An AI chatbot is a program that holds conversations in everyday language: a language model recognises the intent behind your question and generates the answer itself, instead of retrieving a pre-written text. You type a question the way you would ask a colleague and get an answer back, with no menu of options and no queue.
The word “chatbot” is a blend of “chat” and “robot”: a program you have a conversation with. The prefix AI means that a language model builds the answer for each question, whereas an older chatbot picks a pre-written answer from a list.
That single word covers two very different things. ChatGPT is an AI chatbot, and so is the assistant on your energy supplier’s website, but you use the first yourself as a tool, while an organisation deploys the second to talk to customers. The questions that come with each are completely different, and the way AI works explains why.
For a professional in a knowledge-intensive firm, the first type is part of daily practice and the second is an investment decision. Mixing them up leads to requesting quotes for something you don’t need, or to dismissing a tool you could have used.
What is the difference between ChatGPT and a chatbot on a website?
The difference between ChatGPT and an AI chatbot on a website lies in who reads the answer before it lands anywhere. You open ChatGPT yourself, judge what comes out and decide what to do with it. A chatbot on a website answers a customer directly, with no step in between, under the organisation’s name.
Both run on similar technology, and that is exactly why the word causes confusion. The differences are not in the model but in the set-up around it: what information the system is allowed to consult, who sees the output and who is accountable when something goes wrong.
On the user side are the general assistants you open yourself, such as ChatGPT, Claude and Gemini, where you work on documents and draft texts. On the other side is the bot an organisation places on its own channels, connected to a knowledge base and usually to a system the answers come from.
Both also differ from what was called a chatbot five years ago. Those earlier systems searched for keywords and returned a pre-written answer, which worked as long as the question fell within the script and broke down the moment it didn’t. The frustration many people feel at the word “chatbot” comes from that period and says little about what is possible now.
Which of the two you are dealing with determines which questions are relevant. With the first, it is about how you instruct it and whether you check the output. With the second, it is about investment, liability and what the system may know about your clients.
How does an AI-powered chatbot differ from a rule-based bot?
An AI-powered chatbot works with learned patterns, whereas a rule-based bot follows pre-programmed rules. The second recognises keywords and phrases that someone has entered; the first recognises the intent behind a phrasing it has never come across.
In practice, you notice that difference the moment someone strays from the script. If someone asks “when will my order be delivered?”, both systems handle it well. If someone asks “is my parcel on its way yet?”, the rule-based system falls silent, while the model recognises it as the same question. Programming every variant in advance is a hopeless task, because people phrase things in endless ways.
Most chatbots you come across today are AI-powered and run on a large language model. The same technology is behind generative AI, which builds text based on probability instead of looking up an answer in a list.
For organisations, there is an in-between form that matters more than the distinction above. With retrieval-augmented generation (RAG), the system first finds the right passage in your own documents and then has the model formulate the answer based only on that. The model then makes up less, because it draws on a source you supplied rather than on everything it was trained on.
The quality depends almost entirely on the documents you put in.
That set-up is the reason a business chatbot is something other than ChatGPT with a logo on it. The quality depends almost entirely on the documents you put in, and an outdated document produces an outdated answer that looks just as convincing as a correct one.
The table below brings the three forms together, showing for each type how the answer is produced and where it gets stuck.
| Type of chatbot | How the answer is produced | Where it gets stuck |
|---|---|---|
| Rule-based bot | Recognises keywords and returns a pre-written answer | Any phrasing that is not in the script |
| AI chatbot | A language model formulates the answer based on learned patterns | Facts the model doesn’t have but fills in anyway |
| RAG chatbot | Searches your own documents first and has the model answer from them | Outdated or missing documents in the source |
What are some examples of AI chatbots?
The best-known examples of AI chatbots are the general assistants you open yourself, alongside the bots that organisations deploy on their own channels. The general assistants differ mainly in provider, in the underlying models and in how they connect to other software:
- ChatGPT from OpenAI, the chatbot that brought generative AI to the general public in November 2022.
- Claude from Anthropic, strong with long documents and at following instructions closely.
- Gemini from Google, integrated into Gmail, Docs and the rest of Google Workspace.
- Microsoft Copilot, built into Windows, Edge and Microsoft 365.
- Vibe (the new name for Le Chat) from the French company Mistral, a European provider.
On the organisation side, there are three types of application: customer service bots on the websites of online shops, energy suppliers and banks; internal assistants that let employees ask questions of the organisation’s own knowledge base; and project spaces in ChatGPT or Claude where a team sets up its own templates and procedures. That last form requires nothing to be built and is the most realistic starting point for most firms.
What do professionals use an AI chatbot for themselves?
Professionals mainly use an AI chatbot as a work tool for tasks that used to be entirely manual. Summarising documents, drafting a first version, structuring a long text, getting a second look at a line of reasoning: those are the tasks where anyone who works with knowledge and text gains the most.
The time saved is less interesting than the shift in the kind of work. An hour spent drafting an advice letter becomes ten minutes of sketching and fifty minutes of refining, with the thinking shifting from wording to judging. If you only aim for speed, you deliver mediocre work faster.
Which model fits best depends on the kind of work you do, not on which tool gets mentioned most.
The skill that makes the difference is not choosing the tool but instructing it. A vague question gets a vague answer, and an answer you don’t check is an assumption.
When does your firm need its own chatbot?
Your own AI chatbot only pays off if you get enough repeat questions, and most knowledge-intensive firms don’t reach that volume. A bot for customer contact earns its keep by handling the same question hundreds of times a month, not by looking modern.
