A WhatsApp chatbot follows a fixed script of menus and keywords; a WhatsApp AI agent reads the customer’s own words, checks your real data and proposes the next action. The chatbot is predictable but breaks on anything unscripted. The agent copes with open questions, but it needs guardrails so it never invents a price or a promise.
For most of the last decade, “WhatsApp automation” meant a button tree: press 1 for opening hours, 2 for order status, 3 for a person. Large language models changed what a business can automate, and in January 2026 Meta changed who may run AI on the platform at all. This guide covers both shifts in plain terms, with examples from online stores, clinics and service businesses, and ends with a checklist to decide what your number needs. If you want to see the agent side in practice, WACRM AI agents are the product this guide refers to.
What is a WhatsApp chatbot?
A WhatsApp chatbot is a set of rules that maps an incoming message to a prepared reply. The customer taps a button, picks a row from a list or types a keyword, and the bot answers with the text you wrote for that branch. Nothing is generated: every reply exists before the customer asks.
Meta’s own interactive message types shape what these bots look like. A reply-buttons message can carry at most three buttons, each label no longer than 20 characters, and a list message can hold up to ten rows across all its sections, according to Meta’s developer documentation checked on 13 November 2026. That is why so many WhatsApp bots look the same: a greeting, three buttons, then another list.
Meta’s limits on scripted menus (checked 2026-11-13)
Reply buttons: up to 3 per message, labels up to 20 characters. List messages: up to 10 rows in total. Source: Meta for Developers, interactive messages.
Where button bots still work well
Scripts are not obsolete. They are the right tool whenever the set of answers is small and fixed:
- Fixed choices. “Delivery or pickup?”, “Which branch?”, “Morning or afternoon?” are faster to tap than to type.
- Order and booking status. When a WhatsApp automation sends an order update, a button to “Track order” or “Change time” needs no intelligence.
- Opt-in and opt-out. Consent wording should be exact, every time, which is what a script guarantees.
The trouble starts when a customer types something the tree was never built for: “Do you have the blue one in a 42, and can it arrive before Friday?” A keyword bot either matches the wrong branch or replies “Sorry, I didn’t understand”, and the customer leaves.
What is a WhatsApp AI agent?
A WhatsApp AI agent is software that uses a large language model to read the customer’s message in their own words, pulls the facts it needs from your business systems, and proposes a complete action: a reply, a booking, a reschedule, or a handoff to a teammate. The key word is proposes. A well-built agent works within rules you set, not as an open chatbot that can say anything.
Three things separate an agent from a chatbot with a language model bolted on:
- It is grounded in your data. For an online store, that means the product names, SKUs, prices and stock you already keep in WooCommerce or Shopify, both of which expose that data through their own APIs (the WooCommerce REST API and the Shopify Admin API). For a clinic or salon, it is the service list, prices and staff hours.
- It acts through tools, not just text. Booking an appointment, adding a note, or assigning the chat to a person are actions with rules attached. Open standards such as the Model Context Protocol describe how AI applications connect to such tools and data sources.
- It knows when to stop. It hands the chat to a human when the customer asks for one, when it has nothing in your data to answer from, or when the action is too risky to take alone.
What is the difference, side by side?
The short version: a chatbot decides by matching, an agent decides by understanding and checking. The table compares the two approaches, not any particular product.
| Scripted WhatsApp chatbot | WhatsApp AI agent | |
|---|---|---|
| How it reads a message | Button taps, list picks, keywords | Free text in the customer’s own words and language |
| Where answers come from | Replies you wrote in advance | Your synced catalogue, service list and conversation history |
| Unplanned questions | Falls back to “I didn’t understand” | Drafts an answer if your data supports one, otherwise hands over |
| Actions | Fixed: show the next menu | Proposes replies, bookings, reschedules, cancellations, handoffs |
| Set-up work | Build and maintain every branch | Connect data, set approval rules, review early drafts |
| Main risk | Customers drop out of rigid menus | Wrong or invented answers, unless grounded and guarded |
| Best for | Fixed choices, consent, status buttons | Product questions, booking requests, mixed questions |
Rules as of 13 November 2026; WhatsApp platform limits from Meta for Developers.
In practice, the strongest setups combine both. The agent handles the open question, and buttons handle the moments where a single tap is faster than typing, such as confirming a time slot.
What happens to the same customer message?
The clearest way to see the difference is to send the same message to both. Below, a customer of an online clothing store asks a two-part question that no menu was built for.
Scripted chatbot
Customer: Do you have the linen shirt in M, and how much is it?
Bot: Sorry, I didn’t understand. Reply 1 for Orders, 2 for Delivery, 3 for Returns.
Outcome: the customer has to start again, or leaves.
AI agent, grounded in the store’s catalogue
Customer: Do you have the linen shirt in M, and how much is it?
Agent draft: Yes, the linen shirt in M is in stock at the price listed in your store. Would you like the link to order it?
Outcome: the draft quotes only the synced price and stock; a teammate approves it before it is sent.
Illustration: the same WhatsApp question, two outcomes. The product and wording are an example, not a customer conversation.
The same pattern holds outside retail. A clinic patient who writes “can I move Thursday’s cleaning to next week, mornings only?” gets a menu from a bot, while an agent proposes a reschedule into a real free morning slot for the front desk to confirm.
