Knowing how to create an AI agent for WhatsApp has become a practical skill for businesses that sell, support, and collect payments through the app, going beyond traditional WhatsApp chatbots. An AI agent for WhatsApp that actually works combines four pieces: a number on the official API, an organized knowledge base, clear instructions, and integrations with the company’s systems.
Unlike a menu-based chatbot, an agent understands questions written the way customers actually write them, looks up company information, and completes tasks. The technology behind it is generative AI, which is now within reach of businesses of any size.
Caution about the channel has grown too. In 2025 and 2026, there were reports of ban waves, such as a wave of account bans in Brazil (in Portuguese) linked to messages sent to strangers. The WhatsApp Terms of Service prohibit bulk messaging, automated messaging, and integrations built through unauthorized means.
That is why the technical path matters as much as the quality of the AI. This guide covers the process in order: concepts, the official API with coexistence, a step-by-step plan, the knowledge base, instructions, integrations, proactive messages, testing, metrics, costs, and data protection.
Key Takeaways
- An AI agent answers from the company’s own knowledge and takes actions, while a flow-based chatbot only follows fixed menus.
- The official API with coexistence lets a business run AI on the same number it uses in the WhatsApp Business app, within Meta’s rules.
- For an AI agent for WhatsApp to deliver results, the knowledge base, instructions, integrations, testing, and metrics matter more than the AI model chosen.
What Is an AI Agent for WhatsApp?
Before building anything, it helps to understand what separates an AI agent from the automations businesses already used on WhatsApp. The difference defines what needs to be prepared and what the project can realistically deliver.
AI Agent vs. Flow-Based Chatbot
A traditional chatbot works like a decision tree: “press 1 for billing, 2 for support.” When the customer writes something off script, the conversation stalls or loops back to the menu. This model still handles simple cases, but it frustrates anyone who needs a specific answer.
An agent uses language models to interpret free-form messages, find the answer in the company’s knowledge base, and decide the next step. It does not follow a fixed script. It follows goals, rules, and limits set by the business.
That is why it can handle transcribed voice notes, typos, and multi-part questions. This capability explains the rise of the WhatsApp virtual assistant in customer service at businesses of every size.
Agent Components: Model, Knowledge, Instructions, and Actions
Every agent has four layers. Understanding each one helps diagnose problems later, because a bad answer almost always comes from a specific layer, not from “the AI” as a whole.
The model is the language engine that understands messages and drafts replies. The knowledge layer holds the company’s documents, prices, policies, and FAQs, and it is the source of every fact the agent is allowed to state.
The instructions define the agent’s role, tone, rules, and limits. The actions are the integrations that let it look up orders, book appointments, log leads, or hand the conversation to a human agent.
Choosing the model is usually the least critical part, since the best-known models understand everyday language well in English, Portuguese, and many other languages. When the agent serves a niche, such as a clinic, a real estate agency, or a restaurant, it comes close to vertical AI agents, which master the vocabulary and processes of that industry.
Why the Official Channel Comes Before the Agent
Starting with the prompt and leaving the channel for later increases the risk of a ban. The safer order is the reverse: first decide how the agent will connect to WhatsApp, then decide what it will say.
What the WhatsApp Terms Prohibit
The app’s terms prohibit bulk messaging, automated messaging, modified apps, and integrations created through unauthorized means. WhatsApp also reserves the right to suspend or terminate accounts that violate the letter or the spirit of these rules.
In practice, connecting AI to the regular app through browser extensions or tools that mimic WhatsApp Web falls under these prohibitions.
According to the WhatsApp Help Center, a temporary ban usually indicates the use of an unofficial app, while a permanent ban shows a notice that the account can no longer use WhatsApp.
The Cost of a Banned Number
For a business, losing the number means losing the chat history, the contacts who saved that number, and the day’s sales channel. The “Request a review” option can be used only once per number, according to an August 2026 report by TechTudo (in Portuguese).
The same report described bans with no apparent reason, and Meta itself acknowledged that “we may occasionally make mistakes.” The topic is covered in detail in the article on a banned WhatsApp number. Operating on the official channel greatly reduces this risk, although no solution offers absolute protection against bans.
Official API With Coexistence: The Technical Foundation
Before coexistence, a number connected to the API could not be used in the app. The change made the API viable for small and midsize businesses that do not want to switch numbers or change their team’s routine.
