AI agent, chatbot or automation: which one for which need

One axis: who decides the sequence

People often contrast the chatbot, which talks, with the agent, which acts. That is inaccurate. A chatbot connected to your order system looks up a file, opens a ticket and updates a record, exactly like an agent.

The difference is elsewhere, and it is gradual rather than sharp. It comes down to one question: who decides the sequence of operations.

• In a classic automation, you do, entirely in advance. When this form is submitted, create this record, send this message, update this table.
• In a chatbot, you do too, with latitude in the wording. The tool understands different phrasings, but the paths it can take are defined.
• In an agent, the system does, within limits you have set. It chooses what information to fetch, in what order, and what action to take.

Price, framing effort and supervision needs follow exactly that order. Which is why it pays to ask the question in that direction: can the need be covered by an automation? If not, by a chatbot? If not, by an agent?

When classic automation wins

If the handling can be described entirely by fixed rules, classic automation beats an agent on every count: cheaper, faster, fully predictable, verifiable line by line.

Moving data between systems on a constant pattern, triggering a reminder from an event, generating a document from structured fields: in those cases, adding intelligence only adds uncertainty.

The test is simple and reliable: you can write all the rules without ever using the words "depending on the context". If so, stay with an automation.

The common mistake is deploying an agent where fifteen rules would do. It costs more, complicates verification, and introduces variability nobody needed.

When a chatbot is enough, and how far it goes

Two families need separating, because they do not share a domain.

An informational chatbot draws on your public documents. It explains opening hours, terms, return procedures, how a product works. When around twenty questions carry most of the volume, it handles a large share of requests for a fraction of the cost of an agent.

A transactional chatbot is authenticated and connected to your systems. It says where an order stands, opens a ticket, records a service request. So it does far more than talk.

Its limit is not data access, it is that every action stays planned in advance. It does the three or four things it was told to do. It does not decide that a case falls outside the frame and deserves something else.

When an agent becomes necessary

An agent earns its place when the content has no fixed shape and you have to understand before acting.

• Freely written messages, where the intent must be identified before anything else.
• Heterogeneous documents: invoices from twenty suppliers, each in its own layout.
• Cases requiring several sources to be compared before concluding.
• Decisions resting on rules nobody has ever managed to reduce to a tree.

A practical marker: if the answer to "what should be done here?" depends on the content of each case, you need an agent. If it depends only on fields known in advance, you do not.

An agent is also paid for in supervision. Since it decides the sequence, you have to be able to know afterwards what it decided and why. That line does not exist on an automation.

If an agent is indeed the answer, here is the full sequence: how to deploy an AI agent in a company

What each one costs

The budget gaps between the three families are an order of magnitude apart, and they read easily.

• A chatbot grounded in your documents: a few thousand euros to set up, plus a small usage cost.
• A classic automation: a few days of work per scenario, also a few thousand euros.
• An agent in production: from 8,000 euros, plus usage by volume and permanent supervision.

In practice many companies combine all three, and that is the most solid configuration: a chatbot on the website for general questions, automations for regular flows, and an agent at the one place where interpretation is indispensable.

Each block then stays in its place and can be replaced independently. And when something goes wrong, you know which part handled what.

Four questions, in this order

Asked in this order, four questions lead to the right tool with no technical discussion at all.

• Is it only about informing, without consulting data specific to the person? An informational chatbot is enough.
• Can the handling be described entirely by fixed rules? Classic automation is preferable.
• Does the content vary in shape and wording? An agent becomes relevant.
• Does the system have to choose the course of action itself? An agent is necessary.

A fifth question decides the schedule more than the choice: does the volume justify the investment? A rare task, however complex, is often still better handled by a human, and saying so is part of honest framing.

If you are hesitating on a specific need, the quickest route is to gather ten real cases as they arrive with you. The right tool is settled in one session, and it is often simpler than expected.