Why This Reply (AI Reasoning)
Open the reasoning trace behind any AI message to see how and why the agent answered.
What It Is
Every message the AI sends carries a reasoning trace you can open — a "Why this reply?" view. It shows what the agent used for that specific answer: which model ran, how long it took, what knowledge it used, which tools it called, and how it classified the lead.
This is both a transparency feature and a practical debugging tool. Instead of guessing why the agent said something, you can look.
Where to Find It
Open the Inbox and select a conversation. On each AI message, there's a way to open its reasoning trace (the "Why this reply?" panel). It slides in with the details for that one message, so you can inspect any individual reply.
What Each Field Means
| Field | What it tells you |
|---|---|
| Model | Which AI model produced the reply |
| Latency | How long the reply took to generate |
| Knowledge Base passages | Which KB passages were retrieved and used, shown as citations |
| Tools called | Any tools the agent used for this turn (e.g. calendar booking, catalog search) |
| Classification | How the lead was classified on this turn (HOT / WARM / COLD) |
| Fallback path | Which path the reply took when the primary route wasn't available |
Knowledge Base passages (citations)
If the agent used your Knowledge Base to answer, the trace lists the passages it pulled. If this section is empty, the agent answered without any KB support — which is a strong signal that the information may be missing from your Knowledge Base.
Tools called
When the agent books an appointment, searches your catalog, or uses another tool, that call shows here. This is how you confirm the agent actually reached your calendar or catalog rather than answering from general text.
Classification
The trace shows how the lead was scored for that turn, so you can see whether the conversation is being read as HOT, WARM, or COLD.
Using It to Improve Answer Quality
The reasoning trace turns "the agent gave a weird answer" into a specific, fixable observation.
-
No KB passages were used, but the answer should have come from your docs? The information probably isn't in your Knowledge Base, or isn't phrased in a way the retrieval could match. Add or rewrite that content in the KB.
-
The agent gave a vague or generic answer? Check whether any passages were cited. If not, feed the missing facts into the Knowledge Base so the agent has something authoritative to cite next time.
-
A price looked wrong? Confirm the product is in your Product Catalog — catalog prices are the source of truth, and the trace shows whether catalog search ran.
-
The lead was classified incorrectly (e.g. a hot buyer marked COLD)? Refine your classification criteria in AI Agent → Classification so the agent scores conversations the way you want.
-
A tool didn't fire when you expected it to? Review the tools list. If a calendar or catalog tool never ran, the customer's phrasing may not have triggered it — you can adjust your prompt or Knowledge Base to guide the agent.
Reviewing traces on a handful of real conversations is one of the fastest ways to find gaps in your setup and tighten the agent's answers.
FAQ
Is the reasoning trace shown to my customers? No. It's an internal view for you and your team inside the dashboard. Customers only see the reply.
Which channels have it? The reasoning trace is available on AI messages across your channels.
Why is the Knowledge Base section sometimes empty? Because the agent didn't need or couldn't find a matching KB passage for that reply. If the answer should have come from your docs, that's your cue to add the information to the Knowledge Base.
Can I use it to debug wrong answers? Yes — that's a core use. The trace shows what the agent knew and did, which points you straight to what to fix (KB content, classification criteria, or your prompt).