AI agents: what you learn once they start selling

I searched for “AI agents” and read the top ten results. All ten explain what an AI agent is. They are published by Google, IBM, Microsoft, Salesforce and a consultancy. And above all ten, Google now places its own automatic summary with the definition.

So the definition is covered. What I did not find in any of the ten is what happens when you put one in front of real customers and something breaks.

That is my job, so here is the missing part.

What an agent is, in one sentence

An agent reads the question as written, checks whatever it needs before answering — stock, order status, history — and decides what to do. A tree of numbered options is not an agent: it is a form dressed up as a conversation, and the customer spots it in the first message.

The practical difference is not the technology. It is whether the system can answer something nobody wrote in advance.

The one I have put into production most often

The most requested agent is not support. It is sales.

It introduces itself, works out what the customer is looking for, suggests specific products, answers the questions that come up about those products and carries them through to the purchase without a person stepping in. In an online shop, that entire conversation — from “hi, I need a gift” to payment — can happen without anyone on the team reading the chat.

To see why that matters, put numbers on a mid-sized shop. Three hundred enquiries a day, three people answering. If each conversation takes eight minutes and half are the same question about sizes, delivery or availability, the team spends the whole day repeating itself. And the enquiries that land at eleven at night wait until morning.

That is where sales are lost, not on price. Someone asking on WhatsApp is deciding now. If they do not get an answer within minutes, they do not come back tomorrow: they open another tab. There are large brands, with plenty of budget, running an open WhatsApp channel and taking days to reply. Every one of those days is someone buying elsewhere.

What I did not see coming the first time

That the hard part is not getting the agent to answer well. It is getting it to know when to shut up.

An agent that answers everything with confidence is worse than having nothing. If the customer asks something that is not in the data — a warranty exception, an odd delivery case, a promotion that expired yesterday — the agent has to recognise it cannot go there and hand the conversation to a person with the context already summarised. That handover is the most neglected part, and the one that decides whether the customer stays or leaves.

The second thing I learned: an agent inherits your data errors. If catalogue stock does not match real stock, the agent will confidently sell something that does not exist. That is not a model failure, it is a source failure — but the angry customer does not make the distinction.

The effect on the bill nobody mentions

There is a cost angle that gets overlooked. On WhatsApp, messages inside the 24-hour window opened by the customer are not charged. A conversation resolved inside that window generates no messaging cost.

Answering in seconds does not just sell more: it keeps the conversation inside the free window. Answering the next day forces you to reopen with a template, and that one is billed.

When I tell a client they do not need one

When the conversations do not repeat.

An agent pays off when there is volume and pattern. If you handle twenty enquiries a day and each is different — bespoke projects, quotes that depend on twenty variables, clients who need to speak to someone who can decide — an agent adds a layer between you and your customer without solving anything.

And there is an even clearer case: when the problem is the product. If half your messages ask where the order is, support is not the problem, order tracking is. An agent would answer those faster and paper over the symptom for a year.

Before building anything, look at how many of your conversations are the same question repeated. If there are many, an agent will change your business. If there are not, it will cost you money and you will not notice the difference.

How I would build it today

Start with a single task, the most repeated one, and get it working properly before expanding. An agent that flawlessly resolves the most common question is worth more than one doing ten things halfway.

Connect it to your real data from day one. An agent that cannot check stock or order status is a chat with good copywriting, and you already have that.

And decide from the start what it will not answer. That list matters more than the list of what it will.

On where all this lives: the agent handles the conversation, but you need something to route and archive once there is a team behind it. I cover that in what to check in a WhatsApp CRM before paying for it, and why the two do not compete: the CRM organises, the agent answers.

I post examples of agents in the wild on @WhatsMarketing_es.

If you want me to look at your case, get in touch. I work from Buenos Aires with clients in Mexico City, Argentina, the rest of Latin America and Spain.