Run the numbers honestly before you request a quote. At a notary’s or accountancy firm, substantive questions are rarely the same: each question is tied to a case file, a situation and a deadline. A bot that can give the opening hours and directions to the office doesn’t replace a single hour of work.
What does work at that scale is the opposite direction: an assistant that runs internally on your own knowledge base, where employees look up which procedure applies, which template is the latest and what happened last time with a similar case. There is repetition in that, and a mistake stays in-house, where someone can correct it.
Between doing nothing and having a bot built, there is also a step that costs little. In a project space in ChatGPT or Claude, you set up your own templates, procedures and examples once, after which everyone in the firm can ask questions of it without anything having to be built. That is not a chatbot in the commercial sense of the word, but it does solve the problem firms usually consider one for.
So the key question is not whether the technology can do what you want, because it almost always can. The question is whether the same question comes up often enough to justify automation, and whether a mistake in the answer is acceptable where the answer ends up.
What changes once a chatbot speaks on behalf of your firm?
As soon as an AI chatbot speaks on behalf of your firm, responsibility for every answer shifts to you. An employee who makes a mistake corrects it in the next conversation. A bot that makes a mistake does so consistently, with everyone, until someone notices.
From banking practice, I have taken three questions to ask for every AI task: what data goes in, what claims do you base on the outcome and who is liable for the result. With a chatbot for clients, all three become more pressing. The data is your own case file or knowledge base, the claims are statements a client may take as advice, and the liability lies with the firm.
For regulated professions, there is also professional secrecy. A bot that needs case file information to answer a question by definition has access to information that must not reach another client. That is a design question you answer in advance, not a setting you switch on later. What you do and don’t give a system is part of responsible AI use and should be recorded before the first question comes in.
Legislation affects this directly as well. Since 2 August 2026, the transparency obligations of the EU AI Act have applied: a chatbot that talks directly to people must make clear that the user is dealing with an AI system and not an employee, unless that is already obvious from the situation. Those obligations fall under AI Act compliance, together with the arrangements on processing and record-keeping.
Not every task has the same risk profile. A bot that gives the opening hours is miles away from a bot that explains an ongoing procedure. You weigh that up for each type of question, and for the weightier ones the rule I took from my banking years applies: no answer with something at stake goes out without a second check, even if a model wrote it.
What are the limitations of an AI chatbot?
An AI chatbot can produce convincing-sounding nonsense, and that is the limitation that overshadows all the others. The model chooses the most likely answer, not the demonstrably correct one, so an AI hallucination arrives in the same tone as a correct answer.
The quality also depends on the data the model was trained on or draws from. A public model has information about the world up to a certain date and nothing about what happened after that. A business bot has exactly as much information as the documents you supplied, including the errors and outdated versions in them.
Privacy is the third point, and the most underestimated. What you type into a public chatbot is processed on the provider’s servers, and with some account types that information may be used to improve models. Client data and personal data don’t belong there, however convenient it may be at that moment.
The fourth limitation rarely gets attention and costs the most in customer contact. A model is poor at judging when to stop. An employee senses when a question is becoming too sensitive or someone is getting irritated; a bot calmly carries on with an answer that is correct on paper. A good set-up therefore defines in advance which topics and signals trigger an automatic handover to a person.
How do you decide whether to work with AI chatbots?
You decide whether to work with AI chatbots for each task, not as a decision of principle for the whole firm. Start with the question of which tasks come up often and take a lot of time, and only then look at which of them a language model can speed up without the quality suffering.
That order prevents the two expensive mistakes. The first is having a chatbot built for a stream of questions that doesn’t exist. The second is letting employees work out for themselves which tool to use, so nobody knows which information ends up where.
So put the checkpoint in place first. For each task, decide what happens if the answer is wrong and who notices, because that answer determines whether something can run unchecked. The approach I use deliberately starts small, with one to three processes, so that what you learn in the first process makes the second one faster.
Firms that want to set this up in a structured way, with agreements for each type of task and checkpoints that stay in place when a supplier changes something, will recognise this as the first step towards fixed team agreements on AI. The choice of tool doesn’t come first in that process, but follows from the processes you have selected. If you want employees to learn to use chatbots responsibly in their own work first, the AI Literacy Course is a practical starting point.
Frequently asked questions about AI chatbots
Questions about AI chatbots often concern where the term comes from and what you can build with them yourself.
What was the first chatbot?
The chatbot usually named as the first is ELIZA by Joseph Weizenbaum, described in January 1966 in Communications of the ACM. ELIZA ran on MIT’s MAC time-sharing system and made conversations in natural language between people and computers possible. The program searched the input for keywords and built its answer with fixed rules, without a language model.
What is the difference between an AI chatbot and an AI agent?
The difference between an AI chatbot and an AI agent lies in who decides the next steps. A chatbot answers one question at a time and then waits for your next message. Anthropic describes AI agents as systems in which a language model directs its own process and use of tools. An agent can therefore string several steps together, such as opening a file and saving the result somewhere.
Can you make your own AI chatbot without programming?
Yes, you can make your own AI chatbot for personal or internal use without programming by giving an existing assistant fixed instructions. In ChatGPT this is called a GPT, but on 11 September 2026 OpenAI announced that it is phasing out custom GPTs, and creating a new GPT is now only possible in Business, Enterprise and Edu workspaces. In Gemini you create a Gem with a Google account, which tailors Gemini’s responses to a fixed task. Google is replacing Gems with skills: from 17 November 2026, Gems will be read-only; they will disappear from personal accounts on 17 December 2026 and from Workspace in March 2027. An assistant like this runs within the provider’s service and is not a chatbot you place on your own website.
