A WhatsApp AI agent is an AI that runs customer conversations on your business WhatsApp number: it reads each message, finds the answer in your own information, replies in the customer's language and, when allowed, takes an action such as booking a slot or logging a lead. It differs from a classic chatbot because nobody scripts each reply, and it differs from a person because it answers in seconds at any hour. This guide covers what it can do, where the risks sit, how to set one up safely and what to test.
What an AI agent does in a conversation
A typical customer message goes through five steps.
Read. The agent reads the text, and in better products also voice notes, photos and documents the customer sends.
Find. It searches your knowledge, which can include your website, uploaded files and connected systems, and says when it does not know.
Decide. It chooses between answering, asking a clarifying question, taking an action or passing the chat to a person.
Act. It books, quotes, updates a record or asks a colleague, within the limits you set.
Reply. It writes the answer in the customer's language and tone.
Compare that with a WhatsApp chatbot, which usually matches the message to a script and skips the finding and deciding. The practical difference shows when a customer asks something outside the script, such as "can I move my appointment to Friday and add my sister?" An agent that can see your calendar handles it. A bot that only knows a menu says it did not understand.
Jobs it does well
First response outside working hours, so no enquiry waits until morning.
Repeat questions about prices, opening hours, delivery, policies and how to prepare for a visit.
Lead capture and qualification: asking for name, need, budget or date and writing it to the contact record.
Booking, rescheduling and confirming appointments in a connected system.
Reading attachments, such as a photo of a damaged item or a PDF requirement list, and replying about them.
Routing: tagging each conversation and handing it to the right person with the context attached.
Jobs that stay with people
Refunds, discounts and exceptions to policy, unless a person approves each one.
Upset customers and sensitive topics, where tone matters more than speed.
Medical, legal or financial advice that your regulator expects a qualified person to give.
Large orders and negotiation, where a relationship is being formed.
The skill in setting up an agent is writing this division down before launch, so the agent knows when to step back.
The rules around AI on WhatsApp
Meta sets conditions that apply to any agent on the WhatsApp Business Platform.
You can automate replies inside the 24-hour customer service window, which opens and resets with each customer message. You must also offer a prompt, clear and direct way to reach a person: in-chat human transfer, a phone number, an email address, web support, an in-store visit or a support form.
Outside the window, you can send only approved templates, and only to people who opted in.
Meta's terms updated on January 15, 2026 say AI providers may offer general-purpose AI assistants on the Platform only where Meta is legally required to allow it. A business agent focused on its own customers is the use case the Platform is built for. Keep yours on your own topics.
Repeated policy violations lead to restrictions and, in the end, removal. What to do when a WhatsApp Business account is banned explains the ladder.
Also check local rules on telling people they are talking to an AI. Some jurisdictions expect a disclosure, and it builds trust where they do not.
Risks and the controls that reduce them
| Risk | Control to ask for |
|---|---|
| The agent states a wrong price or policy | Answers drawn from your approved sources only, with a clear "I do not know" when nothing matches |
| It promises something it should not | Approval before refunds, discounts and other high-impact actions |
| A customer cannot reach a person | One-tap human takeover, and the agent stops replying once a person joins |
| It repeats a mistake | Corrections saved as lessons that apply to every later chat, with a record of who approved them |
| Private data goes somewhere unexpected | Clear answers on where data is stored and which model hosts see messages |
| It messages people who did not ask | A rule that the agent replies to inbound chats and sends templates only to opted-in audiences |
Setting one up in seven steps
Pick one goal for the first month, such as answering enquiries after hours or booking trial appointments.
Connect your WhatsApp number through the official Platform. If you want to keep using the WhatsApp Business app on the same number, choose coexistence.
Give the agent knowledge: your website, price list, policies and FAQs in files or documents you keep current.
Write instructions in plain language: tone, what it should ask first, what it must never do and when to hand over.
Connect the systems it needs to act, such as a calendar or customer records.
Run in draft mode, where the agent writes and a person sends, and read every draft for a week.
Switch on automatic replies for the topics with clean drafts, keep drafts for the rest, and review unanswered questions weekly.
What to test before you go live
Ten questions with answers in your documents. It should answer correctly.
Five questions with no answer in your documents. It should say it does not know and offer a person.
A customer who is angry. It should hand over with a summary.
A customer who writes in another language. It should reply in that language.
A request that needs approval. The action should wait for a person.
A customer who changes the subject mid-chat or sends a voice note or photo.
A real example
Cake Inspiration, a custom cake studio in Singapore, has Volty qualify every lead on WhatsApp before a designer takes over. The case study reports 91% of conversations resolved by Volty without a human, across 800+ customer messages in one week, and a median reply of 5 seconds against 9 hours before. Kai Yang, Founder, Cake Inspiration, said: "I can fix the agent myself now, without waiting on an engineer. I just tell it what I didn't like about a reply, and it learns. Sometimes it will automatically suggest improvements and it will ask me to approve it. I think that's pretty cool." The Cake Inspiration case study has the full setup.

How Volty approaches it
Volty is an AI helpdesk for builders, and the agent is part of every plan. Each inbox chooses whether Volty replies on its own, drafts for your team or stays out, and you can set it to answer at night and draft during office hours. Your staff choose which actions happen on their own and which come to them first. Volty hands a conversation to the right teammate in the same chat and tells the customer who is taking over. It learns from corrections, with every lesson listed and reversible. A line above its first message tells new contacts it is an AI and a line below offers a person. The WhatsApp AI page shows each part, and pricing lists plans by market.
Frequently asked questions
What is a WhatsApp AI agent?
A WhatsApp AI agent is an AI system that handles customer conversations on your business WhatsApp number. It reads each message, finds the answer in your own knowledge and connected systems, and can act, for example by booking an appointment or recording a lead. Staff can take over any chat and decide which actions need their approval.
What is the difference between a WhatsApp chatbot and an AI agent?
A chatbot follows scripted paths or answers from a fixed list. An AI agent writes each reply from your information, copes with questions nobody scripted, and can take actions in connected systems. The line is blurry in marketing, so test what the product does with an unplanned question and a request that needs an action.
Is an AI agent allowed on WhatsApp?
Yes for a business agent that serves your own customers. Meta's policy allows automation inside the 24-hour window with a clear path to a person. Meta's terms of January 15, 2026 limit general-purpose AI assistants offered by AI providers on the Business Platform, which is different from an agent that answers about your own business.
Will an AI agent make mistakes?
Yes, occasionally, as people do. You reduce the risk by limiting its knowledge to your approved sources, running it in draft mode first, requiring approval for refunds and other high-impact actions, and reviewing the conversations it could not answer. Every correction should become a lesson the agent keeps.
How long does it take to set one up?
Connecting the number and the knowledge can take an afternoon. The slow part is testing: a sensible plan is a week in draft mode with a person approving every reply, then automatic replies for the topics where drafts went out unchanged.










