The short answer
A WhatsApp chatbot with AI handles repeated questions, triage and scheduling well, as long as it runs on the official WhatsApp Business API, hands the conversation to a person when the matter calls for judgment and is connected to the company's calendar or system. Without that, it just answers fast and gets things wrong with confidence.
What defines the result is not the AI model you choose, it is the design of the service: which topics the bot takes on, where it stops, what data it looks up and how you will know it is working. This guide walks through each of these decisions, along with the care that Brazil's data protection law (LGPD) requires for the channel.
What an AI chatbot solves on WhatsApp, and what it does not
The difference between an AI chatbot and the old menu bot is that it understands the message the way the customer wrote it, with typos, transcribed audio or three questions in the same sentence. That greatly expands what can be automated, but not everything. What it usually handles well:
- Repeated questions. Opening hours, address, required documents, delivery times, order status. The AI answers from a knowledge base reviewed by your team, not from what it thinks it knows.
- Triage. Understanding the topic, collecting the basic details (name, order number, location) and routing to the right queue with a summary of the conversation.
- Scheduling and rescheduling. Checking open slots, booking and confirming, when the calendar is integrated.
- After-hours service. Recording requests at night and on weekends so the team picks them up in the morning with everything organized.
What should not be left to the bot alone: negotiation, complaints from an upset customer, health matters that require professional assessment, exceptions to the rule (discounts, refunds, special deadlines) and any answer that creates a commitment the company has not reviewed. In these cases AI can summarize and prepare, but a person answers.
It is the same principle we follow in automation and artificial intelligence: AI organizes the queue, summarizes the history and proposes the answer, and your team remains responsible for what is decided.
Official WhatsApp Business API or an unofficial solution?
There are two ways to connect a system to WhatsApp. The official one is the WhatsApp Business platform maintained by Meta, accessed directly through the API or through a partner provider. The unofficial one is tools that automate the app or WhatsApp Web as if a person were clicking.
Unofficial solutions often look simpler to start with, but they go against WhatsApp's terms of use. The concrete risk is the number being blocked, sometimes without notice, and the company suddenly losing the channel its customers know, along with the conversation history. For a customer service channel, that risk rarely pays off.
The official API comes with rules that are part of the project from the start:
- Service window. After the customer sends a message, the company can reply freely for 24 hours. Outside that window, only with message templates approved in advance.
- Permission for outbound messages. To start a conversation (reminder, confirmation, notice), the company needs the customer's permission to be contacted on WhatsApp.
- Usage policies. Meta defines what can and cannot be done on the platform, including the use of AI, and these rules change. It is worth checking the current business policy before designing the flow.
- Platform charges. Meta charges according to the type and volume of messages, with rules that change over time and by country. This cost goes into the project budget from the start.
How to hand off from the chatbot to a human agent
Handing off to a person is where most automated services disappoint. The customer explains everything to the bot and, on reaching the agent, has to repeat it from the beginning. A good handoff has four elements:
- Clear triggers. The customer asks to talk to a person, the AI is not confident in the answer, the topic is on the list of those that require a human, or the conversation shows frustration.
- Context handed over. The agent receives the summary, the data already collected and the history, without having to ask again.
- Honest expectations. If no one is available right now, the bot says when someone will be and records the request, instead of promising a transfer that will not happen.
- Learning recorded. What the agent resolved feeds the knowledge base, so the AI answers the same question better next time.
The customer also needs to know, from the first message, that they are talking to an automated assistant and how to reach a person.
Integration with calendar, CRM and company systems
A chatbot that only chats solves little. The real gain comes when it reads from and writes to the systems the company already uses: the calendar to book appointments, the CRM to know who the customer is and what they have bought, the order system to report status, the ticketing system to open a case.
These connections follow the same rules as any integration: the chatbot needs an API or another stable means of reading and writing, permissions limited to what it actually uses and a plan for when the other system is down. We detail these choices in the guide on how to integrate systems.
One specific point: everything the bot does on the customer's behalf needs to leave a trail. In a service system like Ágil Cidadão, each request has a case number and each update is recorded with author and date. The same applies to what comes in through WhatsApp: who asked, what the bot answered and what was saved in each system.
LGPD and consent in WhatsApp customer service
A service conversation carries personal data from the first message: name, phone number and often a tax ID, address or health information. The LGPD applies fully to this channel, and AI adds questions of its own, such as where the messages processed by the model go.
- Purpose and necessity. Collect only what the service needs. If scheduling does not require a tax ID, the bot does not ask for one.
- Transparency. Say that the conversation is handled by an automated assistant, what the data is for and where the privacy policy is.
- Legal basis and permission. Consent is one of the legal bases of the LGPD, not the only one. For sending outbound messages on WhatsApp, however, the platform's policy requires the customer's permission, and it is worth keeping a record of when and how it was given.
- Sensitive data. Health information, for example, calls for extra care: limiting what the model receives, anonymizing when possible and restricting who sees the history.
- Retention. Define how long conversations are kept and delete what no longer has a purpose.
- Automated decisions. If the bot decides something that affects the customer, such as denying a request, the LGPD gives the data subject the right to ask for a review. In practice, that is one more reason to leave decisions with a person.
We go deeper into these points in the guide on LGPD and artificial intelligence.
How to measure whether your WhatsApp chatbot is working
The number of messages answered does not measure service. Before going live, agree with the team on which indicators count and what today's starting point is, so you can compare later.
- Resolution without an agent. How many conversations ended resolved by the bot alone, checked by sampling, not by the bot's own assessment.
- Accuracy rate. In a sample reviewed by people, how many answers were correct and complete.
- Handoffs to a human. How many there were, for what reason and how long the customer waited after them.
- Response and resolution time. Before and after automation, comparing the same type of topic.
- Satisfaction. A short question at the end of the conversation, read together with the complaints that arrive through other channels.
Start with a few topics, measure for a few weeks and only then expand. It is the rhythm of short cycles we describe in how we work: each delivery goes into use, is checked in day-to-day use and guides the next one. If the question is where to start, the guide on which processes to automate first helps you choose.