AI Agent vs Chatbot: the Difference That Decides Whether You Sell

A chatbot picks a reply from a list somebody loaded in advance. An AI agent understands the question however it is written, looks the fact up in your company's information, and writes the answer inside the rules you set. Put briefly: a chatbot replies, an agent resolves.

The distinction matters because both are sold under the same word at similar prices, and because the difference only shows up when a sale is on the line. In a demo, they both look fine.

Three generations, not two

The button bot. It shows a numbered menu, and each option leads to another menu. Good for four fixed questions. The problem is not that it is old: it is that customers type what they want instead of choosing, and that is where it jams. The keyword chatbot. It matches strings in the customer's text and returns the linked answer. It works when the person writes the way the chatbot expects. If your catalog says "twenty-litre drum" and the customer writes "the big water one", there is no match and no answer. The AI agent. It interprets the question, searches your knowledge base, composes the answer from what it found, and decides the next step: ask one more thing, build the order, book a meeting, or hand off to a person. The key phrase there is not "interprets": it is "decides the next step".

Where the difference shows

Four situations that happen daily in a business WhatsApp.

Two questions in one message. "Do you have the five-litre one? and do you deliver south of the city?" The chatbot answers one of them, usually the first, and drops the other. The agent answers both. The customer uses their own words. "How much if I take a whole case?" The chatbot finds no match for "whole case". The agent understands this is the per-case price and quotes the per-case price. The customer switches topic mid-thread. They were asking about a product and suddenly ask about payment terms. The chatbot loses the thread and returns to the menu. The agent answers the new question and picks up where they were. The customer sends a voice note. The chatbot has nothing to do with it. The agent transcribes it and carries on.

None of these is an edge case. All four are a Tuesday morning.

What makes an agent trustworthy

There is a trap worth facing directly: a bare language model, with nothing around it, is worse than a keyword chatbot. The chatbot does not know your price and says nothing; the bare model does not know your price and invents one, in flawless prose. On WhatsApp that reads as a promise from your company.

An agent you can put in front of customers has four pieces:

Its own knowledge base. Your catalog, your prices, your terms. The answer comes from there, not from the model's memory. How that works is explained in what a RAG system is. Explicit rules. What it may answer, what it may not, when it has to hand off, what it can never offer. A discount is not left to the model's judgement: it is forbidden outright. Rule-based handoff. No fact found, a negotiation appears, the customer is upset, the customer asks for a person. Any of the four stops the automatic reply. Conversation memory. So it does not ask the customer again what it already asked yesterday.

A product missing two of those four may call itself an AI agent on its website. It is not one in your WhatsApp.

How to tell them apart before buying

Three questions are enough.

First: where does the answer come from when a customer asks a price? If the answer is "the model knows", it does not. If it is "from your loaded catalog", good.

Second: what does it do when it cannot find the fact? If it answers anyway, that is a risk. If it hands off, it is a serious agent.

Third: who can change an answer, and how long does it take? If changing a price means asking somebody to edit a flow, it will not scale. If it means updating the data source, it will.

A test you can run today

If you are evaluating a tool and want to know which of the three generations it belongs to, send it this: one message with two questions in it, using a word that is not in its catalog verbatim.

Something like "do you have the big one, and do you invoice?". Two questions, and "the big one" is nobody's product name.

A button bot returns the menu. A keyword chatbot answers one of the two, or neither. An agent answers both and, if "the big one" is ambiguous, asks which of the two large ones you mean, exactly as a person would.

The test takes thirty seconds and tells you more than any prepared demo.

What about cost

A keyword chatbot has a fixed, predictable cost: a monthly licence, whatever the usage. An AI agent adds the cost of processing each conversation, which tracks volume.

Put that way, the chatbot looks cheaper. The full arithmetic says otherwise.

If the chatbot resolves three out of ten enquiries, your team pays for the other seven in working hours, which is the most expensive line you have. If the agent resolves seven out of ten, the licence difference is repaid inside the first week.

The right comparison is not licence against licence. It is licence plus your team's hours, against licence plus your team's hours. On that arithmetic, the tool that looks expensive is usually the cheap one.

Migrating from a chatbot you already have

If something is already running, you do not have to throw it out on day one.

First, take the one thing worth keeping from the old chatbot: the questions and answers somebody loaded over the months. That is your business's information, already written, and it is exactly what the agent's knowledge base needs.

Second, look at the old chatbot's conversation log, if it keeps one, and find where it jammed. Those unanswered questions are the best map of what to load.

Third, the switch itself, which takes as long as connecting the number. The customer does not see a migration: they see the same WhatsApp starting to answer better.

What you should not do is run both on the same number in parallel. Two systems answering one customer is worse than either of them alone.

One more thing worth planning: keep the old chatbot's transcripts somewhere you can read them. For the first few weeks after the switch, comparing what the old one answered against what the agent answers on the same question is the fastest way to find gaps in the knowledge base. After that they stop being useful and you can archive them.

When a simple chatbot is enough

If you always get the same five questions, a menu costs less and does the job. Hours, address, directions, a link to the catalog. No need to bring a bigger machine for that.

An agent starts paying off when the questions vary, when the catalog is large or changes often, and when the conversation has to end in something: an order, a meeting, a qualified lead.

An example with numbers

Take one hundred WhatsApp enquiries in a month, in a business with a mid-sized catalog.

A well-loaded keyword chatbot resolves about thirty: the ones phrased the way it expects. The other seventy end in "let me connect you with an advisor", which is the same as having installed nothing, except the customer already lost two minutes.

An agent with a knowledge base resolves sixty to seventy, and the thirty it hands off arrive qualified. The difference between the two tools is not the technology: it is forty conversations a month your team did not have to handle, and several that did not go cold while waiting.

What we use

The WhatsApp Sales Agent pack is an agent, with the four pieces above: knowledge base with your catalog and prices, explicit rules, handoff to a person, and a dashboard with every conversation. Implementation is USD 1,000, the monthly plan is ARS 220,000, and delivery takes fifteen business days.

If your case is closer to bookings and orders than to consultative selling, the AI Receptionist pack covers that in ten business days.

And if you want the mechanism behind the answers, the knowledge-base side is covered in what a RAG system is.

Message us on WhatsApp and try ours before deciding.