why our help desk agent says “I don’t know”
A support agent that guesses is worse than no agent at all. How Chatty decides it doesn’t know, what it does next, and how the gap gets closed.
The quickest way to lose a customer’s trust is to answer their question confidently and wrongly. A made-up refund window, a feature that doesn’t exist, a price from last year. Each one turns into a support ticket later, or worse, a customer who quietly leaves.
So the most important thing Chatty’s agents do is not answering. It is knowing when not to.
What “knowing” means here
When a question comes in, the agent searches your knowledge: the web pages, files and videos you added. The search is hybrid. It looks for passages that mean the same thing as the question, and also for the exact words in it, so a product code or an error number is found even when a meaning-based search would miss it.
Every passage the agent uses gets a number, and the answer cites those numbers. Your visitor sees which page each part of the answer came from, and can check it.
If nothing relevant enough comes back, the agent does not fall back on what the model happens to know about the world. It says it couldn’t find the answer in your content. How strict “relevant enough” is can be set for each agent.
Then it offers a person
Saying “I don’t know” is only half of it. Next, the agent offers to pass the question on. If the visitor says yes, a ticket appears on your board with the whole conversation attached, the page they were on, and, inside your app, who they are and which plan they’re on. Nobody has to ask the customer to explain it all again.
And the gap is written down
Every “I don’t know” is recorded as a knowledge gap, and counted again each time someone else asks. The list shows the questions your content doesn’t answer, most asked first.
Closing one is ordinary work: fix the page or add the document, re-sync the source (only the pages that changed are read again), and mark the gap resolved. The next visitor who asks gets an answer with a source.
Why not let the model answer anyway?
Large models know a lot about the world and nothing about your refund window. A plausible answer about a price or a policy is the most expensive kind of wrong, because it sounds right.
There is a second reason. The agent on your website only ever searches knowledge you have marked as public. An internal document, a draft price list or a support playbook never ends up in a visitor’s answer, because that agent cannot see it.
What it costs you
Some questions go unanswered that a looser chatbot would have “answered”. We think that is the right trade. A customer who hears “I couldn’t find that, shall I ask the team?” still trusts you. And the gaps list turns each of those moments into a fix you can make once.