Chatbot or AI assistant: what is the difference for an SME?

What is a classic chatbot?

A first-generation chatbot follows a decision tree: buttons, a few keywords and answers written in advance. It does its job well on a short path that is the same for everyone, such as choosing a service or booking a slot. As soon as a customer asks a question in their own words that falls outside the script, the path can get stuck and send them to a form.

This approach remains useful when the choices are known and stable. A shop can direct a visitor to opening hours, deliveries or returns. The team controls the text of every answer, but has to plan and maintain the branches. Adding buttons is not always enough to handle a request that mixes several topics.

An older idea than most people think

Conversation programs have existed since the 1960s. In 1966, Joseph Weizenbaum, a researcher at MIT, published in Communications of the ACM the description of ELIZA, a program that imitated an interview by spotting keywords and sending back stock sentences. The principle changed little for decades: recognize a word or an intent, then display the answer planned for that case.

In the 2010s, many tools added a layer of language understanding. Natural language processing made it possible to classify a sentence into a known intent ("track my order", "change my address") even if the customer did not type the exact word. The answer was still written in advance: the program understood the question better, but could not answer what had not been planned.

How a scripted chatbot works in practice

A scripted chatbot rests on three elements that the team has to build and maintain:

• A path tree: each button leads to a step, each step to an answer or a new question
• A list of intents or keywords: to recognize what the visitor wants when they type instead of clicking
• A library of answers: fixed texts, reviewed once, displayed as they are

This approach has real strengths. Every answer is known in advance, which is reassuring on sensitive topics. The computing cost is low. The path is predictable, so easy to test. The trade-off lies in maintenance: every new question requires a new branch, and a catalogue that changes means rewriting several answers by hand.

What does an AI assistant do on top of that?

An AI assistant relies on a language model and on your documents to answer freely worded questions. It interprets the question, looks for the useful passage in your pages, FAQs, terms or guides, then writes an answer. The differences that matter for an SME are these:

• It handles unexpected wording: it can find information without a script having been written for every way of asking
• It shows where its answer comes from: source citation lets the customer check and lets your team spot an outdated document
• It can hand over: when the answer is missing or calls for a decision, the instructions must provide for a handover to your team

These capabilities depend on the documents, the settings and the tests. A fluent answer proves neither that the right document was found nor that the assistant is allowed to make a commercial commitment. Its limits must be as explicit as its role.

What happens between the question and the answer

The general public discovered generative AI with ChatGPT, launched by OpenAI on 30 November 2022. A business assistant uses the same kind of model, but within a much narrower frame. Handling a question usually follows four steps:

1. Retrieval. The system searches your knowledge base for passages whose meaning is close to the question, using semantic search that does not depend on the exact words
2. Instructions. The model receives the question, the passages found and a prompt that sets its role, tone and limits
3. Writing. It writes an answer from the passages, in the visitor's language
4. Checking. The answer is shown with its source; if nothing matches, the instructions say to admit it and offer a contact

This method has a name: RAG, for retrieval-augmented generation. It lets you update the knowledge by changing a document, without retraining the model through fine-tuning.

What an AI assistant does not do on its own

A language model produces plausible text. Without documents and instructions, it can fill the gaps with invented information: a hallucination. A serious assistant reduces that risk with several guardrails, but does not remove it. Nor does it know the state of a file, a stock level or a calendar, unless it is connected to the tool that holds that information. And it does not decide on the company's behalf: a discount, an exception or a goodwill gesture remains a human decision.

Chatbot or AI assistant: the comparison point by point

The two approaches are not opposed on everything. The table below compares what matters for an SME that wants to answer the visitors of its website.

The table shows a shift in the work. With a script, the effort goes into writing paths. With an AI assistant, it goes into the quality of the documents and the review of conversations. In both cases, someone must remain responsible for the content.

When is a scripted chatbot enough?

A scripted chatbot is enough when requests are repetitive, choices limited and answers stable. If your visitors mostly look for an address, opening hours or a form, a good FAQ can also do the job. An AI assistant becomes worthwhile when the wording varies and the answers are spread across several documents.

Receiving questions in the evening or at the weekend can justify a document-based answer available at any time. That does not make the team available around the clock: the hours when the team takes over must stay visible.