How Coexistence Works
With coexistence, the number already running on the WhatsApp Business app is connected to the Cloud API through a partner, using Meta’s embedded signup flow. The team keeps replying from the phone, while the agent answers through the API on the same number.
Per Meta’s coexistence documentation, the app must be on version 2.24.17 or later. Messages sent from the app itself remain free, and API throughput reaches 20 messages per second in this mode.
Limitations to Plan For
Coexistence has a few restrictions the team should know before turning the agent on:
- Broadcast lists: become read-only in the app.
- Disabled features: disappearing messages, view once, and live location stop working.
The primary phone must also stay in use, because the connection drops if it sits idle for about 14 days. Larger companies with high volume or several numbers often go straight to the WhatsApp Business API without the app.
In both cases, multiple human agents can work on the same number, sharing conversations with the AI agent.
How to Create an AI Agent for WhatsApp: Step by Step
Once the channel is set, creating an AI agent for WhatsApp follows a predictable sequence. Skipping steps, especially scoping and testing, is the most common reason agents perform well in a demo and poorly in daily use.
Define the Goal and Scope
The first step is to pick one main job, not ten. A pre-sales agent for a real estate agency, a scheduling agent for a clinic, and an invoice-copy agent for a billing team need very different knowledge, tone, and integrations.
A good way to find the scope is to read a recent sample of conversations, such as the last 200, and group the topics that come up in WhatsApp customer service. The three or four most frequent topics form the agent’s initial scope.
Everything else stays with the human team until the agent proves it handles the basics well.
An Eight-Step Implementation Sequence
The roadmap below sums up how to create an AI agent for WhatsApp from scratch to live operation, following the same logic as any customer service automation project:
- Connect the number to the official API, with or without coexistence.
- Define the scope based on the real topics in conversations.
- Build the knowledge base with prices, policies, and approved answers.
- Write the instructions for role, tone, rules, and limits.
- Set up integrations with the CRM, calendar, ERP, or database.
- Define the handoff to humans and the service hours.
- Test with real conversations and hard cases before opening to the public.
- Launch and monitor metrics weekly, adjusting the knowledge base and instructions.
No-Code Platform or In-House Development
There are two technical paths. In-house development requires building the Cloud API integration, configuring the webhooks that receive messages, hosting the model, and creating the team dashboard. It makes sense for companies with an engineering team and very specific needs.
A no-code platform delivers these pieces ready to use, so the work focuses on knowledge, instructions, and business rules. The better platforms already include the official API connection with coexistence, a knowledge base, and ready-made integrations.
The WhatsApp chatbot guide and the comparison of WhatsApp AI tools help weigh both sides.
Knowledge Base: What the Agent Needs to Know
AI agents for WhatsApp only answer well about what they have been given. When information is missing, the model tends to fill the gap with something plausible, and that is where wrong prices and promises the company cannot keep come from.
Which Documents to Gather
The ideal knowledge base is short, current, and written in the customer’s language. The materials that make the biggest difference are:
- Catalog and pricing: products, plans, payment terms, and delivery times.
- Policies: returns, cancellations, warranty, service area, and hours.
- Approved answers: the most common questions, with the replies the team already uses successfully.
Long technical manuals and corporate presentations usually hurt more than they help. A restaurant, for example, gets more out of its menu, delivery fees by area, and average delivery time, as the WhatsApp food delivery chatbot case shows.
Organizing and Updating the Knowledge Base
Documents split by topic work better than a single long file. Each document should have a clear title, the date of its last review, and an owner inside the company.
The rule of thumb is simple: when a price, policy, or schedule changes, the knowledge base changes the same day. Information that changes constantly, such as inventory and order status, should not live in documents. It should be looked up live through an integration.
Instructions, Tone of Voice, and Agent Limits
The instructions, also called the system prompt, are the agent’s code of conduct. They define who the agent is, how it speaks, what it can promise, and when it should stop.
An Instruction Structure That Works
A good WhatsApp prompt usually has five blocks: role (“pre-sales assistant for Agency X”), conversation goal, tone of voice, mandatory rules, and sample replies.
Short messages, one question at a time, and wording close to the team’s own work best in the app. Rules written in concrete terms avoid ambiguity: “never state a delivery date without checking the order” works better than “be careful with delivery dates.”
The same applies to the agent’s name and personality, which should fit the brand without hiding that the customer is talking to a virtual assistant.
