How to Set Up an AI Agent for Customer Service: A 7-Step Beginner's Guide

If you run a small business and spend your day answering the same questions over chat, an AI agent for customer service can take that routine off your plate. The good news: setting one up is simpler than it looks when you follow a clear order and start small. This beginner's guide walks through seven practical steps, from scope to daily operation, including the WhatsApp and data-privacy rules that apply if you serve customers in Brazil.

Before you start: what an AI agent for customer service is (and isn't)

Three terms often get mixed up. A chatbot follows scripts: the customer picks options and gets canned answers. A copilot helps a human agent by suggesting replies, but the person still talks to the customer. An AI agent talks directly with the customer, understands the request and resolves whatever falls inside the scope you define. This is the distinction the CloudHumans guide uses to organize an implementation, from pilot to rollout.

You wouldn't be testing something experimental. The Zendesk blog notes that Brazil already had tens of thousands of chatbots in 2022, handling billions of messages a month. And according to TIC Empresas 2023, 75% of small businesses in Brazil use WhatsApp or Telegram as a digital channel.

In other words, customers already expect to talk to you by message. What's missing is organization. You can often start without heavy engineering by focusing on scope and knowledge; the need for code depends on the tools you choose. It requires good decisions about what the agent does, what it knows and when it hands the conversation to a person.

Step 1: Define a small scope and choose the right channel

The most common mistake is trying to automate everything at once. Underchat lists exactly that among frequent problems, along with ignoring how each channel behaves. LiveAgent recommends the opposite: a narrow scope, expanded in stages, only after the pilot is validated.

In practice, pick one or two topics and start there. Frequently asked questions, business hours and appointment scheduling are good candidates.

What to automate first and what to leave to humans

Automate what is repetitive and has a clear answer: "What are your hours?", "Are you open on Saturdays?", "I'd like to book an appointment." Leave complex complaints, sensitive cases and exceptions to your team, since they call for judgment and empathy.

A simple rule: if you would always give the same answer, it's a candidate for automation. If the answer depends on context, negotiation or a decision from you, it stays with a human.

Why messaging apps are often the starting point

Prioritize the channel where your customers already are. For most Brazilian businesses that's WhatsApp, and the guides from WATI and CloudHumans point the same way.

But not every customer uses the app. Keep equivalent paths, such as phone and email, for people who prefer them or can't use messaging.

Step 2: Organize the information your agent will use

An agent only answers well if it has something to look up. WATI treats a solid knowledge base as a prerequisite, and Intercom recommends maintaining a quality Help Center to improve the accuracy of AI-powered support.

There is a pattern called RAG (retrieval-augmented generation). Put simply, the agent checks trusted documents from your company before replying instead of guessing. Zendesk explains that this raises accuracy, and Google Cloud points out that it helps reduce made-up answers.

Turn FAQs and policies into a knowledge base

Gather in one place what your support team answers every day. A starter checklist:

• business hours and address;
• services and prices;
• cancellation and rescheduling policy;
• payment methods;
• documents or preparation needed before the appointment.

Write in plain language, as if explaining to a new customer. Text that confuses people will confuse the agent too.

Also set limits: what the agent must not answer (for example, diagnoses or discount promises) and what it must not do on its own.

Owners and a review calendar

Give each topic an owner. Whoever handles pricing updates prices; whoever handles the calendar updates rescheduling rules. Schedule a regular review, monthly for example. Outdated information is the fastest way for an agent to start making mistakes.

Step 3: Design the handoff to a human agent

A good agent knows when to step aside. Zendesk advises planning the handoff to a human within your flows and defining how the conversation will be managed afterward, preserving context. There is also a way back: the handoff and handback documentation describes how a conversation moves from the bot to a human and can return.

When to hand off

Set objective criteria:

• the customer asks to speak with a person;
• the agent fails to understand two or three times in a row;
• the topic is sensitive, such as a serious complaint, billing or health.

Chatbots that trap customers in a loop are a common complaint. A clear exit to a human solves much of that problem.

