Chatbot vs AI agent: what actually separates them
These two words get used interchangeably, quotes included, and they describe two different things with different costs and different returns. The short version: a chatbot talks, an agent closes the case. Below we show exactly where that line runs and how to work out what your process needs — before you pay for something from the wrong shelf.
- A chatbot's work ends with the answer; an agent performs an action in a system and records the outcome.
- A chatbot with no access to company data creates work: somebody still has to check and file it.
- For simple, repetitive questions a chatbot is enough — and it costs less. The difference starts with cases that have consequences in a system.
One difference that produces all the others
A chatbot receives a question and returns text. All of its work ends in the chat window. An agent receives a case and takes it to the end: it pulls data from systems, makes a decision within the rules it has been given, performs an action and leaves a trace behind — an entry in the ERP, a changed status, a sent reply. When it isn't sure about something, it hands the case to a human along with the context.
Everything else follows from that one difference. An agent needs access to systems, so it needs permissions, logging and escalation rules. A chatbot needs none of that, because it changes nothing. That is why it is cheaper and faster to launch — and why, on cases with consequences in a system, it leaves the work exactly where it was.
The same case, two scenarios
A customer writes: "When will order 4471 arrive and can I add two more pallets to it?" Let's follow both versions.
Chatbot with no integration. It replies that standard delivery takes three business days and asks the customer to contact their account manager about changing the order. The customer writes a second time, now to a human. Team workload: unchanged, plus one extra conversation.
Chatbot with access to knowledge. It finds the order-change rules in the terms and conditions and quotes them correctly. It still has no idea where shipment 4471 is, because it has no access to the ERP. Team workload: less time explaining the rules, everything else unchanged.
Agent with access to systems. It checks the order status in the ERP, verifies that it hasn't been picked yet, checks the customer's commercial terms in the CRM, replies with a specific date and — if the rules allow it — adds the two pallets to the order. Team workload: none, apart from disputed cases.
What your process needs
Ask yourself one question about a typical case: after the answer has been given, does somebody still have to click something in a system? If not, a chatbot is enough and there is no reason to overpay. If so, a chatbot will move the work rather than remove it, because a human will end up in the ERP anyway — only now later, and without the context of the conversation.
In practice most service processes in B2B companies do have that consequence in a system: a status, an entry, a correction, a complaint ticket. Which is why "chatbot or agent" is usually settled not at the technology stage but while mapping the process — when you can see how many steps are left after the answer.
Frequently asked questions
Isn't an agent just a chatbot wired up to an API?
Put simply, that is the technical difference — but the implementation consequences are far bigger. An agent that changes data needs decision rules, a defined scope of permissions, escalation paths and logging of every action. That layer, not the connection itself, is what takes most of the work in an implementation.
We already have a chatbot on our site. Does it have to go?
Usually not. A common arrangement is to keep the existing chat interface and put an agent behind it for selected case types. The customer sees no change, and cases that used to land with the team now finish inside the conversation.
Can an agent work without a chat window?
Yes, and that is how most of the implementations we build work. The agent runs in the mailbox, in the ticketing system or in the background of a process — with no interface for anyone to talk to. Chat is one possible entry point, not a requirement.