How to create an AI agent: a step-by-step guide to serve customers
By José Pedro · Co-founder of CRPRO
Learn to create an AI agent with a six-step method: goal, instructions, material, limits, testing and review. See an instruction template, what changes on WhatsApp, what you can do for free and how this path works on CRPRO.
To create an AI agent, define a single goal, write the instructions, gather the company's material, set limits, test with real conversations and review after you publish. With ready-made tools, you don't need to code. The main work is explaining the business clearly, and this guide shows how to do that at each step.
What is an AI agent?
An AI agent is a program that understands what a person wrote, decides what to reply and can carry out tasks. It has three parts:
- AI model: the technology that reads and writes, like those from OpenAI, Anthropic and Google.
- Instructions and material: what it needs to know about the company.
- Actions: what it does besides replying, such as booking a time or alerting a salesperson.
A menu chatbot follows paths drawn in advance. An agent handles people who write in their own way, but it can be wrong in a convincing way. That's why the method below spends more time on limits and testing than on technology. If you're still deciding between the two, compare them in how to choose a WhatsApp chatbot.
Step 1: what task will the agent solve?
Start with one task only. An agent that tries to do everything tends to do everything halfway.
Read the company's last 50 conversations and note what repeats. Then pick a goal that fits in one sentence:
- Clinic: answer questions and book the assessment.
- Real estate agency: understand what the customer is looking for and pass them to a broker.
- Service provider: qualify the lead before the quote.
Also define how to measure the result. It can be the number of bookings, qualified leads or conversations resolved without a handoff.
Step 2: what goes into the instructions?
Instructions are the text that tells the agent who it is and how to act. Write as if you were training a new person. This template works for most cases:
- Role: "You are the virtual assistant of Aurora Clinic."
- Goal: "Understand what the person needs and book an assessment."
- Tone: "Short sentences, simple language, one question at a time."
- Required questions: "Name, procedure of interest and best day."
- Handoff: "Call a person if there is a complaint, a request for a discount or a request to talk to someone."
Prefer concrete rules. "Be friendly" changes little. "Greet by name when you know it" changes the reply.
Step 3: what material does the agent need?
The agent only answers well about what it has received. Gather:
- services or products, with what's included;
- frequently asked questions and the answers the team already uses;
- hours, addresses and payment methods;
- exchange, cancellation and delivery policies.
Remove what's outdated before you upload it. If the terms changed and the document didn't, the agent repeats the old information with total confidence.
Step 4: what limits should you write?
A limit is everything the agent can't say or do on its own. Write it explicitly:
- What it can't promise: off-list prices, deadlines, discounts, results.
- What it does when it doesn't know: says it will check with the team, without making things up.
- When it stops: a complaint, a sensitive subject or a request to talk to a person.
- How it introduces itself: as the company's virtual assistant, so nobody feels deceived.
No instruction eliminates error. They lower the chance and leave the exit ready for when it happens.
Step 5: how do you test before publishing?
Build a list of hard conversations and run all of them before any customer sees the agent:
- a simple question, which it must get right;
- two questions in the same message;
- a subject that isn't in the material;
- an angry customer;
- an insistent request for a discount;
- a message trying to change its rules.
When it gets one wrong, adjust the instruction or the material and repeat the test. Publish only when the hard answers are acceptable.
Step 6: how do you review after publishing?
Start small, on one line or one subject. Then read the conversations every week and look for three things: a wrong answer, a handoff that came too early and a customer who gave up. Each problem becomes an adjustment. The business changes over time, and the agent only keeps up if someone updates the instructions.
How do you create an AI agent for WhatsApp?
The method is the same. Two decisions change:
Connection. The official WhatsApp API is the route Meta authorizes for businesses and the most predictable one. A QR Code connection is faster to set up, but carries a higher risk of blocking at high volume.
Sending rules. The agent replies to whoever wrote. To start a conversation outside the 24-hour window, Meta requires an approved template.
Can you create a free AI agent?
To learn and test, yes. Some tools have a free plan or a trial period. To serve real customers, some cost almost always shows up: the AI model's usage, Meta's messages on the official API and the tool that keeps everything running.
If your case fits in welcome messages and fixed replies, you may not even need AI. See what you can do without paying.
How does this path look on CRPRO?
CRPRO's AI agent is part of the CR Pro IA plan and follows the same method:
- Open the agent screen through the menu search.
- Describe the business, the tone and what it can or can't promise.
- Upload the support material, such as a PDF or a page from your website.
- Use the plan's monthly allowance of conversations or your own OpenAI, Anthropic or Google key.
- Check the replies in the test chat on the same screen.
- Put the agent live through a workflow: publishing turns it on, pausing turns it off.
After that, it replies on WhatsApp, through the official API or QR Code, and in Instagram Direct. It follows the script, moves the card in the pipeline, applies a tag, notifies the person in charge and transfers to a queue or a person. With permissions turned on, it checks the calendar and books, reschedules or cancels. It also sends a reminder to people who stopped replying. Phone calls are left out.
See the details on the AI agent on WhatsApp page and the terms in plans. CRPRO has 7 days free.
Frequently Asked Questions
Do I need to know how to code to create an AI agent?
With ready-made tools, no. The main work is writing clear instructions and gathering the company's material, and whoever knows the business usually does that better. Coding comes in when you build everything from scratch or connect your own systems.
How long does it take to create an AI agent?
It depends on the material you already have. With questions, answers and policies organized, the basics move faster. Fine-tuning comes from real conversations, after you publish.
Can an AI agent make up an answer?
It can. Clear instructions, up-to-date material and a rule to check with the team reduce the risk, but don't eliminate it. That's why you test first, review afterward and always keep an exit to a person.
Does the customer need to know they're talking to AI?
The recommendation is for the agent to introduce itself as the company's virtual assistant right at the start. That makes clear who the person is talking to and avoids the feeling of being deceived.
Can you use an AI agent on Instagram?
Yes, if the tool supports the channel. On CRPRO, the agent replies in Instagram Direct. Replying to comments is turned off for now, and audio replies exist only on WhatsApp.