An agent without states is an employee who can only recite the catalog

A customer writes this to a distributor on a Tuesday morning:

"Hi, send me the same as last month but double. And look into it, the last batch came in damaged."

One message. Three separate tasks.

Find what they bought in March. Check stock and today's price. Open a claim for the previous order.

Almost every bot I see answers one of the three. The honest ones hand it to a human and call it a day.

It is not the model. It was never the model. It comes down to two things, and both are solved in the design, not by writing a longer prompt.

The information base

Most agents get loaded with a fifteen question FAQ and sent out to handle customers.

Fifteen questions will not survive the first Tuesday. What you actually need is purchase history, the current price list, real stock, that specific customer's terms, and the exceptions the senior salesperson knows by heart and never wrote down anywhere.

That last part is the one nobody does. And it is the one worth the most.

An agent without access to that does not answer. It improvises. And it does it with a confidence that should scare you, because the model was trained to sound certain, not to know.

The question to ask before writing a line of code is simple: where is the answer going to come from? If there is no concrete answer to that, there is no agent.

The states

This is the part almost nobody thinks about.

A linear conversational flow works like this: greet, ask, answer, done. It is a form with a better vocabulary.

But checking an order, quoting a new one, opening a claim and booking a visit are four different tasks. Each one needs different data, different tools, and a different way of knowing when it is finished.

A well built agent has one state per task. It knows which one it is in, what information it already gathered, what is missing, and it moves to the right state when the person changes the subject. If the customer starts by asking for a quote, drifts into the damaged batch, and then comes back, the agent comes back with them. It does not start over.

That is the difference between a bot that responds and one that does things.

What it looks like when it is done right

With that opening message, an agent with a broad base and real states does this:

It identifies the customer. Enters the history state and pulls up the March order. Moves to the quoting state, checks stock, builds the new order at double the volume at today's price. Detects a pending claim, opens the after sales ticket with the old order number. It returns all three in a single reply and, if the claim is above a certain amount, escalates it to a person.

That is not magic. That is design.

Before it goes to production

I never put an agent in front of people without a set of conversations built to break it.

Not the easy questions. The other ones. The person who asks three things at once. The one who contradicts themselves. The one who asks for something out of scope. The one with typos. The one who wanders off and comes back four messages later.

That set runs every time a prompt is touched or a tool is added. Without it, every adjustment is a blind bet: you fix one thing, break two others, and find out from a customer.

What I see coming

In LATAM most people still use AI to write an email for them. The small and mid sized companies that connect their real data to an agent with states over the next two years will have an operational edge that no marketing budget can buy.

The winner will not be whoever has the best model. Everyone will have the same model.

The winner will be whoever gave it more information and more things to do.