AI sales agent vs chatbot: the difference is what it hands you

Almost every product in this category describes itself the same way, which makes the category almost impossible to shop in. Conversational. Intelligent. AI-powered. Every vendor demo shows a friendly exchange in a chat window, and every friendly exchange in a chat window looks the same.

There is a simple test that cuts through it, and it has nothing to do with the conversation.

Ask what arrives at the end

When a good prospect finishes talking to the thing, what lands in your inbox?

If it is a transcript and a first name, you have bought a widget. Someone on your team now has to read a conversation, work out what was established, decide whether it is worth a call, and start from the beginning when they ring.

If it is a brief, the facts established, the facts still missing, the agreed next step, and anything the software promised the customer; you have bought something that did a job. Your salesperson opens the call already knowing the budget, the timeline and how it is being paid for. Nobody is asked twice for something they already said.

A widget waits. An agent qualifies. The conversation is the visible part; the handover is the product.

Three practical differences

What it is trying to achieve. A chatbot succeeds by answering the question. An agent succeeds by answering the question and establishing the six or eight facts that decide whether a sale is possible, without the conversation turning into a form.

What happens when the visitor leaves. This is the sharpest divide. For a widget, a visitor who closes the tab is gone; the session ends and the transcript is filed. A visitor who goes quiet has not necessarily lost interest, very often they got interrupted. Continuing that same conversation later, by email, without re-introducing itself and without asking again for things already given, is a different architecture rather than a feature.

What it will not say. A general-purpose chatbot will answer a pricing question because answering questions is what it does. That is exactly how a company ends up bound by a discount that does not exist, a tribunal has already ordered an airline to honour one its chatbot invented. An agent built for sales has to treat figures as a controlled substance: read from approved data, checked before sending, refused when absent.

What neither of them is

Neither is a salesperson, and any vendor implying otherwise should be asked what happens in the hard cases. Customers know this better than the industry does. Gartner found, from 3,566 respondents surveyed in early 2026, that 87% say a company using generative AI for customer service must provide access to a human agent.

Read that as a design requirement rather than a criticism. The failure mode people hate is not talking to software; it is being trapped with software while a human is visibly unavailable. An agent that hands over cleanly, with everything it learned attached, is meeting an expectation. One that keeps trying is testing a customer’s patience on your behalf.

The category is being rewritten right now

Worth knowing if you are shopping: Drift, which largely invented the vocabulary of conversational marketing, has been folded into another product, and its original benchmark reports are no longer at the URLs everyone cites. The language survives its author.

Which is a reasonable moment to notice that the interesting part was never the chat window. Live chat was a delivery mechanism. What actually mattered, talking to people while they are still on the page, and knowing something useful about them afterwards, is a problem, and the problem outlived the product category built on it.

Four questions to ask any vendor, including us

1. Where do prices come from? If the answer is “the model knows them”, stop.
2. What checks the output before it is sent? An instruction in a prompt is not a check.
3. What happens when it does not know something? The good answer is a plain admission and a human follow-up.
4. Show me the handover. Not the conversation. The thing that arrives when the conversation is over.

Those four questions are worth asking of anything in this category, from anyone, ourselves included. If a vendor cannot answer the fourth one with a real example, the first three do not matter much.