Chat control

Can you review chat replies before they go to the visitor?

You cannot review every chat reply before it goes to the visitor, and that is deliberate: the visitor is waiting in front of the bubble, so a Muzau agent's chat replies on its own. Each reply is checked automatically before it is shown: a third party's email address, an external link, the name of an internal file or a promise nothing backs up will hold it back, and your team receives it. Upstream, you set what the chat knows and what it does on its own.

Why the chat cannot wait for approval

An email can wait an hour in "To approve": the customer will read the reply later. In the bubble, the visitor has just typed their question and is watching the screen. With a review step, the one writing in the evening would wait until the next day, then close the tab. It is by replying on its own that the chat can answer your website visitors at any hour.

So the review moves to a different moment: before, you set what the chat knows and does on its own; during, an automatic check holds back what should not be shown; after, you read the conversations and fix the source.

How the agent goes about it

Input

The reply the chat has just written, before the visitor reads it

It does

Rereads it, looks for specific signals, holds back any reply that carries one

It delivers

A reply shown within seconds, or a card for your team

The signals that hold back a chat reply

Each chat reply is checked automatically before it is shown. It is held back in particular if it contains a third party's email address (neither the visitor's nor your business's), a link to another site, the name of one of your internal files or a promise nothing backs up. This check backs up an instruction given to the chat: announce no price, lead time, availability or commitment that is absent from your Brain. It looks for specific signals; it does not judge whether the reply is accurate.

When a chat reply is held back, the visitor is told and the team receives it

The held reply is not shown. The visitor reads "I would rather have the team check my reply before answering you. They will get back to you very soon.", and is asked for their contact details if they have not left them. An appointment already booked stays confirmed. Your team receives a card with the question, the held reply, the whole conversation and a reply prepared by the agent. They send it by email to the address left, or call back the number left. If the held reply was right, it only costs that one step; if it was wrong, the visitor never read it.

Upstream, you set what the chat knows and what it does on its own

The chat answers from your Brain and the pages of your site that Muzau has read. What you keep for the team does not appear in its replies. At creation, you choose what it asks the visitor and what it passes on to you. On the Autonomy page, you decide whether it books an appointment or sends an email on its own; otherwise, the appointment request goes to your team. If a visitor writes an IBAN, the chat books no appointment and sends no email on its own.

You try the chat before opening it, then you reread its conversations

Before going live, you try the chat from the cockpit: you see what a visitor would read and what the agent would have done, with no appointment booked and no email sent. Afterwards, the agent's page shows the last 20 conversations in full. From the card, your team can add to the Brain the answer that was missing. See what the chat does when it does not know the answer.

Reviewing the conversations of a chat that replies on its own: the checklist

Whatever the tool, first read the replies held back or passed to the team, then the conversations where the visitor left contact details, and finally a few conversations of one or two messages: they show what makes people leave. For each conversation, ask six questions.

• Did the first question get its answer? From the first message, not after three detours.
• Can every figure quoted be found somewhere? A price, a lead time, opening hours: if it is missing from your site and your rules, the chat inferred it.
• Was what the chat announced actually done? A callback, an email, an appointment: check who took care of it.
• Where did the visitor stop? Read the last reply before they left: either it was enough, or it did not help them.
• Did a request stay in the bubble? A quote, an appointment, a complaint: check that the team received it.
• Is the tone yours? Formal address, length, a phrase that keeps coming back from one reply to the next.

Note each gap on one line: the question, the chat's reply, the right reply, the cause (a vague page, a missing rule, missing knowledge). Reread every week, on a fixed day, and after every change of price, offer or page. For an AI's mistakes beyond the chat, see whether an AI writing to your customers can get it wrong.

What it does not do

• There is no mode where you review every chat reply before it goes out: only a reply held back by the automatic check waits for you instead of being shown.
• The automatic check looks for specific signals, such as an address, a link, an internal file or a promise: it does not judge whether the reply is accurate.
• The chat does not learn on its own from its conversations: a missing answer goes into the Brain through an action by your team.

Frequently asked questions

Is there a chatbot with human approval before every reply?

A chat that waits for human approval before every reply makes the visitor wait in front of the bubble until someone is free. A Muzau agent's chat therefore replies on its own, and only a reply held back by the automatic check waits for your team. For emails, however, you can choose when creating the agent to review everything before sending.

How do you control the replies of an AI chatbot on your website?

Control happens before and after the conversation: you write what the chat knows and what it never promises, then you reread its conversations. A Muzau agent's chat adds an automatic check of each reply before it is shown: a third party's email address, a link to another site, the name of an internal file or a promise nothing backs up will hold it back for your team. This check looks for specific signals; it does not judge whether the reply is accurate.

How often should you reread your chatbot's conversations?

Reread the chat's conversations every week, on a fixed day, starting with those passed to the team or held back before display, and more often after a change of price, offer or page. On a Muzau agent's page, the 20 most recent chat conversations can be read in full, your own tests marked "Test".

Can a chatbot reveal the company's internal information?

A Muzau agent's chat never says what you mark as reserved for the team in the Brain: it treats every visitor as an outsider, even one who has left their email. A reply that contains the name of one of your internal files is held back before it is shown. However, what it reads in a spreadsheet you have it consult, a schedule or a stock list, can feed its replies: that is your choice.