Agent monitoring
AI re-reads what your human agents tell customers and compares it against your brand's knowledge base, leaving a private note in the conversation whenever an answer contradicts what you have documented.
What it does
Your knowledge base is the source of truth your AI already answers from. Agent monitoring points the same knowledge at your human agents: after a monitored agent replies, the AI checks their answer against the brand's knowledge and — when it finds a clear contradiction — leaves a private note in that conversation showing what the agent said next to what the knowledge base actually says.
The visitor never sees the note. Your team does, right where the mistake happened, so it can be corrected while the conversation is still open.
Where to find it
Team → Agent monitoring in the left menu. The page has two tabs:
- Flags — everything the AI has flagged, and where you review it
- Rules — who is monitored, for which brand
A number badge on the menu entry counts the flags nobody has reviewed yet. It updates by itself as new flags arrive and as you work through them, and disappears at zero.
When a check runs
An agent's whole turn — everything they wrote since the customer's last message — is judged once, at the moment the turn ends. That happens when:
- the agent has been quiet for 30 seconds, or
- the customer replies, or
- the agent leaves the conversation
Judging the turn as a whole matters: an agent who states a price in one message and corrects it in the next is read the way the customer reads it, not line by line.
What is never checked
- Private notes between team members
- Messages with no text (an attachment on its own)
- Messages from your AI agent or from the customer — they are context, never the thing being judged
- Replies from agents who have no rule, and any brand with no knowledge to check against
- Replies written before the rule was saved — see Rules
What gets flagged
The AI first asks whether the reply even contains a checkable claim about your company. Greetings, "one moment please", questions back to the customer and one-off promises ("I'll refund you today") are not claims, and the check stops there.
For a real claim, it searches your knowledge base and only a contradiction counts. Before a flag is created, Yaplet verifies the AI's evidence itself: the quoted passage must literally exist in something the search actually returned. Anything the AI cannot prove is discarded rather than shown to you.
What it costs
Each check is one call on your organization's normal AI model, and the tokens are billed like every other AI feature. Checks that find nothing still cost a little — which is why keywords exist, to narrow monitoring to the topics where being wrong actually hurts.
Every check is logged, so the Flags tab can show you how many replies were reviewed, not just how many were wrong.
How long flags are kept
Flags, their private notes and the check log are kept for 120 days and then removed automatically. When a flag is deleted, its note disappears from the conversation with it.
Who sees what
- Reviewing flags and editing rules requires the Agent monitoring permission, granted per team member in Organizations.
- The private note in a conversation is visible to anyone on your team who can open that conversation — including the flagged agent, who can reply to it and dispute the flag.
- An agent without the permission has no flag list of their own; they meet flags inside the conversations they work.
Turning it on
Agent monitoring is included in the Growth plan and available as a $10-a-month add-on on every other plan, the free one included. Switch it on in Settings → Organization settings → Billing, or read more in Plans and pricing. After that, grant the permission to whoever should review flags, and add your first rule.