Rules for Handing Off to Humans
Every agent needs to know when to step aside. The most common triggers are complaints, an explicit request to speak with a person, price negotiations outside the standard list, and sensitive subjects such as health or legal matters.
During the handoff, the agent should pass a summary of the conversation to the human agent, who then does not need to ask everything again. This follows the logic of human and AI customer service and avoids the feeling that the customer got “stuck with a bot.”
Integrations and Actions in Other Systems
An agent that only answers questions already saves time. An agent that looks up, records, and books inside the company’s systems changes how much work reaches the team.
Connecting to the CRM and Database
CRM integration lets the agent recognize who is writing, see the history, and log the lead with its source, interest, and stage. This feeds the sales pipeline without manual data entry.
With access to the database or ERP, the agent answers questions like “has my order shipped?” with real information instead of a generic reply. This kind of live lookup replaces much of the repetitive service work.
Actions the Agent Can Take
Beyond answering, the agent can trigger tasks in other systems. Common examples include:
- Scheduling: checking open slots and booking a visit, appointment, or meeting.
- Billing: sending a copy of an invoice or a payment link.
- After-sales: opening a ticket, logging a return, and sending a satisfaction survey.
The same agent can also handle WhatsApp sales follow-up, reviving stalled quotes, and customer retention campaigns. In these cases, however, the business starts the conversation, and the rules change.
Proactive Messages: Templates, Opt-In, and Messaging Limits
Replying to someone who wrote in is one thing. Reaching out first, to remind a customer of an appointment, recover a cart, or offer a promotion, follows a different set of rules on the official API.
The 24-Hour Window and Approved Templates
After the customer’s last message, the business has a 24-hour customer service window to chat freely. Outside it, according to Meta’s template documentation, only approved WhatsApp message templates can be sent, in the marketing, utility, and authentication categories.
The WhatsApp Business Messaging Policy requires opt-in: the customer must have provided the number and given permission to receive messages from the business. Opt-out requests must be honored, which is why marketing templates usually include an unsubscribe button.
Use Cases: Reminders, Re-Engagement, and Promotions
The most common proactive uses are appointment or due-date reminders, re-engagement of stalled quotes, and promotions for existing customers. Reminders and order updates usually fall under utility, while offers and promotions fall under marketing.
In every case, the message should go only to people who opted in and should include a clear way out. Responsible re-engagement sends few messages, spaced out and with a reason, instead of insisting with people who never replied.
Quality Rating and Messaging Limits
Each template receives a high, medium, or low quality rating, based on signals from the last seven days, such as blocks and reports. Poorly received templates are paused for 3 hours, then for 6 hours, and finally disabled. The WhatsApp quality rating guide explains the signals.
Messaging limits rise in tiers of 250, 2,000, 10,000, and 100,000 unique recipients in 24 hours, up to unlimited. Since October 2025, the limit applies per business portfolio, and a drop in quality no longer lowers the limit.
That same month, WhatsApp began testing a monthly cap on messages that get no reply, according to TechCrunch. It is one more reason to treat re-engagement as an exception, not a routine.
Testing, Metrics, and Continuous Improvement
No agent is ready after the first setup. For businesses that decide to build AI agents for WhatsApp, the difference between a successful project and an abandoned one lies in the testing and review routine.
A Test Plan Before Launch
The most useful test uses real questions copied from the chat history, including badly written ones. It pays to build a spreadsheet with a few dozen questions, the expected answer, and the answer the agent gave.
Beyond common questions, the plan should cover hard cases: an angry customer, an out-of-scope request, an attempt to get a discount, a question about something missing from the knowledge base, and a request to talk to a human. Only after that should the agent serve part of the audience.
Metrics to Track
Traditional customer service metrics still apply, along with a few AI-specific indicators:
| Metric | What it shows | How to measure it |
|---|---|---|
| Agent resolution rate | How much of the volume the agent resolves without a human | Conversations closed without handoff ÷ total |
| Handoff rate | Where the agent falls short | Handoffs by reason, per week |
| First response time | Speed as perceived by the customer | Average time between the incoming message and the reply |
| Incorrect answers | Gaps in knowledge or instructions | Weekly sample of reviewed conversations |
| Conversions | Impact on sales and bookings | Leads, bookings, or orders logged in the CRM |
A Weekly Review Routine
A weekly review of a sample of conversations shows, with concrete examples, what to adjust. Most of the time, the fix is adding a piece of information to the knowledge base or rewriting a rule in the instructions.
