AI Agent or Chatbot? The Distinction Is About What Happens to the Unexpected Question

The words are used interchangeably in marketing and they describe two genuinely different things. The difference is not sophistication; it is what happens when someone asks a question you did not anticipate.

A chatbot follows a tree you drew. Trigger, branch, message. It is deterministic, which means it is predictable, auditable and completely stuck the moment reality goes off-script.

An AI agent reads the question, finds the answer in material you supplied, and composes a reply. It is not following your tree; it is using your knowledge.

Why does the distinction matter in practice?

Because of the shape of real inboxes. Read fifty Direct conversations in almost any business and you get a short head and a long tail: three or four questions that repeat constantly, and then a very long list of things asked once.

A chatbot is excellent at the head. Fixed question, fixed answer, instant, cheap, no risk of invention. For the three questions that make up most of your volume, a rule is the right tool and an AI is overkill.

A chatbot is useless at the tail, and the tail is where the money often is. “I bought the blue one in March, it has started peeling, is that covered” is one question that will never repeat in that form, and no tree has a branch for it.

What does “AI-powered” usually mean on a pricing page?

At least four different products, and they are worth telling apart before you pay for one:

1. Keyword matching with synonyms. Marketed as AI. It matches more phrasings than a plain rule and that is genuinely useful, but it is still picking a branch you wrote.
2. Intent classification. A model decides which of your branches the message belongs to. Better at the head, still nothing for the tail.
3. A generative reply with a prompt. A model writes an answer from a system prompt describing your business. Flexible, and dangerous — it will confidently invent a return policy, because a prompt is not a source.
4. An agent grounded in your material. The model retrieves from documents you supplied and answers from them, saying so when it cannot find an answer.

Only the fourth is what people mean when they say an AI agent replaces a person, and the difference between the third and the fourth is the whole risk profile.

How can you tell which one a vendor is selling?

Three questions, and they cannot be answered vaguely:

• “What is it trained on?” If the answer is “it understands your business” with no mention of documents you upload, it is a prompt.
• “Can I see why it said that?” A grounded agent can point at the paragraph it used. If you cannot trace an answer to a source, you cannot fix a wrong answer — only complain about it.
• “What does it do when it does not know?” “It always finds a way to help” is the wrong answer. The right answer describes a handoff.

So should you use one or the other?

Both, and in this order.

Rules for the head. The three or four questions that repeat daily. Instant, exact, zero cost per answer, no possibility of invention. The method for finding yours is in building an FAQ bot from the questions you already get.

An agent for the tail. Everything else, grounded in your own material, with an explicit instruction to hand over rather than guess.

A person for the rest. Complaints, money, anything where being wrong is expensive.

A setup that uses an agent for the head is paying for a model call to answer “what time do you close”. A setup that uses rules for the tail is losing the conversations that mattered most.

Does an AI agent actually replace a person?

It replaces a share of the typing, and the honest number depends on how good your material is rather than how good the model is.

An agent with a thorough knowledge base — real price list, real policies, real service descriptions — handles a large majority of routine questions without a person. An agent given three paragraphs about how great the company is handles almost nothing and invents the rest.

What it does not replace is judgement. Deciding whether to make an exception for an unhappy customer is not an information problem, and an agent that makes that decision will eventually make an expensive one. That is a design choice, not a model limitation: the handoff condition is a feature you configure. How to think about it is in when an AI should hand off to a human.

What does it cost to run an agent versus a rule?

A rule costs nothing per answer. An agent costs a model call, and usually more than one per conversation: understanding the question, retrieving the answer, sometimes a clarification, sometimes a decision about handing off.

When you compare vendor quotas, the useful arithmetic is calls per conversation rather than the headline number. A quota of 1,000 “AI responses” may be 300 real conversations. It is also worth knowing what happens when the quota runs out — whether the agent stops and someone is told, or whether it stops silently and your inbox quietly gets worse.

Where should the knowledge come from?

From what you already have, not from a document written for the bot. Your price list, your shipping page, your returns policy, your service descriptions, the answers you have typed a hundred times.

Material written specifically for an AI has a habit of drifting away from the real thing, and then the agent is confidently correct about a policy you no longer have. The practical method is in training an AI agent on your own knowledge base, and the Instagram-specific version is in an AI chatbot for Instagram DMs.

Chatonaut’s AI agent is the grounded kind: you give it your own material, it answers in the customer’s language, and it hands over when it should. Plans are on the pricing page.