Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Channels

AI Support on WhatsApp Business: A Practical Guide

AI SNS CHAT EditorialPublished 8 min read

How to set up AI replies on WhatsApp Business: prepare knowledge, set guardrails, hand off to humans, respect the 24-hour window and measure results.

Customers message businesses on WhatsApp the way they message friends: short, informal, at any hour, and they expect a quick answer. A small team cannot staff that around the clock, but an unprepared bot that guesses will damage trust faster than a slow reply. This guide covers how to put an AI agent on WhatsApp Business so it answers what it knows, hands over what it should not decide, and gets better week by week.

Why WhatsApp behaves differently from website chat

A website chat widget sits on a page you control. The visitor is already looking at your product, and the page gives context. WhatsApp is different in four ways that shape your AI setup.

  • The conversation is persistent. The customer returns to the same thread days later. Your agent must cope with "Hi, any update?" without any page context.
  • Messages are short and fragmented. People send three one-line messages instead of one clear question.
  • Media is common. Photos of damaged goods, screenshots of errors and voice notes arrive often. Decide in advance what your agent should do when it cannot interpret something.
  • The channel has rules. WhatsApp limits when a business can send messages, and the customer expects not to be spammed.

Treat WhatsApp as a conversation channel with memory and etiquette, not as a widget in a different colour.

Decide what the AI should handle

Before connecting anything, sort your incoming questions into three groups. A simple exercise is to read the last few weeks of chats from any channel and tag each one.

GroupExamplesAI role
Answerable from documentsOpening hours, delivery times, return policy, how to reset a settingAnswer directly, in your tone
Needs a lookup or a formOrder status, booking a slot, collecting contact detailsAnswer using an integration, or collect details and pass them on
Needs a personRefund disputes, complaints, legal or safety issues, anything emotionalHand off promptly with context

The first group is where an AI agent earns its keep. The third group is where it must step aside gracefully. The middle group depends on your integrations, so start with the first and third and add the second once the basics are stable.

Prepare the knowledge before you connect

The quality of AI answers is mostly the quality of the material behind them. Gather what your best human agent already uses.

  • Policy pages and FAQs
  • Product or service descriptions with current prices and conditions
  • Shipping, returns and warranty rules
  • Internal how-to documents that explain the real process, not the marketing version

Keep it current. If your return window changed last month, update the source now, because the agent will repeat whatever it reads. Write answers in complete sentences where you can, and put exceptions next to the rule they modify rather than on a different page.

Also write down the questions the agent should refuse or defer. "Can you guarantee this will arrive by Friday?" is a question where the right answer is often a careful one, not a promise.

  1. 01Customer messages on WhatsApp
  2. 02AI answers from your knowledge
  3. 03Unsure or sensitive? Hand off
  4. 04Person replies, AI learns from edits
From first message to resolved or handed offIllustrative example

Write a persona and guardrails that fit a chat app

On WhatsApp, a wall of text looks wrong. Define the voice explicitly.

  • Length: a few short sentences per reply, with the most useful fact first.
  • Tone: match your brand, but lean friendlier than your website copy.
  • Boundaries: list what the agent must never do, such as quoting prices that are not in the knowledge base, giving medical, legal or financial advice, or promising delivery dates.
  • Fallback behaviour: say what happens when the answer is not known. A good fallback is honest ("I don't have that information, let me bring in a colleague") and triggers a handoff.
  • Questions back: allow one clarifying question when a request is ambiguous, but not an interrogation.

Test with real messages, including the awkward ones: typos, mixed languages, angry customers, and attempts to make the bot ignore its instructions. Prompt-injection and abuse filters are worth switching on for any public-facing channel.

Hand off to humans without losing the thread

A good handoff has three parts: the right trigger, the right person, and an honest message to the customer.

Triggers

Hand off when the customer asks for a person, when sentiment turns negative, when the topic is on your sensitive list, or when the agent has failed to resolve the issue after a couple of exchanges. Do not make customers fight to reach a human.

Routing

If your team works in shifts across time zones, route to whoever is on duty rather than to a shared inbox nobody watches at night. Alert the team where they already work, such as Slack, Teams, email or a webhook.

The message to the customer

Never tell a customer "someone will reply in five minutes" unless that is true. The better pattern is to state what happens next, based on your actual schedule: a colleague will pick this up during working hours, or the team is available now. Invented wait times are worse than none.

Once a person takes over, the bot should stop replying in that conversation so the customer does not get two voices. When the person finishes, review what they wrote. If their edited reply is better than the AI's, feed it back as the ideal answer so the same question goes better next time.

Respect the 24-hour window and consent

WhatsApp's business rules distinguish between replying to a customer who recently wrote to you and starting a conversation yourself. In general, free-form replies are allowed within a 24-hour window after the customer's last message. Outside that window, WhatsApp expects approved message templates.

For AI support this has practical consequences.

