Chatbot

What is a chatbot?

A chatbot is software that has a conversation with a user through text or speech. The user asks a question or chooses an option, and the bot replies without a human typing the answer in real time.

The word covers a wide range. A website widget with three buttons is a chatbot. So is an assistant that answers open questions about a return policy or internal HR procedure. The experience may look similar, but the technology behind it can be very different.

For many SMEs, a chatbot is most useful as first-line support: it answers recurring questions and routes the hard cases to a person.

Types of chatbots

Rule-based chatbots
These follow fixed flows, keywords, and menus. They are predictable and cheap, but brittle. If the user asks something outside the script, the bot has nowhere good to go.

LLM chatbots
These use a large language model to understand freer wording and write more natural answers. They are flexible, but they can hallucinate: answer confidently with something that is not true.

RAG chatbots
A RAG chatbot retrieves relevant information from your own documents or systems before answering. The language model still writes the response, but the facts come from a knowledge base, such as policies, product pages, manuals, or tickets. This is usually the safer pattern for business knowledge.

Chatbot versus AI agent

A chatbot mainly answers. An AI agent does work. The line is blurry in product marketing, but the difference matters in design.

A chatbot tells a customer how returns work. An agent creates the return label, updates the order in the ERP, and emails the customer. The agent needs tools, permissions, logging, and stronger safety controls because it changes systems, not just text.

A sensible path is to start with the chatbot. Learn which questions users ask, which handoffs happen often, and which tasks are safe enough to automate later.

Where chatbots are useful

• Customer support. Opening hours, delivery times, return rules, invoice questions, order status, and routing to the right team.

• Internal knowledge. HR policies, IT procedures, onboarding documents, and process instructions.

• Lead qualification. A visitor gets first answers after hours and leaves structured context for sales.

• Appointment booking. The bot collects the necessary details and hands off to a calendar or booking system.

A chatbot is irritating when it blocks the user from reaching a person. Always provide an escape route, and pass the conversation history to the human handoff so the user does not have to start again.

What to watch out for with chatbots

Scope
A bot that answers everything badly is worse than a bot that answers one domain well. Start with a narrow set of questions and expand from real logs.

Sources
For LLM chatbots, connect trusted knowledge sources and show citations where possible. If no source is found, the bot should say so instead of improvising.

Escalation
Make handoff to a human part of the design, not an afterthought. The bot should recognise frustration, uncertainty, sensitive cases, and requests for a person.

Privacy
Chat logs can contain names, emails, order numbers, health information, or employee issues. Treat them as personal data, set retention rules, and check where your vendor stores them.

Transparency
Under the EU AI Act, users generally need to be informed when they are interacting directly with an AI system unless that is obvious from context. Do not pretend an AI chatbot is a human employee.

Common chatbot platforms

Microsoft Copilot Studio
A low-code platform for building agents and chat experiences. It can use knowledge sources such as websites, SharePoint, Dataverse, and enterprise data, depending on configuration and authentication.

Helpdesk and ecommerce chat
Many support and webshop platforms include chat and automated replies. They are quick to start with, but limited by what the platform exposes.

Custom LLM implementation
A development team can build a chatbot directly on model APIs, vector search, and internal systems. This gives the most control, while testing, guardrails, evaluation, and maintenance stay with your team.

How to judge whether a chatbot works

Do not judge only by how fluent it sounds. Track containment rate, handoff rate, user satisfaction, wrong-answer reports, unanswered questions, and the cost per resolved conversation. Read samples of real conversations regularly. The logs are where you find both the bot's weaknesses and the process problems your website may be hiding.