Three signs that a script is no longer enough

Some signals show that a button path has reached its limits:

• The contact form fills up with questions whose answer already exists on the website, on a page the visitor did not find
• The team copies the same explanations by e-mail or on the phone, with slight variations depending on the customer's context
• The path grows every quarter: new branches are added, old ones are out of date, and no one knows which are still used

In these situations, the problem rarely comes from a lack of buttons. It comes from the gap between how customers ask their questions and how the information is organized.

Mixed approaches

Many organizations combine the two. A few buttons direct the most frequent requests ("follow up a request", "book an appointment"), and a free text field lets visitors ask the assistant something else. The buttons can also serve as suggestions at the start of the conversation: they show the visitor what the assistant can do without locking them into a path. This combination keeps predictability where it counts and flexibility elsewhere.

How much does each approach cost?

The cost is not limited to the subscription. It includes preparation, maintenance and the team's time spent reviewing. A scripted chatbot costs mostly in writing and updating paths. An AI assistant costs in preparing documents, model usage and follow-up.

Self-service tools often show entry plans of a few dozen dollars a month, with a bill that rises with the number of agents, conversations or generated answers. An accompanied set-up adds an upfront amount but delivers an assistant already fed with your documents. Detailed public prices, checked and dated, are in our article on the cost of an AI assistant for a Belgian SME, and tool-by-tool comparisons in the comparisons.

To compare fairly, calculate the cost over a year rather than a month. Add the time your team spends preparing, reviewing and correcting. A cheap solution that needs two hours of maintenance a week can end up costing more than a higher plan that includes that follow-up.

Which documents should you prepare before choosing?

Prepare the questions that keep coming back and the reliable documents that already answer them. The quality of an AI assistant depends above all on what you give it to read. Gather your commercial terms, practical guides, opening hours and answers approved by your team, then remove outdated versions.

For each topic, name a person who can confirm the answer. If no one knows which condition applies, the assistant is likely to reproduce that ambiguity. Also separate publishable information from internal notes and personal files. A document that is useful to a colleague is not automatically suitable for a public conversation.

Write exceptions with as much care as the general rule. "Delivery available" leaves open questions about the area, the conditions and availability. A precise text reduces interpretation and makes answers easier to check.

A simple method to list the questions

Before comparing tools, spend an hour noting the real questions. Three sources are often enough:

1. E-mails and messages from the last three months: note the questions as they were worded, without rephrasing them
2. Phone calls: ask the person who answers to list the ten questions they hear most
3. The website's internal search, if there is one: the terms typed show what visitors cannot find

Then group these questions by topic and, for each topic, note the document that contains the answer. Topics without a document are the first work list. Topics whose answer depends on a file or a decision are the list of what will go to the team.

What does the difference look like in a Belgian SME?

The difference shows mostly when a request combines a particular context with information found in the documents. The following situations are fictitious examples, not real customers.

In a shop in Brussels, a visitor asks whether an item ordered online can be exchanged in store. A script can send them to the "Returns" section. A document-based assistant can find the matching condition, explain it and cite it. If the customer asks for an exception, it hands over to the team instead of granting it.

In a coliving house in Ghent, an English-speaking applicant asks whether a late arrival is possible and how to collect the keys. The assistant can bring together the useful passages of the arrival guide and answer in English. It must, however, tell a general procedure apart from a permission tied to a specific booking.

A property manager in Liège can explain the general steps for reporting a problem in the common areas. Information about an individual lease belongs in the appropriate exchanges with the team. The choice of tool therefore depends as much on the document scope as on how it converses.

The same example with both tools

Take a fictitious accounting firm in Namur. A business owner writes on a Sunday evening: "I have just set up my company, do you also handle VAT and how long before we can start?"

With a scripted chatbot, the visitor sees three buttons: "Our services", "Our rates", "Contact us". They click "Our services", read a list, then fill in the form. On Monday the firm receives a message without context and has to call back to understand the request.

With a document-based AI assistant, the answer says that the firm handles VAT returns for companies, cites the services page, explains that the start date depends on a first meeting, then offers to leave contact details. On Monday, the team finds in its inbox the full conversation, the type of company and the need expressed. The decision to take on the file and the price stay with the firm.

How do you check that an assistant really meets your needs?

Test the assistant with representative questions, including questions your documents do not answer. Ask for a demonstration based on your own content rather than a rehearsed exchange on a generic topic.

