Testing the AI agent

AI Agent → Test runs the real agent against your real knowledge. Nothing is saved and no customer sees it, so use it freely: after adding a source, after writing a snippet, after changing the tone.

Asking

Type a question the way a customer would and choose Ask. To see how the agent behaves for a particular customer, use Preview as with their email or user ID. That matters for actions with a verified identity, since the answer depends on who is asking.

Reading the result

Each run tells you what the agent decided:

• Answered: it found enough knowledge and replied.
• Small talk: it treated the message as a greeting or chit-chat.
• Asked to clarify: the question was too vague to answer.
• Handed to a human: it had no knowledge for this, or was asked for a person.

Below the answer:

• Knowledge it considered lists what retrieval found, best match first. If the right article or snippet is missing here, the problem is your content, not the answer. Add or fix the material and ask again.
• How it searched shows the ways the agent rephrased the question to look for matches.
• Self-review shows how the answer scored when the agent re-read it: whether it is backed by your knowledge and whether it answers the question. Below 70 the answer is rewritten once, then goes to a person. If you see a rewrite, the first draft failed review.
• Where the time went breaks the reply down by stage. The model calls are the cost; everything else is milliseconds.

A simple routine

1. Collect the ten questions customers ask most.
2. Ask each one here.
3. For every weak answer, look at Knowledge it considered. Add the missing article, source or snippet.
4. Ask again until the answers hold.