Chatbot: definition, types, uses and limits

In France, the official French term is “dialogueur”, published in the Journal officiel of 9 December 2018 as part of the scheme for enriching the French language, with “agent de dialogue” as a synonym. The definition given on FranceTerme, the official French terminology database, is software specialised in natural-language dialogue with a human, able in particular to answer questions or trigger the execution of tasks. The same entry notes that the expression “agent conversationnel” (conversational agent), which is often seen, is not recommended.

The CNIL, France’s data protection authority, in advice published in February 2021, describes chatbots as software that allows a user to hold a dialogue with a program designed to provide them with information, in particular to answer the most frequent questions in a targeted and interactive way. The two definitions have one thing in common: a chatbot is defined by its function, holding a dialogue, not by the technique it uses. That is why the same word covers both a button menu and an assistant based on generative artificial intelligence.

A rule-based chatbot follows a path planned by its designers. It offers buttons, recognises keywords or asks a series of questions according to a decision tree: “Would you like a quote?”, then “For what type of service?”. Its behaviour is predictable and easy to check, since every answer has been written in advance. On the other hand, it can only handle what has been planned: a question worded differently, or one that falls outside the script, defeats it.

A language-model-based chatbot works differently. It relies on a large language model, a program trained on very large amounts of text, which writes its answer from the question asked and the information it is given. It understands varied wording, follows the thread of an exchange over several messages and expresses itself naturally. This flexibility comes at a price: since the answer is no longer written in advance, it has to be framed by instructions and reference content.

Many tools combine the two approaches: a language model to understand and answer, and fixed steps for the moments that demand rigour, such as entering an email address or confirming an appointment.

The most common use is the one the CNIL mentions: answering frequently asked questions. On a website, a chatbot can present your services, give your opening hours, explain a procedure or list the documents to prepare, at the moment the visitor has the question and without them having to dig through the pages.

It is also used to direct and collect: understanding what the person is looking for, asking them the right questions, then recording their request and contact details so that the team can follow up. Connected to a calendar, it can offer a time slot. Internally, the same principle applies to employee support, for example to find a procedure. Points of contact vary: a chat window on the website, messaging apps, a customer area.

A language-model-based chatbot can be wrong with complete confidence. In a research paper published in September 2025, Adam Tauman Kalai and his co-authors observe that language models, like students facing a hard exam question, sometimes guess instead of acknowledging their uncertainty, and argue that this is because their training and evaluation reward attempting an answer. This risk is reduced by restricting the chatbot to the information supplied by the business and explicitly asking it to say when it does not know.

The second limit is data. The CNIL points out that a chatbot often processes personal data, if only to keep a record of the conversation. The tracker it needs in order to work can be exempt from consent if it is placed at the user’s express request, for example when they open the chat window, and is used only for that purpose. The retention period must be suited to the purpose, and a conversation without human intervention cannot on its own lead to an important decision for the person, such as refusing a credit application or making it impossible to submit a job application. Because users write freely, they may also share sensitive data, which calls for particular care.

Finally, a chatbot does not replace a team: it is best to provide a way through to a person at all times when a request goes beyond its scope, and not to hide that it is a program. The transparency obligations linked to artificial intelligence are covered in our guide to AI regulation.

A chatbot is not an AI agent. A chatbot holds a dialogue; an AI agent pursues a goal by using tools to act, for example checking a calendar or creating a record, and can work without any chat window at all. The boundary between the two is covered in detail in our guide to the definition of an AI agent.

Nor is a chatbot synonymous with conversational AI: that term covers all the techniques that allow a machine to understand and produce language in a dialogue, and a chatbot is only one application among others. When the dialogue happens by voice, over the phone, the terms callbot or voicebot are more usual, two concepts covered in dedicated guides.

Nor is it a search engine: a well-designed business chatbot does not answer everything, but works within a defined scope. To move on to setting one up, our guide to building a chatbot sets out the steps.

A few questions are enough to judge a chatbot. What information does it rely on, and who keeps it up to date? What does it say when it does not know? Can you reach a person or leave your contact details at any time? Where are conversations stored, and for how long? The most telling test is still to ask it your customers’ real questions yourself, including the clumsiest ones and those furthest from your business.

Alyas offers a chat agent that is installed on your website as a widget, and can then answer on WhatsApp, Messenger or Instagram once those accounts are connected from your dashboard, subject to your accounts’ permissions and an availability check. It answers from the information you provide (website, text, PDF documents); if it does not know, it says so and offers to pass on the request. It records visitors’ contact details, can book an appointment if Google Calendar or Cal.com is connected, and when a conversation needs to be handed over to a human, it is flagged and you are notified by email. The trial lasts 10 days, with no credit card required.