The first temptation for anyone who decides to build an AI agent in their company is to start with the model. Which one to pick: ChatGPT, Claude, Gemini, Llama? That is the last thing that matters. The first is the process. The second is the data. The third is the integration. The model is a decision for week five, and whoever flips that order burns money in the first four.
This text is for those just getting started. It is not a technical guide. It is what nobody tells you at the moment of the initial decision, and what makes the difference between a project that gets off the ground and a project that becomes just another slide filed away. If you are an owner, director or operations leader thinking about taking the first step, read this before signing any proposal.
1. Myth number 1: "I need to train an AI"
Almost nobody does. Training an AI model (from scratch, or a deep fine-tune of an existing model) costs millions and requires a specialized team. For 95% of companies, the path is to use an already trained model (from OpenAI, Anthropic, Google or a good open-source one) and adjust its behavior through instructions and context data. It works better, costs less, and takes far less time.
When you "train" in the homegrown sense, you are usually doing one of two things: giving the agent access to your documents (that is RAG, search over a knowledge base) or teaching business rules through a detailed prompt. Neither of them is real training. Both solve the problem. Confusing the terms gets expensive when negotiating with a vendor.
(To understand from the inside how this works, see how an AI agent works under the hood.)
2. Myth number 2: "it has to be perfect before showing it to the customer"
Whoever waits for it to "get good" before putting it into production never puts it into production. An agent gets good by being used, not by being prepared. The real rule is different: define the acceptable floor ("the agent can never make up customer data, can never give a discount without authorization, always escalates case X") and release it in small volume as soon as that floor is guaranteed.
Small volume means 5% to 15% of the real total, with a human backstop reading every conversation in the first days. You learn in 2 weeks what would take 6 months of simulation. And you learn under real conditions, with a real customer, with swearing, with a change of subject, with the case nobody predicted.
The risk of waiting for perfection is delivering late, wrong and expensive. The risk of releasing early (with a floor and a controlled sample) is hearing the occasional complaint and fixing it. The first one kills the project. The second is the way.
3. Myth number 3: "the agent will replace the team"
It will not. Whoever comes in with that expectation breaks twice: they spend a fortune trying to force the agent to do what a human does better, and they lose the good team out of fear.
A well-built agent does the repetitive, standardized, high-volume, low-variation work. A good human does the non-obvious work, with nuance, with hard decisions, with empathy in tense moments. A company that puts both to work in layers (the agent filters and solves the easy stuff, the human takes the rest) gets a real gain. A company that tries to replace humans with agents everywhere delivers worse service and loses useful people as a side effect.
The right question is not "how many humans do I cut?". It is "which part of the human work turns into margin if I take it off their hands?". That difference is what separates efficiency from layoffs.
4. The right start: one process, one channel, one objective
Anyone building an AI agent from scratch should choose, in the first project, exactly these three things: one process (not two, not five), one channel (WhatsApp OR email OR the website chat), one number objective (reduce AHT (average handle time) from X to Y, increase conversion from A to B).
Artificial focus seems restrictive. It is not. It is what gets the first project across the finish line. A company that starts with a fat scope wastes time, spends money, and never gets the internal success case that unlocks the next projects. A company that starts small, closes well, earns credibility, and expands in the second cycle with 3x more ease.
A typical good first project: an agent that answers about order status via WhatsApp, integrated with the order system, with a goal of resolving 60% of messages without a human and cutting AHT from 18 min to 3 min. It is small, it is clear, it is measurable. It is the case that unlocks everything else.
5. The first 30 days: what has to hit the street
For those just starting out, the first 30 days have three deliverables (not five).
Week 1: choosing the process, choosing the channel, choosing the indicator, and a measured baseline (what is the number today, before the agent). Without a baseline, any goal is a guess.
Weeks 2 and 3: mapping the current process with the team, gathering the tacit rules, listing the systems the agent will need to consult. This is fieldwork, not technology. Whoever skips this pays for it later.
Week 4: a rough prototype that talks with real messages (anonymized) and shows how far it gets right on its own. It does not need to be pretty. It needs to show numbers: "in 50 test conversations, it got 32 right, got 11 wrong, escalated to a human in 7". That is the basis for deciding whether to continue or pivot.
A company that delivers these three milestones in 30 days is at the right pace. A company that is still debating "which AI vendor to choose" at the end of month 1 will spend 6 months to deliver what fit in 3.
6. A sign that you are on the right track (and the 4 red flags)
You are doing well if, every two weeks, the internal team responsible for the agent can clearly answer three things: what changed, what improved (with a number), what is still bad. If they cannot, they are operating in the dark.
Four red flags that show up early and need immediate action.
Red flag 1: no real data has entered the project by week 4. It is being built in a vacuum. Stop, bring in data, adjust.
Red flag 2: nobody in-house can explain what the agent does. It is being done "by the consultancy" without absorption. It will become a black box.
Red flag 3: the scope grew from one to three processes before the first one delivered. It is bloating. Pull back.
Red flag 4: the vendor cannot show you a sample of conversations where the agent got it wrong. They are hiding what does not work. A big red warning. (To understand why this is so critical, read why an AI agent soars in the demo and dies in production.)
Whoever ignores these red flags and pushes ahead "to meet the deadline" delivers a project that has to be redone 6 months later. Pausing a week to address it costs far less.
Frequently asked questions about building an AI agent from scratch
Is it possible to build an AI agent without being a programmer?
For a simple proof of concept, yes. Platforms like OpenAI Assistants, Anthropic Projects, Google Vertex and several no-code solutions let you put together basic agents without code. For an agent that connects to your internal systems, takes actions in an ERP/CRM and runs in real production, a technical team is needed (in-house or contracted).
How much does it cost to build an AI agent from scratch?
Simple prototype with no integration: R$ 5 mil to R$ 15 mil. Agent connected to 1 system: R$ 30 mil to R$ 80 mil. Agent connected to 3+ systems with real actions (create an order, generate a payment slip, open a ticket): R$ 80 mil to R$ 250 mil. Plus the monthly operating cost of the model and infrastructure, usually between R$ 2 mil and R$ 20 mil.
What is the first thing to do before building an AI agent?
Choose the problem, not the vendor. Define a single operational process that has high volume, clear repetition and a measurable indicator. That is the starting point. Without that choice made well, any project stays vague and turns into spending with no return.
How long does it take to see the first result from an AI agent?
A prototype talking with real data: 4 weeks. A pilot with real customers in controlled volume: 8 to 10 weeks. A measurable and stable result (indicators in the green for 4 weeks in a row): weeks 14 to 18. Promises of results in under 30 days are marketing, not execution.
Can I build an AI agent using free tools?
For a personal test and prototype, yes. Free ChatGPT, free Claude and free Gemini are enough to experiment. For a business agent in production, no. A free plan does not offer an SLA, contractual privacy guarantees, usage limits compatible with real volume, or support. A serious company comes in with a paid plan from the pilot onward.
