Chat is unforgiving in a way email and even phone are not. A customer who emails you expects to wait. A customer typing into your widget expects a reply before they finish their coffee, and a lot of what they are asking has already been answered, in a Knowledge Base article, a dozen times this month alone. That is a strange place for a support team to still be typing every reply by hand.
We built Candor AI because that gap did not have to exist either.
What it actually does
Candor AI is a toggle, not a project. Turn it on from Settings, or right from the Inbox sidebar, and it starts answering new widget conversations using the same semantic search over your published Knowledge Base that the Voice Agent uses mid-call. Add an optional personality prompt if you want a specific tone, and it is live.
When it has a real answer, it replies in the thread like any other message, labeled Candor AI so nobody mistakes it for a teammate. When it does not, it says so once, lets the customer know a teammate will follow up, and leaves the conversation exactly where an unanswered conversation already goes: your unassigned queue. No confident guessing, no repeating "let me check on that" forever.
Why this matters more than it sounds
The moment any teammate replies to a conversation, Candor AI steps back permanently for that thread, whether or not anyone formally assigned it. It never talks over a human, and it never needs to be told to stop.
It also never lives in a separate "bot inbox" you have to check alongside the real one. Every reply it sends is a normal message in a normal conversation, sitting next to your team's replies in the same thread, the same Inbox, the same history a customer sees when they scroll up.
Where we are taking it next
Candor AI replies are metered per plan the same way voice minutes and seats already are, with the same grace period before anything locks, visible from Settings > Billing. And because "why didn't it answer that" is the first question you will ask the first time it hands off, Settings > Candor AI now shows exactly which published articles it can see and how much of each is actually indexed, with a one-click way to re-embed an article that predates any of this.
If you have not turned it on yet, do it from Settings and ask it something a real customer would before they do.