Grounded AI Reply Suggestions: Why the Draft Shows Its Sources
Chat's AI-suggested replies retrieve from your knowledge base, cite what they used, and never go to the visitor without an agent reviewing them first.
Ask an AI to draft a support reply with no other instruction and you tend to get something fluent, confident, and occasionally wrong — an answer that sounds like it knows your refund policy without ever having read it. Chat's suggest-reply feature is built specifically to avoid that failure mode, and it's worth explaining plainly because the fix is unusually concrete.
What "grounded" means here, mechanically
When an agent requests a suggested reply, the system retrieves relevant passages from the workspace's own knowledge base — not the model's general training, your actual published documentation — and puts those passages directly into the prompt. The model is then asked to answer using that material and to cite inline which passage backed which part of its answer.
That citation step is the part that matters most. Those inline citations are stripped out of the reply text before it's shown to the agent, and returned separately as structured sources. The result is a draft that reads naturally, with the receipts kept alongside it rather than baked into the prose. An agent can glance at the sources and know immediately whether the draft is actually backed by something real, or whether it's thin — a wrong suggestion is checkable rather than plausible, which is the entire point.
What it doesn't do
The suggested reply is a draft, full stop. It's never sent to the visitor automatically — an agent has to review it, edit it if needed, and send it themselves. This isn't a minor caveat; it's the boundary that keeps a hallucinated or misapplied answer from reaching a customer. The AI's job here is to make writing a good reply faster, not to remove the human from the loop on what actually gets said.
Why this is worth trusting more than a generic "AI-powered" claim
Grounding and citations have become the standard framing the industry uses to separate production-ready AI support tools from riskier ones, and for good reason: an AI reply with no visible sourcing asks you to take its accuracy on faith. Chat's suggest-reply path doesn't ask for that. It shows its work by construction — retrieve first, cite what you used, strip the citations into a structured field an agent can actually inspect — rather than as an afterthought bolted on to make the feature sound safer than it is.
What this means day to day
If your knowledge base is thin on a topic, the suggested reply for a question in that area will be thin too, or the sources returned will make that obvious. That's a feature, not a bug: a suggestion with weak sourcing is a signal to write the reply yourself and, ideally, go fill the gap in your knowledge base afterward so the next similar question gets a better draft.