Your agent on WhatsApp

The agent that answers your calls is the same one that replies on WhatsApp: same script, same actions, same data. There is no second agent to configure, and the only thing that changes is how it writes. If the chat needs something different — another script, another voice, fewer actions — you can separate it per channel, and below is how.

What changes when the same agent writes

A script written to be spoken out loud reads badly. In writing the agent adjusts the form, not the content:

• It writes short. Two or three sentences per message. If there are options or steps, it uses a brief list instead of a long paragraph.
• It does not spell things out. It writes "August 25, 10:00" instead of "the twenty-fifth of August at ten in the morning".
• It does not talk about the call. No "before I hang up" and no "thanks for calling": it says goodbye the way you would in a chat.
• It does not rush. An hour can pass between one message and the next, and in a chat that is normal: it does not treat silence as the end of the conversation.
• It answers once and waits. It does not send three messages in a row.

What you can separate per channel

In the agent editor, above the tabs, there is a control with Phone and WhatsApp. It is a lens over the same agent: switching it edits nothing, it only changes what you are looking at. With the lens on WhatsApp, three tabs show you what is inherited from the phone and let you separate it.

• Identity → Script. By default it reads *Inherits the phone’s, plus the add-on*. Write your own copies the phone script over so you start from what already works, and from then on the two are independent.
• Voice & language → Voice. The voice for its voice notes. By default it is the same as the phone’s, and Use another voice picks a different one for the chat only. Calls keep theirs.
• Tools. By default it offers the same ones as the agent. Choose which lets you turn off the ones that make no sense in writing, without unconfiguring them: they keep working on the phone.

While something is inherited the control is not even rendered: there is nothing to touch by accident. And anything you separated goes back with Inherit again, which discards the channel’s own value and restores the phone’s.

With its own script the add-on stops applying and the screen hides it: whatever the add-on used to adjust is already written inside the script, and keeping both would produce a third text that cannot be read in full anywhere.

When it answers from your knowledge base

If the agent has a knowledge base attached, in writing it answers with what is in it and in its script, and nothing else: it does not fill in a catalog with what a business like yours "usually has", it does not build links that are not written in your material, and if asked something that is not there, it says it does not have that information and offers the contact that appears in your material. When the search finds nothing about what it was asked, the agent knows it, and answers that it does not have that information instead of filling in. If the base holds your whole site, it can send the link to the exact page: always an address that is in your material.

The same review keeps running after linking:

• If an agent that already handles a number ends up in that state when you save (you removed its base, or added "check the website" to the script), the change is not saved: the list of what to fix appears at the top of the editor, each item with a button that takes you to the tab where it is fixed.
• Nor can you delete a base if a WhatsApp agent would be left without any and its script promises that information: the screen tells you which agent depends on it. Unlink the number from the agent or fix its script and try again.

When the agent offers options

When the answer it needs is short and closed — confirm or cancel, pick a time, pick a branch — the agent sends options the person taps instead of a question they have to type out.

• Up to 3 options show as buttons. From 4 to 10, as a dropdown list. The count decides it; there is nothing to configure.
• When the person picks, the agent receives exactly the option they tapped, not an interpretation of what they wrote. That is the difference between "I understood they said yes" and "they tapped Confirm".
• After offering options the agent writes nothing else: it is the person's turn to choose.
• In the inbox you will see the options that were offered and, below, which one they picked. They look different from a typed message, because the person did not write it.

The agent does not use options for everything: not for open questions ("how can I help?") and not when it needs something that has to be typed, like a name or an address.

Blue ticks and the typing indicator

When the agent takes the conversation it marks the message as read and shows that it is typing. Both happen when they are true: as it starts replying, not when the message arrives.

• If your plan blocked the turn, it is not marked as read: that would tell the person someone saw it when the platform decided not to handle it.
• If you took the conversation yourself, it is not marked either: you are the one reading, and we do not know when.
• The typing indicator turns off on its own once the reply goes out.

When the agent sends a button with a link

The agent can send a message with a button that opens a web address — the details of an appointment, for example. It must start with https://.

How long it takes to reply, and why

People write in several bubbles in a row: "Hi" · "I need to change my appointment" · "the one on Tuesday". That is three messages, but one single idea.

So the agent does not reply to the first bubble: it waits a few seconds after the last message it received and answers all of them together. Every new message restarts that wait.

How long it waits depends on how the last message ends:

• A message that looks finished —a question with its mark, a "yes", a "no", a "thanks"— goes out in about 3 seconds.
• If they tapped one of the options the agent offered, it replies right away.
• A lone "Hi", a sentence that was cut off, a photo without text or a voice note wait about 6 seconds, in case they keep writing.
• If the person keeps sending messages non-stop, they still get an answer: half a minute after the first one, the agent replies with what it has.

