Agent or chatbot? The difference that decides the budget

Two quotes for the same request — "we want to automate support": €380 a month in one, €14,000 to implement in the other. Neither was wrong. One proposed a chatbot, the other an agent.

The difference between them isn’t technical. It’s what each one is allowed to do.

Answering isn’t acting

A chatbot answers from what you gave it: an FAQ, a document base, a script. Question in, answer out. When the conversation leaves the script, it hands over to a person.

An agent acts: it reads your CRM, changes a booking, issues a credit note, puts something in the calendar. The conversation is only the interface — the work happens inside your systems.

flowchart LR A([Customer question]) --> B{Needs to touch
a system?} B -->|No| C[Chatbot answers
from what it knows] B -->|Yes| D[Agent reads, decides
and executes the action] D --> E([Human approves
anything irreversible])

The cost is in the action, not the model

The language model costs roughly the same either way — cents per conversation. What separates €380 from €14,000 is everything the action forces you to build:

Integrations with every system the agent touches, and their upkeep when an API changes.
Permissions — what it may change on its own and what has to wait for approval.
An audit trail for every action, for the day someone asks why that booking moved.

A chatbot has none of this because it touches nothing. That’s why it’s cheap — and it’s also why, when it’s enough, it’s genuinely enough.

Two questions before you decide

How many of the questions you receive end in an action inside a system? If it’s fewer than one in five, you’re paying for an agent to do a chatbot’s job.

Can anyone be harmed if the action is wrong? If so, you need human approval — and the agent stops saving the time it promised to save.

When a chatbot is perfectly enough

Opening hours, prices, order status, what a service includes: questions whose answer is already written down somewhere. An agent adds nothing here — it just costs more and fails in more creative ways.

The rule we use: start with the chatbot and let it fail. The questions it handed to a person over one month are the exact list of actions worth an agent. Without that list, you’re buying capability by guesswork.