All guides
WhatsApp

Set up the WhatsApp AI assistant

Turn on an AI agent that answers inbound WhatsApp messages on your number, give it what it should know, test it properly before customers see it, and control when it hands over to a person.

Connecting your number gets messages into the CRM. This is the next step: an AI agent that answers those messages instantly, from what you have told it about your business, and steps aside the moment a person takes over. Setting it up takes about twenty minutes, most of it spent writing down what it should know.

Before you start

  • The AI assistant is on the Business plan. See pricing.
  • Your WhatsApp number must already be connected. See Connect your WhatsApp number.
  • AI replies are paid from your Wallet, which is separate from your plan. An empty wallet means the agent stays silent rather than sending something half-finished.
  • Have your prices, timings and the answers you repeat every day ready to paste. This is the part that decides whether the agent is useful or embarrassing.

Turn it on

  1. Open Marketing, then Integrations, and find your number on the WhatsApp card. Next to it is an AI reply button. A lock on it means the plan does not include AI yet.

  2. Click it. If you have no agent yet, the panel offers to build one on the spot: give it a name your customers will see, such as Priya, and paste what it should know into What it should know. One click creates the agent and switches it on.

  3. Set the three choices on the same panel once an agent exists:

    • Auto-reply. On means the agent answers new messages on this number. Off means they wait for a person.
    • Agent. Which agent answers on this number. One agent can serve more than one channel.
    • Reply language. Match the visitor replies in whatever they wrote, Always Hinglish is Roman-script Hindi, and Always English is exactly that.
  4. Decide what happens when a human replies. Either the AI stays off on that thread for good, or it resumes after a number of minutes you set. Resuming after a while suits most teams. Never means a thread a person has ever touched stays AI-free permanently.

The button reads AI on once it is live.

Give it something to know

There are two places to put knowledge, and they do different jobs.

Knowledge on the agent itself is the text you paste into the builder, under AI, then AI Agents, then Agents. It is always in front of the model on every message, which makes it the right home for the things that must never be got wrong: your prices, your timings, what you do and do not sell. It has a size limit, so keep it to the facts rather than prose.

A Knowledge Base is for the long material. Go to AI, then AI Agents, then Knowledge Base, create one, and upload files. PDF, Word, Excel, CSV, plain text and Markdown are all read. Each file shows processing, then ready, or failed if it could not be read. Attach the knowledge base to the agent and it gets searched when a question needs it.

Use both: the short critical facts on the agent, the long reference material in a knowledge base.

What it does when it does not know

This is the part worth testing hardest, because a confident wrong answer about price is worse than no answer.

The agent runs in a mode that only lets it state facts that appear in its knowledge or in what it finds in the knowledge base. Asked something outside that, it says the team will follow up rather than inventing a figure. It will not invent discounts, offers, refunds or commitments that you have not written down.

That protection is only as good as what you gave it. A price that is nowhere in its knowledge is a question it has to deflect, so anything you would hate to see guessed belongs in the agent’s own knowledge, spelled out.

Test it before your customers do

Message the number yourself from another phone and work through these. Each one checks something different.

Ask itWhat good looks like
An ordinary question you get every dayAnswers it, in the language you chose
Something priced, exactly as a customer would askThe real figure, matching what you wrote down
Something deliberately not in its knowledgeOffers to have the team follow up. It must not invent an answer
The same question in Hindi or HinglishHandled, per your language setting
“Are you an AI?”Says yes. It is set up to be honest about that
Push for a discount it was never givenDeclines and offers to connect you to the team

Then test the handover: reply to that thread yourself from the CRM as a human would, and send another customer message. The agent should go quiet, and come back only after the window you set, if you set one.

When it stays silent

Silence is usually deliberate. In order of how often it is the cause:

WhyWhat to do
A person is handling that threadWorking as intended. It resumes after your window, if you set one
Auto-reply is off, or no agent is selected on the numberOpen the AI panel on the number
The message was not text, such as an image or a voice noteNothing. The agent answers text
The plan no longer includes AICheck your subscription
The wallet is emptyTop it up. It never mentions billing to your customer
Daily limits for that conversation or that number were reachedWorking as intended. They exist so one sender cannot flood you

Keep an eye on it

Read the real threads in your inbox for the first week. Synthetic tests find the problems you expected; actual customers find the ones you did not. Every reply the agent sends is in the thread like any other message, so you can see exactly what it said and correct its knowledge where it was thin.

Still stuck on this step?

Send us what you are seeing and we will sort it out with you.