• Without that wait, those three bubbles would produce three overlapping replies and a conversation nobody can follow.
• The inbox shows you the countdown and the reason: "The agent replies in 6s" and, below it, why it waits that long. When it starts writing, it changes to "The agent is replying".
• If another message arrives while the agent is writing the reply, that reply does not go out: it is discarded and the agent answers everything together. The exception is when it already booked, moved or cancelled an appointment: that confirmation goes out anyway, because it goes with something that already happened.
• If someone on your team answers from the inbox while the agent is waiting or writing, the agent does not reply on top of them.

When the agent does not reply

Each case is different and is solved differently. The most common ones are:

• Someone on your team took the conversation. While it is set to "Me", the agent stays out.
• The person asked not to receive more messages. If they write "stop", "unsubscribe", "remove me" or one of the other opt-out words we recognize, we stop writing to them: neither the agent nor your team. Their message still reaches the inbox.
• More than 24 hours have passed since their last message. That is a WhatsApp rule, not ours: outside that window nobody can reply with free text and an approved template is required.
• The conversation is closed. The agent closes it when the topic is finished, and also when a voice call where people talked ends. If the person writes again, a new conversation opens and the agent does reply there.
• The number has no agent assigned. Messages arrive and wait for your team.
• The channel already used its share of the plan. The agent stops replying in writing so you keep balance for calls (see below).
• Your account is suspended or out of balance. It is the same block that applies to calls: while it is active, the agent does not reply on any channel.

If the conversation was closed and the person writes again, a new conversation opens in the inbox. The previous thread stays as it is, with its history.

Taking the conversation

At the top of every conversation there is a control that says who answers: The agent or Me. Set it to "Me" and the agent stops replying in that thread. The same button, Take the conversation, also appears in the notice right above the text box, which is where you are looking when you decide to step in.

• When you take it, the conversation is assigned to you, so two people on your team do not overlap.
• When you hand it back to the agent, the assignment is released and the agent replies to the following messages again.
• The notice is a notice, not a wall: you can write without taking the thread. Stepping in is sometimes exactly right.

What the agent can do in writing

• Reply using its script and whatever it has loaded in its knowledge base. With large bases it looks up the fragment that matches each question, the same as on the phone.
• Read the recent history of the thread: the last messages from the last few days. If there is older history it can no longer see, it knows, and it will not claim this is the first time you talk.
• Tell who said what: if a person on your team replied earlier, or if you wrote from your phone, the agent sees it and does not contradict it.
• Know that a photo, an audio, a document, a location or a contact arrived, with its type and its name.
• Treat the conversation as finished when the topic is resolved: it says goodbye and only then closes it.

What it cannot do yet

• See images. It does listen to audio —voice notes are transcribed automatically and it answers what you said— but with a photo it only knows one arrived and what it is called. If the case depends on that photo, a person should take it.
• Query your systems. Actions that reach out to a website or an external system are not available on this channel yet.
• Hand the conversation to another agent. Handing it to a person, yes: that is the "who answers" control.
• Send files or photos. It replies with text and, if you spoke to it, with a voice note plus that same text. Attaching a file is not available yet.
• Write first. The agent only replies: it never starts a conversation on its own. It is started by the person, by a campaign, or by someone on your team from the inbox with an approved template.

What it uses from your plan

The unit is the agent reply, not the message. If it answers three bubbles with a single message, that is one reply.

• It does not use anything when a person on your team writes from the inbox.
• It does not use anything to receive messages.

Agent replies draw from the same balance as your calls and your meetings: one pool, one usage bar. How much each reply draws depends on your plan; if your plan does not have a rate for this channel yet, replies draw nothing.

There is a limit on purpose: the text channel can use at most half of your plan balance in each cycle. Once it reaches that ceiling the agent stops replying in writing, and your phone keeps working. It frees up on its own when the cycle closes, or when you upgrade your plan.

On the free trial

During the free trial, each text reply from the agent draws 3 seconds from your trial balance: the same balance as your calls and your meetings. Used only in writing, the 20 trial minutes are good for 400 text replies; if you also make calls, each minute of calling is 20 fewer replies. You can see how many you have left on the WhatsApp screen and on the meter in the side menu. The half-of-the-balance limit does not apply here: you can spend the whole balance in writing.

Audio draws more, depending on its length. When the agent replies with a voice note, or has to listen to an audio message you received, that turn also draws what the audio lasts — just like a call, which is charged for how long it lasts. A short voice note from the agent is about 20 seconds of balance. If you want your trial to go further, set Voice notes to Text only in the agent's WhatsApp settings.