It is tempting to think an AI agent's quality comes entirely from the model behind it. In practice, the model is rarely the bottleneck. The bottleneck is whether it has anything real to search when a customer asks something specific, like your refund window or how a particular setting works. Without that, it either makes something up or falls back to "let me connect you with someone," which defeats the point of having an agent at all.
Candor's Knowledge Base is built to be that source of truth, for customers browsing it directly, for Candor AI answering in chat, and for the Voice Agent searching it mid-call.
One set of articles, written once
You write Knowledge Base articles the same way you would write any help-center content: organized into categories, using a rich-text editor with headings, lists, and code blocks. Save one as a Draft while you are still working on it. Only Published articles are visible to customers, and only Published articles are ever searched by either AI agent.
Search built for meaning, not just keywords
The search box customers use to browse the help center directly still combines full-text search with a fuzzy word-similarity match, so a search for "verifcation" still finds "Verifying your identity" without needing to spell it correctly.
But an AI agent answering on your behalf needs more than keyword matching. A customer asking "why was I charged twice" and an article titled "Duplicate transaction holds" share no words in common, yet one clearly answers the other. So both Candor AI and the Voice Agent search a different way: every article is split into focused sections and embedded as a vector that captures what a section means, not just the words in it. A customer's question gets embedded the same way, and the closest-matching sections come back ranked by actual relevance, with a slight preference for ones your team has reviewed recently. It is the same retrieval underneath a typed chat message and a spoken question, so the quality bar does not change depending on how the customer reaches out.
Turn it on, and mean it
Enable Candor AI from Settings and every reply it sends is grounded in that same search, never citing something it cannot find. Enable Knowledge Base grounding on the Voice Agent and it does the same thing mid-call. Either way, the practical effect is the same: the sentence you wrote once to answer a customer's email now also answers their chat message and their phone call, without you writing it three times.
Write the article once. Let it answer the question everywhere a customer might ask it.