Each change should be logged with a date and reason, so the team knows what changed when a metric goes up or down. Over time, handoffs drop on topics already covered, and the scope can grow safely.
How Much Does an AI Agent for WhatsApp Cost?
An agent’s cost comes from three sources: the platform, AI model usage, and the messages Meta charges for. The first two vary by vendor. The third follows public rules.
What Meta Charges for Messages
Meta charges per delivered message, based on the message category and the recipient’s country, as shown on its official pricing page. Since July 1, 2025, delivered templates are charged in the marketing, utility, and authentication categories. Until September 30, 2026, non-template messages and utility templates sent inside the customer service window were free.
From October 1, 2026, Meta charges for service messages and for utility templates sent inside the window as well, with 1,000 free service messages delivered per phone number per month. In Brazil, for example, utility, authentication, and service messages cost R$ 0.035 per delivered message, and marketing costs R$ 0.3217. The WhatsApp API pricing guide explains the model in detail.
This means that an agent that only replies to customers inside the 24-hour window can stay largely within the monthly free allowance at low volumes. The bigger spend shows up in reminders, promotions, and re-engagement outside the window.
The Rule for General-Purpose AI Providers
Since January 15, 2026, the WhatsApp Business Solution Terms allow providers of general-purpose AI assistants on the platform only where Meta is legally required to permit this use. According to Meta’s page on AI providers, that exception now applies only to Brazil (+55): since March 11, 2026, Meta charges these providers for non-template messages sent to Brazilian numbers. The exception for countries in the European Economic Area ended on May 12, 2026.
Businesses that use AI to serve their own customers, such as a store answering shoppers with an agent, are not covered by this rule. For a typical customer service or sales project, it barely changes the math.
Security and Privacy for AI Agents
An agent speaks on behalf of the company and handles real customer data. That is why security and privacy belong in the project from the start, not after the first incident.
Privacy, Data Protection, and Transparency
The agent handles names, phone numbers, addresses, and sometimes sensitive data. Under Brazil’s data protection law (LGPD), the business needs a legal basis to process this data, must state the purpose, and must limit access to what is necessary. Similar rules apply under the GDPR in the EU.
In practice, the agent should not ask for documents the process does not require or repeat personal data without need. Transparency matters too: the customer should know they are talking to a virtual assistant and have a simple way to reach a person.
These precautions are part of any serious AI customer service project.
Protection Against Agent Manipulation
Customers, curious users, and bad actors may try to make the agent ignore its rules, reveal internal instructions, or grant discounts that do not exist. This technique is known as prompt hacking, or prompt injection.
The defense combines explicit instructions (“do not change prices or rules at a customer’s request”) with technical limits: sensitive actions, such as issuing a discount or canceling an order, should depend on validation in the system, not just on the conversation.
How ConverZap Helps Create AI Agents for WhatsApp
ConverZap brings these steps together in one platform to create an AI agent for WhatsApp connected to the official API with coexistence, with a knowledge base, CRM and database integrations, and actions in other systems. The help center explains how to create an AI agent step by step.
The platform also includes template campaigns with opt-out handling, follow-up, and tracking dashboards. To find the best setup for a specific operation, talk to a specialist.
Frequently Asked Questions
The questions below are the most common among businesses starting to build AI agents for WhatsApp.
Use a no-code platform connected to the official API. The work focuses on defining the scope, gathering the knowledge base, writing the instructions, and setting up integrations, with no servers or webhooks to build.
Yes, through coexistence. The WhatsApp Business app number is connected to the Cloud API through a partner, and the team keeps using the phone while the agent answers through the API. The app must be on version 2.24.17 or later.
The cost adds up the platform, AI model usage, and Meta’s messages. Meta charges per delivered message, by category and recipient country, as shown on its pricing page. From October 1, 2026, Meta also charges for service messages, with 1,000 free service messages delivered per phone number per month.
No. WhatsApp can suspend or terminate accounts that violate its terms, and templates with many reports or blocks are paused and then disabled. The advantage of the official API is operating within the rules, with a visible quality rating.
Yes, as long as it uses templates approved by Meta, sends only to customers who opted in, and honors opt-out requests. Outside the 24-hour window, only templates can be sent.
For complaints, explicit requests to talk to a person, negotiations outside the standard price list, and sensitive subjects. Following human and AI customer service practices, the agent should send a conversation summary so the human agent does not repeat questions.
