Repetitive customer questions were the first thing one small business asked us to help with. Their support team of two started every morning with an inbox of the same few questions: what time do you open, do you deliver to my area, how much is the standard package, is the promotion still on. None of them was hard. Each one had been answered hundreds of times. But they arrived all day, and the one message that genuinely needed a person sat behind the rest.
They had tried a keyword chatbot. It handled "opening hours" and failed "are you open on the public holiday?", so customers ended up in the queue anyway, now slightly annoyed.
What they told us
The answers already existed: on the website, in a policy document, in a price list, in the head of the person who had been there longest. The problem was that customers would not go and read them. Asking is faster than searching, so they asked, and a person had to answer.
What the team wanted was simple to say and hard to get right. Answer the routine questions the way the business would answer them. Never make up an answer the business never gave. And leave the unusual ones for the team.
How we approached it
We spent a long time on the second requirement, because it is the one that decides whether a business can trust an assistant in front of its customers. An assistant that answers from what a model picked up in training will eventually tell a customer something the business never said. So we set three conditions before building anything:
- Answers come only from the business's own content, and the assistant says plainly when that content is silent.
- The team can see how an answer was reached before any customer sees it.
- Nothing changes for live customers until the team decides it should.
How it works
The assistant answers from the knowledge the business gives it: the FAQs it writes, the documents it uploads, and pages imported from its website. A visitor asks in their own words and gets the answer the business would give.
It answers from your content, and declines when it cannot. Since Gydr 2.0 the assistant answers strictly from the knowledge base, and when the answer is not there it says so instead of guessing.
You can check an answer before customers see it. The test page lets the team chat with the widget as a visitor would, and its debug view shows which knowledge each answer drew on. A wrong answer usually points to missing or outdated knowledge, and fixing the knowledge fixes the answer.
Changes go live when you decide. The team edits a draft, saves named versions, and publishes when ready.
Where it leaves the team
The inbox still fills up. Most of the routine questions now get their answer immediately, at any hour, and what reaches the team is the remainder: the complaint, the unusual request, the customer who asks for a person. The assistant does not handle those, and it is not meant to.
Frequently asked questions
Where does the chatbot get its answers?
From the knowledge you give it: FAQs, uploaded documents, and pages imported from your website. It does not answer from general training when your content is silent.
What happens when a customer asks something the knowledge base does not cover?
The assistant says it does not have that information instead of guessing, and the question can go to your team.
Can I test answers before customers see them?
Yes. The test page lets you talk to the widget as a visitor would, and its debug view shows the knowledge behind each answer.