WhatsApp assistant

People message you at ten at night, at the weekend, and while your team is on a call. The assistant answers them — in their language, using what you already know about them — or writes the reply and waits for you to press send. It is not a chatbot menu. It reads the conversation and the contact’s record, answers the question, and hands over to a person when the answer is not one it should be giving.

​

One assistant per number

The number someone messages already says which desk they reached: a market, a language, an audience. So each Cloud API number gets its own assistant, with its own name and its own instructions. That is what lets a Turkish line and a Malaysian line answer completely differently, and it is what lets you switch one on while the other stays off. Find them under Settings → Agents. Every Cloud API number is listed whether or not it has an assistant yet.

​

Off, Suggest, Live

Every assistant is in one of three states, set on its row or on its page.

• Off — nothing happens. This is where every number starts.
• Suggest — when a contact writes, a draft reply appears in the conversation, visible only to your team, with Send and Discard. Nothing goes out on its own.
• Live — it replies by itself, about fifteen seconds after the contact stops writing.

Start on Suggest. Read a week of drafts before you let anything reach a customer by itself. Suggest is useful on its own — a reply written and waiting is faster than one you have to think of — and the assistant’s page shows what share of its drafts your team actually sent, which is the number that tells you whether it’s ready.

​

Writing its instructions

Click the assistant’s name to open its page. Instructions is where you say how this desk answers: who it’s talking to, what it should offer, what it must never say, and when to get a person involved. Instructions do not save as you type. A half-written sentence would otherwise be live on the very next message a customer sends. The page says when you have unsaved work and offers Save or Discard changes. A short brief that is specific beats a long one that is vague. The most useful lines are usually the refusals — the figures it may not quote, the promises it may not make.

​

What it knows

You don’t have to tell it who it’s talking to. Before it writes anything it already has:

• The contact’s record — their name, and the details on their card, in plain wording rather than raw values.
• What your calls collected — the answers your voice agents captured, plus a short summary of recent calls. A chat can pick up where a phone call left off.
• The last ten messages in the conversation.
• What the business knows — see below.

It deliberately does not see your internal notes, the contact’s status, their emails, or full call transcripts. Those are how your team talks about someone, not how you talk to them.

​

What the business knows

Under the instructions is a shared list of facts: opening hours, parking, refund policy — the answers you give twenty times a week and that don’t live in a field or a template. It is shared by every assistant, so write it once rather than into each one’s instructions. Don’t retype things Nudge already holds, like a contact’s own details: prose that repeats a record can contradict it, and the contradiction is invisible until someone is told the wrong thing. Write it in whatever language you think in. The assistant replies in the customer’s.

​

Saved messages it can send

The assistant can send your saved messages — a price list, a booking link, a brochure — but only the ones you allow. On a WhatsApp template, tick The assistant can send this. You can add one plain line saying when it applies, like “send when they ask about opening times”. That is guidance, not a keyword rule: “what time do you open” and “are you open Sundays” both land on it. Nothing is offered by default, and that is the point. Most of a library is outreach copy, and one of those going out in the middle of a live conversation is exactly what this prevents.

​

When it hands over

Sooner or later someone asks for something the assistant shouldn’t answer — an exact figure, a commitment, or simply a person. When it hands over decides what happens next.

• Keep answering — it flags the conversation for your team, says so once, and carries on helping with everything else. Not being able to quote a price is no reason to stop answering the opening hours.
• Stop replying — it says a colleague will follow up and then goes quiet until someone hands it back. The right choice when half an answer is worse than none.

On Keep answering, a flagged conversation shows a line above the composer saying it needs a person, and when. It says it once — a conversation that announces a colleague on every message is the thing this is designed to avoid. Either way, a person typing in the conversation takes it over immediately. No setting, no confirmation: the reply is the instruction.

​

Changing it for one conversation

The setting on the number is a default. Any single conversation can disagree with it, from the strip above the message box: the same Off / Suggest / Live, for this chat only. That covers the ordinary cases — a delicate conversation you want to write yourself on an otherwise live number, or one straightforward chat you’re happy to let run while the rest are drafts. To put a conversation back to normal, pick whatever the number itself is set to. There’s no separate “back to default” button because there doesn’t need to be: a conversation you never touched follows the number, including when you change it later, and one you set by hand keeps your choice.

​

Introducing itself

Introduce itself is off by default. Switch it on and the first reply a contact ever gets opens with a line naming the assistant as a digital one. Whatever you choose, it will never claim to be a person, and if a contact asks outright whether they’re talking to a bot it says so plainly and carries on helping. That part isn’t a setting.

​

Trying it before anyone sees it

Try it on the assistant’s page talks to the real assistant with your saved instructions. Nothing is sent and no contact is involved. It answers “does this sound the way I want”, which is the question you ask twenty times while writing a brief. It does not answer “will this handle my customers” — there’s no contact record behind it. Suggest mode on real conversations answers that one.

​

What it costs

Replies are charged by the AI provider you connected, per message, and it is small — fractions of a penny for a typical reply. Recent replies on the assistant’s page shows every turn with its cost, and the month’s total against your budget. The budget is workspace-wide, because the bill is. It’s checked before each reply, so the ceiling can be passed by at most one reply’s worth rather than found out about afterwards.

​

What it won’t do

Worth knowing, because these are deliberate:

• It never changes the contact’s record. If someone tells it something worth keeping, it acknowledges and carries on — your team reads the conversation and records it.
• It answers at most ten times an hour per contact. A loop can’t run away with your budget or your customer’s patience.
• It waits about fifteen seconds after the contact’s last message, so three messages typed in a row get one answer to all of them rather than three replies to a question nobody finished asking.
• It can’t see attachments. Sent a photo or a voice note, it says so and gets a person rather than guessing.
• If the AI provider fails, the contact gets nothing — never an error message from a system they didn’t know was there. The conversation just stays unread and your team answers it.
• Meta’s 24-hour rule still applies. The assistant answers people who have messaged you, so it is almost always inside the window — but a reply that lands outside it needs an approved template like any other. See WhatsApp.

​

Getting started