Talk to a human, and rate the answers

Two things make the website chat accountable: a visitor can reach a person when the agent is not enough, and every answer can be rated or reported so you learn where the agent goes wrong.

Let visitors reach a human#

With human takeover enabled on the agent's web widget, a chat can pass to your team in three ways:

• the visitor asks for a person from the chat window,
• the agent itself decides it cannot help (repeated failed answers, or someone insisting on a human),
• you take a live chat over from the console.

The moment a chat asks for a person, the agent goes quiet and the visitor sees a notice with their place in line. While waiting they can cancel and continue with the agent. If nobody on your team is available, the visitor gets a fallback message you control, and your automation webhook is notified so you can page someone.

If you do not see the takeover option on the agent's widget settings, ask your administrator to switch it on.

Answer waiting chats#

Open Live chat in the console:

1. Set yourself Available and choose how many chats you can hold at once. Your availability fades after a day, so keep the console open or set it again.
2. Waiting chats appear in the inbox, oldest request first. Claim one; exactly one person wins a chat, a colleague who was a moment late simply picks the next.
3. Read what came before. The chat window shows the conversation as it stood while the visitor was waiting, with a marker where you joined, so you are not answering a question you cannot see. It is the most recent part of the wait rather than the whole conversation: if older messages were left out, the window says so, and the complete transcript lands on the call record when the chat ends.
4. Reply in the chat window. The visitor sees your messages as coming from a person, clearly distinct from the agent's bubbles, and everything lands on the conversation transcript under your name.
5. When you are done, hand back to the agent (it continues with your replies in its context) or close the conversation as handled.

If you ask for a suggested reply, the draft is written against the whole conversation including the part before you arrived, not just the messages since you claimed it.

Every takeover shows up on the call record: who took it over, when the visitor asked and when someone arrived (the gap between the two is how long they waited), how many human messages were exchanged, and how it ended.

Set your opening hours#

The Opening hours card on the same page answers one question: could a visitor reach a person at all right now.

With no hours set, every hour is open, which is what your project did before this existed. Turn the calendar on and give it a time zone and the windows you are staffed, day by day. A weekday you leave empty is closed, and closed days you list beat the weekly hours, so a public holiday shuts the desk even on a normal Tuesday.

Outside those hours a visitor who asks for a person is told straight away, and the agent carries on with the conversation instead of leaving them waiting for a desk nobody is at. A chat a colleague has already picked up is never cut off by closing time.

The card shows the verdict as the platform reads it, and the local time it read it at, so you can see which clock you are being judged against. An armed calendar with no open window at all is refused, because nobody could ever reach a person through it.

Collect ratings and reports#

Visitors can rate any answer in the chat with a thumbs up or down, and can report a wrong answer in one tap. A report carries the exact answer and the knowledge documents it relied on, captured at that moment, so the reviewer sees precisely what happened even if the content changed later.

Your own team can rate and report answers on any finished conversation from the call log.

Review what was reported#

Open Answer feedback in the console:

• The review queue lists every reported answer with its sources. Decide each one once: confirmed wrong (the answer really was wrong), correct (it was fine), or fixed (you corrected the knowledge content, so the answer changed). You can link the knowledge document that caused it.
• Analytics show satisfaction per agent: how many ratings, the up and down split, and the worst rated answers to look at first. Ratings visitors leave on your team's human replies are kept but never counted against the agent.
• Every answer you confirm wrong becomes a regression case, exported in a stable format so future evaluations can prove the mistake stays fixed.