Chatbot or AI Agent for Your Business: Which One, and When a Chatbot Is Enough

A chatbot converses; an agent works. The first answers in a chat window, from scripts or documents it was given. The second is connected to your tools (mailbox, CRM, management software): it finds the customer, reads the history, prepares the reply or the document, and flags what it cannot handle.

There are degrees in between, and vendors blur them. Three levels are fairly distinct.

The same distinction, seen from the service side, is covered in our article on what an AI agent actually does in a business, and what it does not.

When the answer already exists in writing and does not depend on who is asking. Opening hours, service areas, documents needed for a file, how to file a complaint: a scripted chatbot, or a well-kept FAQ page, handles these without connecting to any of your tools.

It is the cheapest case to set up and the most predictable, and it is often the right starting point. The limit shows quickly, though: as soon as a customer asks "where is my order?" or "can you redo the quote with the other reference?", the chatbot can only point to a form or a phone number. The customer made the effort to write and has to start again somewhere else. If those requests are frequent in your business, the chatbot moves the work instead of removing it.

When answering means fetching information specific to the customer, from a tool the chatbot cannot see, and someone currently spends time doing it. Three signs tell you.

• The answer depends on the file. Order status, negotiated price, history of exchanges, availability: the information sits in your management software or your CRM, not on an FAQ page.
• Requests arrive in writing, not only in a chat window. Many small businesses receive most requests by email or web form. A chatbot on your website does not see them; an agent reads the mailbox.
• The expected result is a document or an update: a proposal, a quote, a customer record brought up to date. A chatbot informs, an agent prepares. Our requests and quotes agent page describes this, from the email received to the proposal ready to approve.

One point comparisons tend to skip: an agent does not replace a fixed rule when a rule is enough. An acknowledgement of receipt, a sort by keyword, a reply template in your mailbox all work without any AI, more simply and at lower cost. An agent earns its place when something has to be read, retrieved and written.

It can get things wrong, and what it writes commits your business. A chatbot that announces a delivery date you cannot meet, an availability that is not accurate or a discount nobody approved creates a dispute, not a time saving.

Three rules are set before go-live, whatever tool you choose.

1. What it may say on its own: information already written down, with its source. Nothing else.
2. What goes through a person: anything that commits the business (a price, a deadline, a goodwill gesture, an answer to an unhappy customer). The system prepares, the person reads and sends.
3. What it does when it does not know: it says so and hands over to someone. It does not guess.

When the system talks directly to your customers, they also need to be told they are speaking to an AI: that is a transparency obligation under the European Union's AI Act, detailed in our article on answering customer requests with an AI agent (this is not legal advice). The easiest mode to control is the one where the agent prepares and a person sends: the reply then goes out in the name of someone who has read it.

By looking at your real requests, not a brochure. Take the last fifty customer requests you received, from every channel, and sort each into one of three boxes.

• The answer is written down somewhere and the same for everyone. That is the ground of a scripted chatbot or an FAQ page.
• The answer depends on the customer or the file, but a person finds it in two minutes. That is the agent's ground: it does the same lookup, and the person reads the result.
• The answer needs judgement: negotiation, complaint, special case. That stays with a person, with at best a prepared file.

The split between the three boxes tells you where to look. If the first dominates, start with a chatbot or an FAQ page. If the second weighs most, it is the agent. If the third dominates, no tool will save you much, and it is better to know before spending anything. This count is part of what we look at with you, on one task, during the discovery audit.

Through a discovery audit on a single task, then a written scope before any paid work. The first system is built inside your existing tools, with no extra tab to open, and it runs in trial for two weeks before go-live.

• The discovery audit (20 to 30 minutes, by video call or on site): the task, your tools, the decisions that must stay human. You leave with a written opinion: fixed rule, AI agent, or nothing worth building.
• The written scope: what the system does on its own, what goes through a person, what it never does, and the quote, before work starts.
• The build: read-only access to your data first, approval rules, tests on your past requests.
• The trial period, two weeks: the system prepares without sending anything, you compare its proposals with what your team actually answered, then you decide on go-live.

To see how we choose the first task to give an agent, and the ones better left alone, read AI agents for small businesses: where to start, and when to skip them.