What changed on WhatsApp in 2026?
Two things changed: what AI can do in a chat, and Meta’s rules about who may run AI on the platform. The underlying messaging rules for businesses stayed the same.
The general-purpose chatbot ban
In October 2025, WhatsApp added an “AI Providers” section to its Business Solution Terms. It prohibits providers of large language models, generative AI platforms and general-purpose AI assistants from using the WhatsApp Business Solution when that technology is the primary functionality being offered. According to TechCrunch’s report, the rule took effect on 15 January 2026, and Meta confirmed it does not affect businesses that use AI to serve their customers, giving a travel company’s customer-service bot as an example.
What this means for a business (as of 13 November 2026)
An AI agent that answers your own customers about your own products, bookings and orders, on your own WhatsApp number, is the use Meta says stays allowed. A public “ask me anything” assistant offered through WhatsApp is the use the terms bar.
What stayed the same: the 24-hour window
Meta’s customer service window still governs every reply, scripted or generated. When a customer messages or calls you, a 24-hour timer starts and resets each time they write again. Inside it, you can send free-form messages; outside it, you need an approved template. Service replies inside the window are also the cheapest messages to send: from 1 October 2026, the first 1,000 service messages per business phone number each month are free, per Meta’s pricing page as recorded in our free service messages guide.
WhatsApp’s Business Policy adds one more condition that applies to bots and agents alike: you may automate replies during the 24-hour window, but you must also offer a prompt, clear and direct way to reach a person.
How does an AI agent avoid inventing prices or promises?
It answers only from data it can check, and something outside the model checks the answer. The language model writes the sentence; your catalogue decides the facts. This is the difference between an agent you can let customers talk to and a demo.
A practical checklist of guardrails, all of which WACRM applies today:
- Grounded answers. Product answers are built from the product, SKU, price and stock data synced from your store, and booking answers from your own service list and staff hours.
- A price guard. Every money figure in a draft is compared with the stored prices. If one does not match, the draft is replaced by a safe reply that names the products without a price and says a team member will confirm.
- Approval first. By default every reply and booking the agent proposes waits for a teammate to approve it. The owner can opt in to auto-send for replies inside the 24-hour window, and bookings have their own setting.
- Outside the window, a free-form AI reply always needs a teammate’s approval.
- A pause switch. One switch refuses every AI action at once; nothing is sent or charged while it is on.
- A loop guard. The agent stops itself after three proposed sends to one chat within five minutes.
You can see these controls step by step in the AI reply drafts guide. For an online store, the WhatsApp AI agent for ecommerce guide shows how the same agent answers product questions, recommends items and moves a chat towards an order.
When should a person take over?
A person should take over whenever the customer asks for one, or whenever the agent would have to guess. A good agent treats handoff as part of the job, not as a failure, and it leaves the teammate enough context to carry on without asking the customer to repeat themselves.
In WACRM, the agent hands the conversation to your team when:
- the customer asks for a human in their own words (phrases are recognised in several languages, including English, Spanish, Portuguese and Vietnamese);
- it reaches the repeat limit on one chat;
- the AI wallet balance cannot cover the next draft;
- the proposed action is high-risk and needs approval, such as a free-form send outside the 24-hour window.
The handoff turns AI off for that chat, assigns it to a teammate or leaves it unassigned for the team, and adds a private note with the reason and the most recent messages, so the teammate can reply without asking the customer to start over.
Which one fits your business?
Use a script when the answers are few and fixed, and an AI agent when customers ask open questions about things you already keep in a system. Most businesses that sell or book over WhatsApp end up needing both.
Decision checklist
| If most of your WhatsApp chats are… | Start with |
|---|---|
| Choosing from a short, fixed set of options | Buttons and lists |
| Order or booking status that a system already knows | Automated updates with a button or two |
| Product questions: size, colour, price, stock | An AI agent grounded in your WooCommerce or Shopify catalogue |
| Booking, rescheduling or cancelling appointments | An AI agent with your service list and staff hours |
| Complaints, refunds or sensitive cases | A person, with the agent drafting notes at most |
Checklist as of 13 November 2026.
Two questions settle most cases. First, do customers type questions that your menus do not cover? If yes, a script alone is losing them. Second, is the answer already stored somewhere, such as your store’s product list or your booking calendar? If yes, an agent can answer it safely; if not, the answer should come from a person.
How does WACRM run an AI agent on your WhatsApp number?
WACRM is an AI-native WhatsApp CRM built on Meta’s WhatsApp Business Platform. Its AI agent drafts answers to product and service questions from your own data, proposes bookings, reschedules and cancellations, and hands chats to your team inside the shared inbox. Product data comes from your connected WooCommerce or Shopify store.
Because WACRM supports coexistence, you can keep using the WhatsApp Business app on your phone on the same number while the agent and your team work in the inbox. AI usage is paid per use from your workspace wallet, so the agent stops rather than running up a bill when credit runs out; plan details are on the pricing page. Developers can also reach the same workspace from their own AI tools through the WACRM API and MCP server.
If you would rather not set it up alone, the Setup Commerce service connects your store and order messages for you. To see what the agent proposes and how approval works, explore WACRM AI agents.