  • Design for the customer to start the conversation. Support is naturally reactive.
  • If you follow up proactively (an appointment reminder, a restock notice), do it only for people who have messaged you first, keep it one-to-one, and include a clear way to opt out.
  • Do not use WhatsApp as a mass-marketing broadcast tool. It harms deliverability, annoys customers and can put your number at risk.

A good rule: if you would not be comfortable explaining to the customer why they received the message, do not send it.

Measure, then improve weekly

You do not need a dashboard full of metrics. Review a small set every week.

  • Unanswered or wrongly answered questions: these show gaps in your knowledge sources.
  • Handoff reasons: if one topic dominates, add documentation or a lookup for it.
  • Customer ratings: thumbs up and down on individual replies point you at the exact answers to fix.
  • Sentiment and topics: these show what customers care about and where frustration builds.
  • Credit usage: know how many replies you consume so the plan fits.

As a purely hypothetical example, suppose a shop receives 300 WhatsApp conversations a month and the AI sends an average of four replies per conversation. That is about 1,200 replies a month. If each reply costs one credit, the shop needs roughly 1,200 credits, which is useful to know before choosing a plan.

Hypothetical shop usage
1200
Hobby plan credits
500
Standard plan credits
4000
Example: hypothetical monthly credit need vs plan allowanceIllustrative example

A weekly routine of fifteen to thirty minutes is enough: read the flagged chats, fix the source document, and re-test the question in a sandbox.

Doing this in AI SNS CHAT

AI SNS CHAT lists WhatsApp among its deployment channels, alongside website chat, Messenger, Instagram, LINE, Telegram and others. Here is how the workflow above maps onto the product.

  1. Build the agent. Create an agent with the guided wizard, or start from one of the starter personas. Write the persona and guardrails, including what the agent must not promise. Use the Playground to test awkward messages before going live, and the Playground A/B model compare if you want to see two models answer the same question.
  2. Add knowledge. Under Data sources you can add files, text snippets, a website, Q&A pairs and more. Connectors under Channels can pull from tools such as Confluence, GitBook, Zendesk Guide, Google Drive and Notion, so you do not have to copy documents by hand.
  3. Connect WhatsApp. Open Sidebar → Channels, choose WhatsApp and follow the connection steps shown there. Incoming webhooks from Meta are checked with signature verification. For an overview of supported connections, see /connections.
  4. Choose the model. Under Agent → Settings → AI, pick the provider and model. Advanced models and model fallback need the Hobby plan or higher. If one provider is unavailable, replies can fail over to another in a fixed order.
  5. Set up handoff. In Helpdesk → Scheduling, define shifts per person and time zone. Handoffs are auto-assigned to whoever is on shift, and alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. The message the customer sees is built from your rota, and the server does not invent a wait time. Helpdesk needs the Standard plan or higher.
  6. Take over and teach. WhatsApp conversations appear in Activity → Chat logs (the Inbox is for email). From there you can take over a chat, which pauses the bot, use live translation, insert saved replies, and edit a reply into the ideal answer to train the agent. You can also filter, flag, escalate, resolve and export, and review thumbs up and down ratings.
  7. Capture details. The Collect leads action and custom forms store details under Activity → Leads, and each lead also creates a contact.
  8. Follow up carefully. Outbound Beta (Standard or higher) sends only to people who messaged first, one-to-one, with a STOP opt-out on every message. WhatsApp sends are limited to the 24-hour window. There are no SMS or email marketing campaigns.

Billing is simple: one reply uses one credit, and usage per agent is on the Usage page. See /pricing for current plans and /features for the full list. Note that WhatsApp voice calling is not available yet; it is listed as coming soon. For data handling, see /security.

When you are ready, you can start on /signup and test the agent in the Playground before connecting any live channel.

Frequently asked questions

Can the AI send WhatsApp messages to customers who have never contacted us?

In AI SNS CHAT, no. Outbound Beta on chat channels reaches only people who messaged you first, one-to-one, and every message carries a STOP opt-out. WhatsApp sends are also limited to the 24-hour window.

Where do I see my WhatsApp conversations?

In Activity → Chat logs. The Inbox in AI SNS CHAT is email-only. In chat logs you can filter, flag, escalate, resolve, export and rate conversations, and take over any chat so the bot pauses.

Will the agent invent a response time when it hands off?

No. The server reads your Helpdesk rota and adds the line about what happens next itself, so the customer is not given a made-up wait time. Shifts are set per person and time zone under Helpdesk → Scheduling, which needs the Standard plan or higher.

Can customers call my business on WhatsApp and talk to the AI?

Not yet. WhatsApp voice is listed as coming soon, with a waitlist only. Text conversations on WhatsApp are supported today, and AI phone calls are available separately through a Twilio number.

How many credits will WhatsApp support use?

One reply uses one credit, and credits reset each billing cycle. Estimate by multiplying your monthly conversations by the average number of AI replies per conversation, then compare that with the plan allowance on the pricing page.

#WhatsApp#Channels#AI support#Human handoff

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.