How to pass context so customers don't repeat themselves

Nothing is more annoying than telling your story twice. At handoff, send the human agent a summary of the conversation, the data already collected (name, service, preferred date) and the reason for the transfer. Also define who picks up the queue and how quickly, so the customer isn't left waiting.

Want an assistant that answers and books on its own? Create yours in minutes

Step 4: Set up scheduling within your routine

For clinics and service businesses, scheduling is where automation saves the most time. Go2Flow describes flows in which the agent connects to the calendar and CRM to book and confirm appointments, with human handoff and metrics. Google Calendar has a native appointment scheduling feature that can be integrated with assistants for automatic booking.

Before connecting anything, decide your business rules:

• Active confirmation: does the agent ask for confirmation a day before?
• Rescheduling: how many hours ahead can a customer reschedule?
• Squeeze-ins: who approves slots outside the normal schedule?
• No-shows: what happens when a customer doesn't show up?
• Waitlist: how do you notify someone when a slot opens?

Reminders sent outside WhatsApp's 24-hour window require approved templates and customer opt-in. The next step explains that.

Step 5: Handle WhatsApp rules and privacy law

This is the part that scares beginners most, but it boils down to a few rules. The rules below come from Brazilian sources and apply if you serve customers in Brazil. This content is informational and is not legal advice: talk to a professional about your specific case.

The 24-hour window, templates and opt-in

According to the WhatsApp Business policy, outside the 24-hour customer service window a business can only send messages using approved templates, and customer opt-in is required for business-initiated contact. AWS details that the window restarts with every new message from the user. Outside it, approved templates are subject to charges for each conversation the business initiates.

There is one exception: a public Prodemge/Serpro contract records a free 72-hour window when the customer comes in through a click-to-WhatsApp ad or the Facebook and Instagram buttons (in effect since July 1, 2025).

On weekends and holidays, set up an out-of-office message and plan to pick the conversation back up on Monday so you don't lose it.

LGPD, ANPD and the SAC Decree

Brazil's data protection law, the LGPD, defines legal bases for processing personal data, requires appointing a data protection officer (DPO) and calls for transparency when handling data subjects' rights. In practice, tell customers they are talking to an AI and explain what their data will be used for.

ANPD Resolution CD/ANPD No. 18/2024 aims to prevent LGPD rights from being exercised exclusively through a chatbot: there must be an option for service with human interaction. If your company falls under the SAC (customer service) rules, Decree 11.034/2022 also comes into play: oversight bodies cite free service, cancellation through the same channel and meeting deadlines.

If you operate elsewhere, check which privacy and messaging rules apply to your business before you launch.

Step 6: Run a pilot and measure results

Start with a small group of customers or a specific time slot. Before turning the agent on, write down how support performs today. Without that baseline, you won't know whether things improved.

Essential metrics

MKOM explains the most common customer service metrics:

• TME: average wait time;
• TMA: average handling time;
• FCR: first contact resolution;
• CSAT: customer satisfaction after the interaction;
• NPS and CES: customer recommendation and effort.

If you're just starting out, wait time, handling time, FCR and CSAT are enough.

Set a baseline before the pilot

Measure over weeks, not a single day. Set goals for each stage: how many conversations the agent resolves on its own, how many appointments it books, how many handoffs are needed. Only expand to new topics or channels after validating the pilot, as LiveAgent and CloudHumans recommend.

Step 7: Daily operation and continuous improvement

Setting up an agent isn't a project that ends. Intercom recommends continuously evaluating the agent's performance, and the routine can be light.

Weekly conversation review

Set aside half an hour a week to read the conversations that failed or were handed off. For each one, ask: was information missing from the knowledge base? Were the handoff criteria right? Was the answer clear? Update the knowledge base and adjust the rules.

The new role of your human team

With the agent handling the repetitive work, your team resolves exceptions and complex cases, which CloudHumans describes as the split between Tier 1 and Tier 2. The failures your team finds become improvements for the agent.

That's the key point: with a small scope, organized information and a clear way out to a human, getting started fits into the routine of any beginner business. Start with one topic, measure, and expand from there.

Ready to start? Talk to our team