• Common question: the answer is clear and points to the relevant passage
• Rephrased question: the customer's vocabulary leads to the same information
• Ambiguous question: the assistant asks for a useful clarification before concluding
• Missing information: it admits the limit and offers a concrete next step
• Sensitive request: it hands over to the team with the necessary context
• Change of language: conditions and caveats keep the same meaning

Write down the expected answer before the test so you do not judge the style alone. A friendly wording can hide a mistake about a condition. Repeat the relevant tests after any significant change to the documents. Our guide on stopping an assistant from inventing answers gives a full test grid.

The questions to ask the supplier

A convincing demonstration does not tell you everything. Ask these questions too, and ask for written answers:

• Where are the conversations and documents hosted? And which subcontractors can access them
• Are conversations used to train a model? At the tool's supplier and at the model's supplier
• How long is data kept? And can you set that retention period yourself
• How does the bill change if volume doubles? A busy month should not become a surprise
• Who reviews the questions left unanswered? You, the supplier or no one
• How does the team take over a conversation? By e-mail, in a shared inbox or through another tool

Can an AI assistant replace the person who answers customers?

The assistant can take on documented answers, but decisions, exceptions and sensitive situations call for a human organization. Human handover lets the team read the history and continue the request in a shared inbox.

A plumber can thus receive a description of the need and the useful contact details before calling back. The assistant must not turn a call-out request into a confirmed appointment without reliable information. For the business, the aim is fewer pointless exchanges and a better-prepared answer from the professional.

In practice, the role of the person who answers changes. They spend less time copying opening hours or terms, and more time on requests that call for judgement. They also become the first source of improvement: every time they correct an answer or complete a document, the assistant answers the next question better.

Should you tell visitors they are talking to an AI?

Yes. The European regulation on artificial intelligence, known as the AI Act (Regulation (EU) 2024/1689), provides in Article 50 that people must be informed that they are interacting with an AI system, unless this is obvious from the context. The published text sets this transparency obligation to apply from 2 August 2026. The official text is available on EUR-Lex.

Beyond the obligation, it is a matter of trust. A visitor who knows they are talking to an assistant better understands why it cites its sources and offers to hand over. A name for the assistant, a clear mention at the top of the conversation and an honest greeting are usually enough.

Conversations also contain personal data: a name, an e-mail address, sometimes details about a situation. The GDPR applies just as it does to a contact form. Our article on AI assistants and the GDPR covers the points to settle before going live.

Common mistakes when choosing

The same mistakes come up often, whatever the tool:

• Choosing on a generic demonstration. An assistant that answers well on documents prepared by the supplier says nothing about how it behaves on yours
• Putting everything in the knowledge base. Internal notes, drafts or old price lists end up appearing in the answers
• Forgetting the human handover. An assistant with no path to the team leaves the visitor at a dead end, exactly like a script that is too short
• Not naming an owner. Without someone to review unanswered questions, quality drops as soon as the documents age
• Measuring the number of conversations rather than their outcome. The useful figure is the share of requests settled or properly passed on, which the resolution rate measures

The decision grid in five questions

To decide between a script, an AI assistant or a mix of both, answer these five questions with your team:

1. Do your customers' questions look alike? If they fit into ten stable phrasings, a script or an FAQ is enough. If every customer asks in their own way, an AI assistant fits better
2. Are the answers written down somewhere? A document-based assistant needs documents. If the answers only live in the team's heads, write them down first, whatever the tool
3. Do the answers change often? Rates, terms or opening hours that move every season favour the document-based approach, where you change the source once
4. Do you receive questions in several languages? In Belgium, that is common. One path per language multiplies maintenance; an assistant that answers in the visitor's language from the same documents reduces it, provided you check the meaning in each language
5. Who takes over the conversation when the tool does not know? Without a clear answer to this question, neither tool will give good results

Two "yes" answers to the first four questions usually mean an AI assistant is worth a trial. The fifth question needs an answer in every case.

How do you choose without launching a project that is too broad?

Choose a limited scope with clear answers and success criteria. Start with frequent requests, then widen it when the documents and the team's organization allow.

A reasonable first scope often fits in one sentence: "answer questions about our services, our public rates and our lead times, and pass the rest to the team". Set two or three success criteria before you start, for example the share of questions settled without the team, the quality of the requests passed on and the review time needed each week. After a month, review the conversations, complete the documents and decide whether to widen the scope.

KernChat, developed by KERN-IT in Brussels for SMEs, starts with a one-hour audit of your questions, documents and tools. Its set-up method combines sourced answers in French, Dutch and English, testing and preparation of the human handover.