Chatbot, Conversational AI, AI Agent: Three Words, Three Different Machines
Vendors use the three terms interchangeably. They describe genuinely different systems — and the difference is exactly what decides what you're buying and what it's allowed to do when no one is watching.
A vendor demo, 2026: the slide says "AI Agent," the assistant greets you by name, and three clicks later it hands you a help-center article and a "Did this solve your problem?" button. That is not an agent. It is a decision tree with a language model bolted to the front — and the industry already has a name for the relabeling.
Gartner calls it "agent washing": rebranding chatbots, RPA, and assistants as "agents" without the capability underneath. The words chatbot, conversational AI, and AI agent now get used as synonyms in the same sentence — often the same slide. They are not synonyms. They describe three different machines, and the difference decides what you are actually buying, and what it is allowed to do when no one is watching. If you are shopping, the vendor scorecard is the antidote to the deck.
Chatbot: the interface
A chatbot is the oldest and broadest term, and it names a channel, not an intelligence. The classic version is rule-based: decision-tree logic or keyword matching that maps inputs to a fixed menu of preset replies. The CFPB, reviewing chatbots in consumer finance, described them plainly — one-way streets that cannot respond to requests outside the scope of their data inputs, and that trap customers in "doom loops" when the question drifts off-script.
They are everywhere for a reason: cheap, predictable, fine for password resets and store hours. But "chatbot" tells you the surface — a chat window — and nothing about what sits behind it. A chatbot can be a 2015 decision tree or a 2026 language model. The label is silent on the thing that matters.
Conversational AI: the comprehension
Conversational AI is the capability that lets a system understand language instead of matching keywords. Underneath sit natural-language understanding and intent detection: the model works out what you meant, not just which word you typed, and handles paraphrase, context, and sentiment.
This is a real step up, and it is where most "AI support" lived until recently. But note the ceiling. Conversational AI is about comprehension and response, not action. It can understand that you want a refund, explain the refund policy fluently, and surface the exact right article — and still not issue the refund. It talks well. It does nothing on its own.
AI agent: the autonomy
An AI agent — the agentic layer — is defined by what conversational AI lacks: the ability to act. A real agent plans, reasons over context, calls external tools and APIs, holds memory across the conversation, and works toward a resolution rather than a reply. It reads the order, checks the policy, issues the refund, and updates the record — steps, not just sentences.
That is the true dividing line, and it is not eloquence.
The line between a chatbot and an AI agent isn't how well it talks. It's what it's allowed to do when no one is watching.
On paper, the market is sprinting toward this layer.
And vendors quote the payoff: Intercom markets a 76% average resolution rate for its Fin agent.
Treat marketed numbers with the usual suspicion — production deployments routinely land lower, and "resolved" is a slippery word. What "70% resolved" actually means and the resolution number behind Fin are each their own argument.
Why the blur costs you
The distinction is not pedantry; it is a risk boundary. A chatbot that surfaces the wrong article wastes a minute. An agent that acts on a wrong conclusion issues the wrong refund, changes the wrong address, closes the wrong account. The more autonomy you buy, the more governance you need — guardrails, confidence thresholds, a human offramp, and observability into what the thing did and why. When the model is fluent but wrong, it makes things up with conviction. Buy "an agent," get a chatbot, and you overpay. Buy "a chatbot," get an unsupervised agent, and you have shipped a liability. The hype is already correcting:
That is not a reason to sit out — it is a reason to know which of the three you are actually deploying, which is the whole point of separating the hype from what agents do now.
The one-question test
Skip the datasheet. Ask what the system can complete without a human: read the account, process the return, reschedule the delivery. If the honest answer is "it surfaces the right article," you are looking at a chatbot with better manners — conversational AI, perhaps, but not an agent. If it can take the action, and be wrong in ways that cost real money, it is an agent, and the conversation should be about guardrails, not vocabulary.
Pick the word that matches the risk you are taking on. The vendor picked the word that closes the deal.