Reviewing flags
Work through what the AI found — confirm a genuine mistake, close a false alarm, read the passage it judged against, and answer an agent who disagrees.
The Flags tab
Team → Agent monitoring → Flags opens on four tiles for the period you are looking at:
| Tile | What it counts |
|---|---|
| Turns checked | Agent replies the AI reviewed — the denominator for everything else |
| Unhandled | Flags nobody has reviewed yet. This is the number on the sidebar badge |
| Confirmed inaccuracies | Reviewed, and the agent really was wrong |
| Closed | Reviewed, and decided not to count |
Each tile counts its own status, so changing the status filter does not change the tiles — but the date range, agent and brand filters do.
Flags are listed unhandled first, then newest first, 25 to a page, so the work waiting for you is always at the top.
What a flag shows
Every flag is a card with the agent's name and the time, a status pill, and the contradiction side by side: Agent said (their literal words) against Knowledge base says (what your knowledge actually states). Underneath, each source is listed with a short quote and a View passage button.
A blue Disputed by agent badge means the flagged agent has written on the thread — read that before deciding.
Expanding Details & comments reveals the AI's one-line explanation, the claim as it understood it, and the full discussion.
Deciding: confirm or close
An unhandled flag offers two decisions, and both let you leave an optional note for your team:
Confirm inaccuracy
The agent really was wrong. The flag leaves the unhandled list, still counts as an inaccuracy in reports, and the private note stays in the conversation, now marked as reviewed with your name.
Close — not an inaccuracy
It should not count — your knowledge base was outdated, or the AI misread the answer. The private note is removed from the conversation and the flag stops counting as an inaccuracy.
Nothing is final: Reopen puts a flag back in the unhandled list, and re-creates the private note if closing had removed it.
The evidence view
View passage opens the exact passage the AI judged against, with the contradicting sentence highlighted, and tells you whether it is still what your knowledge says:
- Current copy — the passage is still indexed exactly as shown.
- Changed since — the source has been edited, re-read or removed since the flag was raised, so what you are reading is the snapshot taken at the time.
From here, Open source to edit jumps straight to the right editor — the knowledge base article, the documentation page, the brand's Knowledge screen or Products — and Open page opens the public page if the source has a web address. Fixing the wording at the source is usually the real resolution; see Writing articles.
Discussion and disputes
Every flag carries a comment thread, and it is the reason monitoring does not feel like surveillance to the people being monitored.
- The flagged agent can dispute a flag straight from the note in the conversation — "my answer was right, the knowledge base is out of date" — without needing the Agent monitoring permission.
- Reviewers answer from the flag card.
- Comments are tagged with their author, and the flagged agent's own replies are marked, so it is obvious who is arguing what.
Comments never change a flag's status. Only a reviewer can confirm, close or reopen.
What the agent sees in the conversation
Inside the chat, the flag appears as an amber Possible inaccuracy card — Confirmed inaccuracy once reviewed — marked not visible to visitor, with the same side-by-side comparison, the source, and buttons to View source or Dispute / comment. It also appears in the mobile app.
Private notes and visibility are explained in Visibility.
Per-agent view
Open any agent from Reports → Live chat and their page carries an Agent monitoring (last 30 days) card: how many inaccuracies they have, how many are still unhandled, and across how many checked replies. View flags opens this page filtered to that person.